Who I am

Most of my work is about the same thing regardless of what we call it: helping organisations tell the difference between a shift that will matter and one that won’t. In AI, that means knowing which capability is genuine and which is a demo in a good suit. In product, which interface survives the AI transition and which doesn’t. In creativity, which tool gives you leverage and which just gives you more output. Closing that gap is worth doing carefully.

Why this exists

Most writing about technology is reactive. It chases the launch and the share price, and by Friday it is landfill.

This is built the other way round. Each week I take one shift and stay with it. A shift in AI. In how products are built. In how creativity works. In how organisations make decisions. I stay with it until it is useful to you: where it stands now, where I think it is heading, and the one move worth making.

Researched. Openly opinionated. Honest about what we do not yet know. Written for people who need more than a headline and have no patience for hype.

The name

Every major technological wave is the comet: vast, cold, fast-moving, and very easy to be swept along by. Mobile was one. Social was one. AI is the current one. The comet changes. The patient instinct doesn’t.

Strategy is the opposite instinct. The deliberate position you hold while everyone else reacts. Hold that position. Let the noise burn off, and the impact arrives the way a comet does. Rare, and impossible to miss.

Patient Comet is a standing reminder: patience and precision outlast speed and panic.

How each article is built

Every piece follows the same discipline. It opens on something real and documented, never a hypothetical. It makes one argument, stated plainly.

Every load-bearing number is checked against a primary source and cited, and any figure that is shaky gets flagged as shaky rather than dressed up to look certain. Before I land my own view, I give you the strongest version of the case against it. Then it ends on one thing you can actually do on Monday.

No filler. Nothing you have to take on faith.

The Best Demo Anyone Had Ever Seen

Vibe coding , the term coined by Andrej Karpathy (a researcher who co-founded OpenAI and previously led AI at Tesla) in early 2025, is a way of building software by describing what you want to an AI in plain English, rather than writing the code yourself. You tell the model what you want, it writes the code, and you refine by describing problems rather than debugging lines. The model types. You provide the vision.

It works. Brilliantly. For the first layer.

There is a moment in every demo room, physical or digital, where something clicks.

The product loads. It responds. Someone types something and it does the right thing. Faces change. The questions stop. Somebody says “when can we have this?” and everybody else is already nodding.

That moment happened millions of times in 2025 and 2026. It happened with unprecedented frequency because anyone could now create a working, responsive, convincing product in an afternoon.

In February 2026, a man called Moltbook into existence this way. An AI social network (a place for AI agents to interact and post) built without a single line of human-written code. The founder said so publicly. Karpathy praised it. The internet arrived. It loaded. It animated. It responded.

It looked, by every measure a normal person would use, completely finished.

Three days later, a security firm called Wiz (a US cybersecurity company) opened the JavaScript that Moltbook’s server sent to every visitor’s browser. Inside it: a database key (a password to the entire database) sitting in plain view. Moltbook was built on Supabase (a popular cloud database service), which requires Row Level Security (the policy that controls which users can read which rows) to be explicitly switched on. It had not been. Default: everything open to anyone with the key. The key was in the code every visitor downloaded automatically.

A researcher exploited the full breach in under three minutes.

Behind those rows: 4.75 million records. One and a half million API keys (passwords to other connected services). Thirty thousand email addresses. Thousands of private messages (Wiz, 2026).

Three days from launch. Under three minutes to exploit.

The demo had been perfect. The product had never existed.

That gap has a name. And it is not a bug in the tools.

Frontend and Backend Is Not Full-Stack

The misconception at the centre of almost every vibe coding failure is not about the tools.

When a non-technical founder, product manager, or first-time builder thinks “full-stack” (the complete thing, front to back) they picture two layers: the interface the user sees, and the server that powers it. Frontend and backend. The AI builds both. The prototype demonstrates both.

A production-ready product requires all thirteen. Auth & Permissions (controlling who can access what). Security & RLS (the policies that enforce those controls at the database level). Rate Limiting (stopping a single user from overwhelming the system). CI/CD & Version Control (a pipeline of automated steps for updating the product safely without breaking what already works). Error Tracking & Logs (knowing when something fails, how, and why). Availability & Recovery (continuing to run , and returning to normal when it does not).

None of these appear in a demo. The demo is a performance of layers one and two. The other eleven have no role in it.

And the diagram does not show everything. It shows the technical stack. What it does not show is the design layer , the discipline that determines whether anyone actually wants to use what you built.

A vibe-coded interface assembles functional elements without a design intention running through them. It is like a room furnished from eleven different houses. The chair came from a kitchen, the desk from a bedroom, and the light was designed for a hallway. Each piece works on its own. The room never does. This is not a cosmetic problem. It shows up in conversion rates, churn, and the gut feeling every user has that something is slightly off, before they can name it.

In an audit of twelve production-ready vibe-coded apps, four had live payment credentials or database master keys shipped directly in the front-end JavaScript, visible to anyone who opens the browser developer tools (getautonoma.com, 2026).

These are not edge cases from one bad actor. Moltbook was not even the worst of them.

The Gap a PM Can't See

Moltbook is the story everyone cites. The Quittr story is the one that matters more.

In September 2025, a habit-tracking app called Quittr launched with AI coding tools. Within ten days it reached one million dollars in revenue. Oprah Winfrey mentioned it. It was the kind of launch that makes investors reach for their phones at midnight.

What no one knew was that Quittr’s Firebase database (the cloud service where it stored all its user data) was publicly readable. Any authenticated user could access the full backend, where the records of more than 600,000 people were stored openly. The records included ages, triggers for behavioural habits, and personal confessions about users’ struggles (Cybernews, 2026).

A security researcher found it and contacted Alex Slater, Quittr’s co-founder, on 10 September 2025. Slater’s response: he would fix it “in the next hour.”

He did not fix it for months.

This is the product manager trap. As of May 2026, 63% of the people building applications with vibe coding tools have no coding background (TechTimes, 2026). The majority of people making build-and-ship decisions about AI-generated products cannot see the eleven layers that are missing. Not because they are careless. Because they have no frame for what is behind the wall.

Slater had built something that worked, that was making real money, that Oprah had mentioned. In his mental model, things were mostly working. The security misconfiguration was invisible to him. So was its severity.

The same blindspot hides design failures. A product manager without design experience cannot see that the first few screens of the app are losing eight in ten users before they ever reach the product. They cannot see that error messages are written in developer shorthand that no user understands. They cannot see that the mobile layout is technically responsive but requires a precision of thumb movement real users do not have.

A building inspector knows what to look for behind the wall. A designer knows what to look for in a user’s first sixty seconds. Someone who has only ever toured the show home sees neither.

This Pattern Keeps Repeating

If Moltbook and Quittr were isolated incidents, you could call them bad luck.

In late 2025, Enrichlead (a software product sold on subscription) watched users find a complete paywall bypass within 72 hours of launch, by changing one value in a hidden developer panel built into every web browser. The check that verified whether a user had paid existed only in the front end , the layer the user controls. Anyone could simply remove it (getautonoma.com, 2026).

Also in 2025, Lovable’s AI generated access control logic across 170 production applications (the rules that decide who can touch what) and got it backwards. Users who had logged in were blocked. Everyone else had open access. The AI had implemented the security layer. It had simply inverted it (CVE-2025-48757, a formally recorded software vulnerability).

Amazon’s internal teams, deploying vibe coding tools with experienced engineers, produced four Sev-1 failures (the most critical level of production incident) in ninety days, including a six-hour outage.

And in April 2026, Lovable the platform (used by eight million people, valued at $6.6 billion) ran for forty-eight days with a broken authorisation flaw. Every user’s source code, database credentials, and AI chat histories were readable by any other logged-in user. A bug report was filed. Lovable’s team reviewed it, marked it “intentional behavior,” and closed it without escalation. They fixed it in two hours once the story went public (The Next Web, 2026).

Across 200+ independent audits: 91.5% of vibe-coded apps contain at least one vulnerability (GuardMint, Q1 2026). More than 380,000 publicly accessible vibe-coded apps on the open web, 2,000+ actively leaking data , no attacker required (Red Access, 2026). AI-generated code now causes one in five enterprise security breaches (per industry analysis, 2026). These are not rare edge cases. They are the default outcome of shipping eleven unbuilt layers.

Four Failure Modes

These incidents look different. One founder, three days. Another founder, months. A platform, forty-eight days. Amazon, ninety days. But they all fail in the same four ways.

Technical failure is the most documented. Missing authentication, exposed database keys, access rules never configured , or configured backwards. These are fixable. The thirteen-layer framework tells you exactly where to look. Technical failure happens when the layers under the interface are never built by people who know what building them requires.

Design failure is less discussed and equally real. A vibe-coded interface assembles functional elements without a design intention running through them. The user journey was never mapped. Error states were never written in language users can act on. Onboarding was never tested against how a real person reads and moves. Mobile “works” but is not designed for how people actually use their thumbs. Users leave products that feel wrong before they can name why. And a non-technical PM cannot see this failure any more than they can see a missing security layer , because they have no frame for what a well-designed product experience looks like from the inside.

Organisational failure is quieter and more consequential. Slater was not negligent. He was operating in a mental model where the demo was the product. When the person making decisions about readiness has never built or designed any of the layers underneath, those layers are permanently invisible to the decision-making process. The same blindspot covers security and design.

Perceptual failure is the one that makes the other three inevitable. It is last on this list, but first in the chain, everything else flows from it. It happens in the demo room. The prototype loads, clicks, and does the right thing. The investor nods. The client says “when can we have this?” The PM says “soon.” Everyone has experienced the same convincing illusion: a working product. What they experienced was two layers of thirteen, with no design system behind it.

Prototypes used to look like prototypes. Now a vibe-coded prototype loads, responds, and delivers the product’s core value proposition completely. The gap between it and a real product is invisible. Invisible gaps are the most expensive kind.

What the Prototype Is Actually For

None of this is an argument against building fast. The speed is real and useful.

The prototype is the most precise brief you will ever hold.

It communicates to engineers, to designers, to security specialists, to stakeholders exactly what you are trying to build, with a fidelity that no slide deck or written spec can match. You built the interface. You can demonstrate the flow, the feel, the user journey. You have eliminated the most common cause of failed software projects: the gap between what the people building it imagine and what the people wanting it actually need.

That is where the prototype’s job ends.

Treating it as a brief resolves all four failure modes at once.

It fixes the technical failure, because the brief goes to specialists before launch, not after the breach. It fixes the design failure, because the prototype is handed to a designer who uses it as a reference for the actual design system: the user journey, the error states, the onboarding, the mobile experience, the copy. The prototype shows what is wanted. The designer builds what will actually work. It fixes the organisational failure, because the PM’s role shifts from “is this ready to ship?” to “have the engineers and the designer finished their layers?” Those are questions anyone can ask and verify. And it fixes the perceptual failure, because when the prototype is a brief, the demo room’s question becomes “is this what we want to build?” not “when can we launch this?”

The right workflow: prototype → brief engineers and a designer → engineers build layers 3-13 → designer builds the design system → security review → launch.

The wrong workflow is: prototype → someone who cannot see the gap says “it looks ready” → launch.

Most vibe-coded products that reach production fail. Not because the tool failed. Because the prototype was shipped as the product, without the engineers who build the technical layers, without the designer who builds the product experience, and without the security review that confirms both are done.

The Cost of Looking Finished

The Bill That Changed the Conversation

In May 2026, the story that changed how enterprise AI is discussed did not make every front page.

The number that changed every CFO conversation in 2026 surfaced quietly. An unnamed company had run up a $500 million bill on Anthropic’s Claude in a single month. Not through negligence, exactly , through deployment. They had given their entire workforce access to the model, with no usage caps and no governance. The agentic workflows (automated loops where the AI plans, acts, reads the result, and acts again, re-sending the full conversation at every step) had turned what looked like a modest per-query cost into a figure that exceeded most companies’ entire annual technology budget (byteiota.com, May 2026).

Microsoft’s story broke around the same time and was more specific. The company had invited thousands of engineers in its Experiences and Devices division to use Claude Code starting in December 2025. By May 2026, the tool had become the preferred choice inside the division. Token costs per engineer began outpacing what Microsoft paid those engineers in salary. The company set a June 30 cutoff and redirected engineers to Copilot CLI (Cybernews; The Next Web, 2026).

Uber had got there first. Its operations chief told The Information in April that the company had burned through its entire 2026 AI coding budget in four months , roughly 5,000 engineers on Claude Code, the heaviest users spending $500 to $2,000 each per month individually (Tom’s Hardware, 2026).

None of these are outliers. They are case studies in a structural problem: the price of AI has collapsed, so every new use case looks affordable. Then you do the arithmetic on volume.

The term for the dynamic: LLMflation, coined by a16z (Andreessen Horowitz, a major Silicon Valley venture fund) in 2024 to describe the paradox where AI costs per call keep falling while AI bills keep rising. A model good enough to pass a standard knowledge test cost roughly $60 per million tokens (each token is roughly three-quarters of a word) in late 2021. By 2024 it was $0.06 , a fall of roughly 1,000× in three years, faster than the PC era’s decline in compute. But AI bills per action, not per seat. Lower prices triggered far more actions, not lower bills. Token prices fell roughly 1,000× over three years while total enterprise AI spend rose more than 300% in two (a16z, 2024; Ramp, 2026).

The Meter Never Stops

Three things are happening simultaneously inside most enterprises. They look unrelated. They are the same problem.

The first: usage consistently outruns the forecast. In 2025, 85% of companies missed their AI cost projections by more than 10%; 24% missed by more than 50% (Mavvrik / Benchmarkit, 2025). Two years ago, 31% of finance teams actively managed AI spend. Today it is 98% , the number doubled not because companies got disciplined, but because the bills arrived and demanded attention (State of FinOps, 2026).

The second: agents turn the meter into a fire hose. A standard AI chat interaction uses a few thousand tokens. A million tokens is roughly a 750,000-word document , the equivalent of asking the model to read and write War and Peace from scratch, for every complex task. An agentic workflow re-sends the entire conversation context (the AI’s working memory) at every step. Industry data puts agent token use at 10-100× a standard chat. Stanford’s Digital Economy Lab measured the most complex agentic tasks at 1,000× a standard reasoning call. Simple tool-calling agents consume 5,000-15,000 tokens per task; complex multi-agent systems consume 200,000 to over 1,000,000 tokens per task (Stanford Digital Economy Lab, 2025). Goldman Sachs estimates agents will multiply total enterprise token demand 24× by 2030.

The third: the most expensive model is rarely the right model. The price gap between the cheapest production AI model and the most expensive frontier reasoning model is currently about 4,500×. The cheapest runs at around $0.04 per million tokens. The most expensive frontier reasoning models run at $180 per million tokens. Most enterprises have deployed the top tier to everyone. Analysis consistently shows that roughly 85% of enterprise queries could be handled by budget-tier models (the ones at the $0.04 end) with no meaningful quality loss (FinOps Foundation (the industry body tracking AI infrastructure costs) 2026).

The hidden overhead compounds everything. OpsLyft (an AI cost analytics firm) found that most teams underestimate their true AI bill by 40-60%. A call that looks like $0.05 typically lands at $0.20 once embeddings (text to searchable numbers), retrieval, re-ranking, output validation, retries, and prompt overflow are counted (OpsLyft, 2026).

This Comes Straight Off Your Margin

Inference is not a technology line item. It is cost of goods sold, the costs that come directly out of gross margin, the number investors use to price companies.

84% of companies reported that AI was already eroding gross margins by more than 6% (Mavvrik / Benchmarkit, 2025). Bessemer’s portfolio data puts AI-native gross margins at 50-60%, against 70-90% for mature SaaS, software economics is quietly turning into manufacturing economics (Bessemer, 2025). Even OpenAI, the company whose products power much of this spending, ran $13B in revenue against $22B in costs in 2025 , a $9B net loss at scale (Fortune, 2025).

The pattern is consistent: the more AI gets used, the more it costs, and the less of that cost is recovered in measurable output. This is not a price problem. It is a volume and governance problem. The companies that have protected their margins are not the ones with cheaper models. They are the ones that controlled who was using which model for what. There is a reason for that, and it comes before usage.

The Access Problem

The piece most AI cost discussions miss is the one that makes every other problem harder to fix.

In 2026, Meta created internal leaderboards ranking employees by token consumption , a practice that quickly became known as “tokenmaxxing” (spending tokens competitively, not productively). Employees competed to spend the most tokens, throwing entire projects at agentic AI without any discipline about whether the output was good, whether a cheaper model would have worked, or whether the task was suited to AI at all. The leaderboard rewarded consumption, not outcomes (industry reporting, 2026).

Microsoft’s experience tells a version of the same story. The engineers who used Claude Code most heavily were not necessarily the ones producing the best software. Token costs outpaced salaries. And Faros AI found that code churn (lines of code deleted versus added, a measure of rework) increased by more than 800% in teams with high AI adoption. The teams spending the most on AI were producing the most waste (Faros AI, 2026).

The data points to something the economics alone do not explain: AI amplifies what you bring to it.

It is worth saying plainly that broad access to powerful tools has genuinely produced unexpected value. The marketer with no technical background who caught a product flaw through an unusual prompt. The junior analyst who used AI to try an approach a veteran would have ruled out on instinct, and was right. Ungated access is not without merit. The argument here is not for restriction; it is for matching model tier to demonstrated task competence, with a clear path to qualification for anyone willing to earn it.

That said, the amplification argument is real and the data is hard to ignore. A senior engineer with fifteen years of experience using Claude Code produces faster, better-reviewed code at lower cost, because they can evaluate the output, catch the errors the AI cannot catch, and prompt with precision. Evaluation means spotting when the model has hallucinated a citation (invented a source that does not exist), reversed a logical condition in code, or produced a plausible-looking answer that is wrong in a way only domain knowledge reveals. A junior developer with six months of experience using the same tool produces faster, worse code, because they cannot tell good output from plausible-looking output, they iterate blindly on prompts they cannot evaluate, and they burn tokens on approaches that an experienced engineer would immediately dismiss.

The 4,500× price gap between the cheapest and most expensive models is not the real cost driver. The real driver is the gap between what a skilled user produces per dollar and what an unskilled user produces per dollar. A budget model in the hands of someone with deep domain expertise costs almost nothing and produces real value. A frontier reasoning model in the hands of someone without the background to evaluate its output costs $180 per million tokens and produces fast, confident, unreviewed work that someone else will eventually have to fix.

This is not a controversial idea when applied anywhere else. We do not give junior surgeons operating theatres without supervision. We do not give junior analysts sole authority over financial models that go to a board. We understand, in every other context, that tools amplify skill, and amplifying the absence of skill at speed is not progress. We have simply forgotten to apply this to AI.

Access to frontier AI models (most powerful, most expensive) is a professional resource. It should be treated like one.

Govern the Usage and the Access

The companies that have escaped LLMflation, the roughly 6% McKinsey calls “AI high performers”, are not the ones with the cheapest models. They are running two governance layers simultaneously: usage controls and access controls.

Usage governance: four controls that compound

Route every task to the cheapest capable model. A cascade architecture (routing tasks to cheaper models) sends simple calls to the budget tier, escalates complex ones to frontier models only when the task requires it. Analysis consistently shows this cuts inference (the cost of running the model) cost 40-70%, because most queries never need to leave the budget tier (FinOps Foundation; arXiv, 2025).

Cache the answers you have already paid for. Semantic caching (reusing stored AI answers) removes about 31% of repeat queries before they reach the model. Prompt caching (reusing unchanged prompt sections) makes cached tokens roughly one-tenth the standard price. Add batch pricing on top (roughly 50% discount) and a call that would have run at $0.20 runs at about $0.01 , about 5% of standard cost (OpenAI; Anthropic API docs, 2026).

Cap the spend before it surprises you. Token budgets as hard ceilings per team and per workflow. Fix retry logic that can loop indefinitely. Define maximum context windows for each use case. The $500M bill happened because there were no ceilings.

Measure cost per outcome, not cost per call. “Cost per call” tells you almost nothing. “Cost per resolved customer query,” “cost per code review,” “cost per draft approved” , these are the numbers that tell you whether the spend is working. The vendors that price this way, Intercom at $0.99 per resolved query, Salesforce Agentforce at roughly $2 per conversation, already know their cost per outcome. The discipline turns the meter into a dial (Intercom, 2026; SaaStr, 2026).

Access governance: the layer most companies skip

Tier model access by role and demonstrated competence. Not everyone needs frontier reasoning. Not everyone should have it. The qualification for moving up a tier is not seniority, it is demonstrated domain knowledge. The ability to evaluate what the model produces in that domain and catch what it cannot catch.

The upgrade is available to anyone. A junior developer on the budget tier earns frontier access when they can demonstrate they can evaluate what a frontier model produces in their domain, spotting the errors, assessing the quality, catching the hallucinations. That is not hierarchy preservation. It is competence development with a measurable gate.

The advantage is real, and it is unequal: when skilled engineers have frontier access and unskilled ones are routed to budget tiers, the skilled engineers work faster and the budget recovers its balance. The companies with ungated access to everything get neither the speed nor the savings.

Run both governance layers and the cost reduction runs to 60-90% against ungoverned baseline, according to industry analysis (FinOps Foundation, 2026).

Before the verdict, the three honest positions deserve a fair hearing.

Bubble, or Breakthrough?

Three honest positions, all with something right in them.

Ungoverned Access Is the Real Bill

The AI Subscription Is Breaking

For the past five years, shipping an AI feature meant one thing.

You built a product. You had an idea for an AI feature (a smart search, an automated summary, a document assistant). You signed up for an API (a connection to someone else’s AI server), and it worked. Impressively, immediately, well.

Your customers loved it.

What they did not see was what happened on every query. Their data (their documents, their messages, their records) left your product, travelled to a server owned by Microsoft, Google, or Anthropic, got processed by a model you did not own, and came back as an answer. You paid for every token, every word, every interaction. The cost scaled with your users. The faster you grew, the larger the bill.

The intelligence was rented. Your product was the front end. The AI lived somewhere else.

For most teams, this was the only option. The models worth using required infrastructure so large and expensive that building or owning them was simply not realistic. You connected to the cloud because the cloud was where the intelligence lived.

That assumption just broke. And it broke on three separate occasions, between January 2025 and April 2026.

Three Events That Changed the Math

Each one dismantled a different part of the old model.

January 2025: DeepSeek releases R1. DeepSeek (a Chinese AI research lab) released a reasoning model (an AI system built to think through complex problems) and published the full method openly. When the training cost emerged in a peer-reviewed Nature paper, it landed like a small bomb: $294,000 (Nature; CNN, 2025). Not $294 million. Less than the price of a family car. For a model that matched the best in the world, then given away free.

The markets understood immediately. The first R1 release had already wiped roughly $589 billion off Nvidia (the company that makes the specialist chips AI runs on) in a single day , the largest single-day loss in US market history (CNBC, 2025). Investors were not frightened of one lab. They were frightened of the assumption under their entire position: that only the very largest companies could build intelligence worth using.

April 2026: Google releases Gemma 4. Google’s Gemma 4, at 27 billion parameters (a rough measure of how much the model has absorbed, like a proxy for its depth of knowledge), beats models ten times its size on human-preference testing (including Llama 405B, Meta’s large open model , and does this on a single GPU (the specialist processor AI uses) (Google, 2026). A model you could host yourself, on one card you own, winning blind taste-tests against rivals that need a computing cluster.

Ongoing 2025: Apple ships AI on the device. Apple builds a roughly three-billion-parameter model directly into its devices, available to every developer, with inference that costs nothing and runs locally , meaning the data never leaves the phone (Apple, 2025). No contract. No server. No log sitting somewhere waiting for a legal request.

Three events. One implication. The ingredient that made powerful AI expensive and remote has become something a product team can own, fine-tune, and ship. The question is no longer whether you can build AI into your product without the cloud. The question is whether you are going to.

What Builders Can Now Do

Until recently, shipping an AI feature meant one thing: connect to an API and rent someone else's intelligence. An API (a connection that lets your product borrow intelligence from an external service) was the only viable route. Your product was the shell; the intelligence was always somewhere else, always rented, always metered.

Open models (AI models whose design and weights are published freely) change the product equation entirely.

A software team can now take one of these models, fine-tune it (adapt its behaviour by training it further on their specific domain and data), and ship it as a permanent, built-in part of their product. The model travels with the software. When a customer buys the product, they get the AI too.

Not a subscription to the AI. The AI itself.

Think carefully about what this removes.

No per-query token costs at scale. Once the model is built in, the marginal cost of an AI interaction drops to the compute the customer already owns. No external API dependency. The product works offline, in environments where data cannot leave the building: hospitals, law firms, government offices, banks. And no third-party subscription invisibly embedded in your pricing.

The customer owns what they paid for. Completely.

Now think about what it creates.

A product with a fine-tuned model built in is structurally harder to replicate than one that connects to a shared API. A competitor cannot switch to a better API endpoint and close the gap overnight. The model, trained on your domain knowledge, shaped by your users’ actual needs, integrated into your product’s logic, becomes part of what you ship, and part of what makes it yours.

The pricing model changes too. SaaS products with AI features charge recurring subscriptions partly because they pass through API costs. When the model is built in, that cost disappears. You could sell the software once. Or with a simpler subscription. The AI is included, like a camera in a phone, not a streaming service on your phone.

One honest note: fine-tuning a model for production is a genuine engineering effort. Tools like Hugging Face and Unsloth have made it achievable without a research lab, but it requires a competent ML engineer, proper evaluation, and a realistic timeline. It is not a weekend project. It is, however, now within reach for any well-resourced product team, which would not have been true two years ago.

Apple understood this at the operating system level. The on-device model in every device is not an add-on you pay extra for. It is the product. Every software builder now has the same option at the application level.

That option opens markets that the subscription model could not reach at all.

What This Unlocks for Regulated Industries

There is a version of the builder opportunity that is not just about economics. It is about which markets you can serve at all.

For a significant and fast-growing portion of the software market, cloud AI is not a choice. It is off the table.

A medical device company cannot sell a diagnostic tool that sends patient data to a US cloud server under GDPR and HIPAA. A legal technology firm cannot win enterprise contracts in regulated jurisdictions if their AI feature sends every query to OpenAI. A government software supplier cannot pass a security review if the intelligence in their product lives in a data centre they do not control.

For these markets, the product that wins is the one where the AI runs locally, the data never moves, and the intelligence ships with the software.

This is not a compliance headache. It is a competitive opening, and it just got significantly larger.

In June 2025, the legal counsel of Microsoft France was asked under oath at a French Senate hearing whether he could guarantee that data stored in France by Microsoft would never be passed to US authorities without French approval. His answer was four words.

“Non, je ne peux pas le garantir.” No. I cannot guarantee that (The Register, 2025).

Under the US CLOUD Act (a 2018 law giving American authorities the right to demand data from US-headquartered companies regardless of where that data physically sits), an EU data region gives you lower latency and a reassuring label. It does not give you jurisdiction. A Microsoft executive said so, on the record, to a parliament.

The legislative response has followed. On 27 May 2026, the European Commission proposed restricting Microsoft Azure, AWS, and Google Cloud from processing financial, judicial, and healthcare data across all 27 EU member states (CNBC, 2026). Those three providers control roughly 70 per cent of Europe’s cloud market. The proposal carves out the exact categories of data where the most valuable enterprise software operates.

A product maker who ships with a local, fine-tuned model is not just removing an API dependency. They are entering markets that their cloud-dependent competitors structurally cannot. That is a durable advantage, because the legislative direction is accelerating, not reversing.

When Cloud Still Wins

Local models do not win everything. The honest version of this decision has four camps.

The router wins the argument, but only when the routing is designed rather than defaulted.

The owner is right that the old assumption has expired. The renter is right that the frontier gap is real on the hardest work. Both observations are correct and neither is a complete policy on its own. The mistake is letting either one become the answer for everything.

The organisations, and the product teams, that come out ahead will be the ones that make a genuine per-workload decision: sensitivity, volume, capability required. Write it down. Apply it consistently. Do not revisit it every time a new model is announced. That one-page document is worth more than almost any model selection you make this year.

Four Moves for Builders

The Era of Renting Intelligence Is Ending

When the Reader Was Always Human

The web had always been built for a simple transaction: a person looks, a page loads, a person reads. The entire architecture of search engine optimisation was built around that transaction. SEO (search engine optimisation) is simply the discipline of making your page findable: visible to the right person at the right moment.

That model lasted thirty years, which, in technology, is an eternity.

What ended it was not ChatGPT. It was not Perplexity, or Google’s AI Overviews (AI-generated answer summaries that appear above search results), or any single product announcement. It was something quieter: the slow accumulation of a new primary reader.

Search engines have always had crawlers. Crawlers read your pages constantly, invisibly. They understood what you had written and surfaced it to humans who searched for it. The crawler was always there; the human was always the destination. The crawler served the human.

What changed is that the machine stopped serving. It started answering.

Google’s AI Overviews now appear on 48% of all search queries as of March 2026, up from 34.5% just three months before. In those queries, Google no longer sends the user to a page to read. It synthesises an answer from many pages, presents it at the top, and the user never clicks through. They have their answer. They are gone.

The click-through rate for pages that appear in AI Overview queries has dropped 58% according to Ahrefs (an SEO analytics and research platform) in their December 2025 study of 300,000 keywords. Seer Interactive (a digital analytics agency) measured 61% in their September 2025 research. The precise number depends on who measures it, but the direction is not ambiguous.

The machine is no longer a courier. It is the reader. That particular reader still has a name you know well.

AI Reads First Now

Google still controls 90% of the global search market. As of March 2026, StatCounter (a web service that tracks global search market share) puts it at 90.01%. All AI search engines combined, including ChatGPT, Perplexity, Claude, and Gemini Search, account for somewhere between 0.28% and 0.9% of referral traffic, depending on how you measure. The death-of-Google narrative is empirically wrong.

What is happening is not Google dying. It is Google absorbing its own replacement.

Google is building the machine reader into itself. AI Mode, the full conversational search interface Google launched at I/O 2026, crossed one billion monthly users, with queries more than doubling every quarter. The product that replaced the blue links is Google. The company cannibalising the click-through economy is the same company that created it.

The distinction matters practically. Google’s AI features run on RAG, or retrieval-augmented generation (a technique where Google’s AI searches its own index first, then composes an answer from those pages, rather than inventing one from scratch). In plain terms: it reads real pages before it writes a summary. Being well-indexed by Google still matters for being cited by Google’s AI.

But ChatGPT and Perplexity run their own independent crawlers (the software bots that read websites) and do not use Google’s index. Good Google SEO does not automatically translate to citation by those systems. The machine reader is plural, and each instance of it has its own appetite.

The structural shift, stated plainly: the pages you publish are now read, primarily, by a machine that decided your content is worth summarising to a human, rather than worth sending the human to read directly. The human may never arrive. The machine took what it needed.

For thirty years, the relationship was: search engine helps human find page. The page is the destination.

The relationship is now: AI engine helps human get answer. The page is the source material.

That shift, from destination to source material, is not a minor update to how search works. It is a different model of what the web is for.

The Traffic Number That Changes the Math

That data has a specific implication for anyone running a content-dependent business.

Roger Lynch’s “plan for zero” directive deserves to be understood in its full weight.

Lynch did not say Google traffic will become unimportant. He said his teams should build their editorial and revenue strategies as if it does not exist, because counting on it will produce a plan that fails. You plan for zero not because zero is certain, but because every year you plan for less-than-zero the actual number undershoots your model.

He is not alone. Chartbeat (a web analytics platform that tracks traffic for major publishers) data shows search referrals fell 60% for small publishers, 47% for medium publishers, and 22% for large publishers over two years. News publishers collectively lost more than 600 million monthly visits. Chegg (an online homework and tutoring company) declined 49%. Forbes, HuffPost, and Business Insider each lost roughly half their search-driven traffic.

These are not struggling operations that made wrong bets. These are established publishers with large audiences, real editorial teams, and decades of SEO expertise. They did everything right by the old model. The old model changed.

What Lynch named the shape of what remains: a barbell. Heavy at both ends. Nearly empty in the middle.

Large, authoritative brands with genuine audience trust are holding up. Small, niche publications with devoted direct subscribers are holding up. The middle, the vast category of SEO-dependent content built on traffic arbitrage, ad revenue, and Google as the primary distribution channel, is collapsing.

This is not a media story. It is a business model story. The barbell predicts the shape of every knowledge business over the next five years. The question for anyone who publishes, teaches, advises, or sells expertise on the web is: am I the authoritative end of the barbell, the loyal-niche end, or the exposed middle?

The practical answer requires understanding what the machine is actually reading, and why most of what you currently publish is invisible to it.

What the Machine Actually Reads

The practical question, what do you do about this, runs directly into a problem: the machine reader does not rank things the way Google did.

With Google, you could build a strategy. You had data: which keywords brought traffic, how you ranked, how competitors ranked, where you had gaps. It was imperfect, but it was measurable. You could spend Tuesday afternoon in Search Console and come out with an action list.

There is no equivalent for AI citation. According to practitioner tracking across AI platforms, forty to sixty per cent of the sources cited by AI systems change month to month. There is no stable “position one” in ChatGPT or Perplexity. There is no dashboard showing how often AI Overview cited you. There are no reports showing which questions users asked conversational AI, every conversation is unique and private.

The most counterintuitive finding from all of the available data:

Reuters and The Guardian are two of the most frequently cited sources in ChatGPT and Perplexity. They receive less than 1% of referral traffic from those systems. They are cited. They are not visited. The machine read their work, extracted what it needed, and kept the visitor. Being cited by AI does not mean being visited. Citation is an awareness signal. Traffic is a separate transaction, and for most publishers, it is not happening.

This breaks the old logic entirely. You cannot optimise your way from “cited” to “visited” using the tools that worked in search. The game has changed.

Google’s own guidance, published in May 2026 as the company’s first official AI optimisation guide, is notable for what it debunks. Specifically: llms.txt files (a proposed standard for declaring your content to AI crawlers, similar in concept to robots.txt, the file that tells web crawlers which pages to access); “chunking” content into small pieces; and rewriting pages specifically for AI. Google says to ignore all three, and independent testing confirms no major AI crawler reads llms.txt. The consultants selling these services are selling solutions to problems that do not exist as described.

What Google says does help is simpler and harder: create content that a human would find genuinely useful, that could not be produced by an AI summarising other people’s work. The word they use repeatedly in the guide is “non-commodity.” Generic how-to articles, roundup posts, and content that merely restates what others have written will not be cited, because the AI can produce it itself without needing your page.

The Princeton KDD 2024 study (the only large-scale academic empirical test of AI citation optimisation, covering 10,000 queries) found three things that increase citation probability. Statistics improve AI visibility by 41%. Named expert attribution with quotation marks improves it by 28%. Inline citations to primary sources improve it by 115% for pages that do not already rank in the top results. Keyword stuffing degrades it. The machine reads for credibility signals, not keyword frequency.

The Four Moves

The investment has four destinations. If you run a publication, a brand, or any knowledge-based operation that depends on being found and read, all four apply.

None of this is new. It is a return to what good journalism always required: expertise, attribution, sourced claims, a clear identity. The difference is that now the first audience those things need to satisfy is a machine.

Three Ways to Think About This

None of this settles the argument entirely. There are three coherent positions on what all of this means, and they do not easily resolve , but all three are worth hearing before committing to any of them.

The machine cannot steal a subscriber.

Each of these is partly right. The converter data supports the first. The citation-without-traffic paradox supports the second. The market share numbers support the third. The mistake is picking one and ignoring the others.

The Play Is Identity

What the Pipeline Actually Looks Like

Google Flow was not the starting gun. It was the confirmation.

The pipeline now fits in one window. In March 2026, ElevenLabs (a voice and creative AI platform used by 41 per cent of the Fortune 500) launched Flows: a single canvas connecting script, footage, voice, music, and distribution, without a single external hire. These launches were months apart. The capability gap that once separated a marketing team from a broadcast production facility had already closed.

Let us look at the mechanics. The speed of the change is still not fully understood.

Three years ago, a brand producing a 30-second video advertisement needed a creative brief, an agency, a production company, a shoot day, a post-production house, a sound designer, and a music licensing deal. Timeline: eight to twelve weeks. Budget: five figures minimum, climbing quickly.

Today, a marketing team at a mid-sized company can move from brief to published video in under a working day. The 2026 production stack looks like this.

Script. ChatGPT (OpenAI’s AI writing assistant) or Claude (Anthropic’s AI assistant) writes a first draft from a brief. Tone, length, platform format, all specified in the prompt. Turnaround: under a minute.

Images and storyboard. Midjourney v7 (from a San Francisco AI company of the same name) creates cinematic concept frames and editorial illustrations , the standard for artistic, mood-driven imagery. Flux 1.1 Pro (from Black Forest Labs, a German AI research company) leads on photorealism, producing images that look like actual photographs rather than generated artwork. For any image requiring readable text inside it (a poster, a product label, a social card) Ideogram v3 (from a New York AI company) is the specialist. It is the only major image tool that reliably renders legible typography; every other model still produces distorted or misspelled text.

Video. Google’s Veo 3.1 (Google’s AI video generation tool) produces broadcast-quality video clips from text descriptions, generating dialogue, ambient sound, and music in a single pass. It is the only major AI video tool that creates audio natively alongside footage without a separate step. Runway Gen-4.5 (from Runway, a New York AI company) is the professional alternative for work requiring precise creative control, consistent characters across multiple scenes, and post-generation editing. Seedance 2.0 (from ByteDance, the Chinese technology company behind TikTok) accepts text, images, video, and audio together as input (up to twelve assets simultaneously) and produces multi-shot, cinema-quality video with native audio sync and frame-level character consistency. Its April 2026 API launch made it directly accessible to brand production teams at scale. One notable absence from this list: Sora, OpenAI’s video generator, was shut down in April 2026. Per industry analysis, the product was generating significant operating losses against minimal revenue.

Voice. ElevenLabs generates voiceovers in more than 70 languages from text input. Voice cloning is available for brands that have established audio identities. Their technology is now embedded in IBM’s enterprise AI systems.

Music. Suno v5 (an AI music generation platform) creates complete original tracks with vocals, instruments, and lyrics from a text description, in seconds. Following Suno’s November 2025 settlement with Warner Music Group (one of the three largest recorded-music companies in the world), paid subscribers hold commercial rights, though as explained in §04, the legal picture is more complicated than that agreement suggests. Udio (another AI music generation platform) settled similarly with Universal Music Group (the world’s largest recorded-music company) in October 2025.

Assembly. ElevenLabs Flows, launched March 2026, connects all of the above in a single visual workspace: image generation, video, voice, music, and sound effects linked together in a pipeline. Non-destructive: change the voiceover without rerunning the video generation. The entire sequence can be triggered from a brand team’s form submission.

Distribution-ready design. Canva AI 2.0 reads your brand’s colour palette, fonts, and logo rules before generating anything. It has Google’s Veo 3 embedded directly for video, and outputs social-ready assets in every required format simultaneously.

That is the 2026 creative production stack , the generation layer. The tools that create the content.

There is a second layer that most brands discover only after they start producing at scale. It is the orchestration layer: the pipeline that takes a brief at one end and delivers a finished, distributed asset at the other, without a human copying files between applications.

Every Brand Is a Studio Now

Unilever (the consumer goods company behind Dove, Lynx, and Hellmann’s) runs 18 AI creative studios across 18 markets. A brief goes in. Campaign-ready video assets , in local languages, featuring the product, formatted for every platform come out. No production company booked. No crew scheduled. No shoot. The system uses 3D digital models of each product, known as digital twins, to generate the imagery. Since deployment: creative assets produced 30 per cent faster. Video completion rate doubled. Click-through rate doubled. A parallel system called Sketch Pro, built with IPG Studios (the production arm of Interpublic Group), delivers testable assets in two hours at three times the previous production speed. Scaling to 21 markets by the end of 2026.

This is not exceptional. It is now the direction of travel.

82 per cent of members of the Association of National Advertisers (the main US industry body for major brands) now operate an in-house agency , up from 78 per cent in 2018. 60 per cent of US senior marketing leaders told Forrester Research (a global market research firm) in 2025 they spent less on external agencies as a direct result of AI. Forrester projects a 15 per cent job reduction across agencies in 2026, following 8 per cent cuts in 2025.

The numbers at the holding-company level tell the story directly.

WPP (one of the world’s largest advertising networks, owning agencies including Ogilvy and Grey) cut close to 10,000 jobs in a single year. Its “Elevate28” restructuring programme targets £500 million in annual savings by 2028.

Omnicom (another major global advertising group), following its $13 billion acquisition of IPG (Interpublic Group, also one of the world’s largest advertising companies), announced more than 4,000 job cuts and doubled its cost-saving target to $1.5 billion over 30 months. Iconic agency names were retired.

Nike built Nike Icon Studios in Culver City, Los Angeles: a single facility consolidating every global brand pre-production and post-production function under one roof. Not an external agency. A production house owned by Nike.

For a mid-sized brand, the equivalent is a small internal team running ElevenLabs Flows, Midjourney, Runway, and Canva AI, producing monthly content volumes that would previously have required a full-service agency retainer.

Every brand is a production house now. The question is what they are producing.

When Everyone Has the Same Pipeline

What happens when everyone has the same pipeline is not obvious until the feeds fill up.

McDonald’s Netherlands released its 2025 Christmas advertisement using generative video tools. It was pulled within days. Viewers described it as “AI slop” , the term internet users coined for AI-generated content characterised by warped movement, uncanny expressions, and a synthetic quality that is difficult to name but impossible to ignore. Comments said it “ruined Christmas spirits.” Tom Williams, global creative director at Incubeta (a digital growth agency): “It seemed to lack intent. The whole thing gave off Uncanny Valley vibes.” Valentino (the luxury fashion house), Skechers, Colgate, Samsung, and Burger King received similar audience backlash for AI imagery during the same period.

iHeartMedia (one of the largest radio and digital audio companies in the United States) launched a “guaranteed human” tagline after its own research found 90 per cent of listeners wanted their media made by people, not machines.

The Journal of Business Research (a peer-reviewed academic journal) confirmed in 2025: when consumers believe marketing communications are AI-written, they judge the content as less authentic, feel moral disgust, and show weaker engagement and weaker purchase intentions, even when the content is otherwise identical to human-made equivalents.

Europol (the European Union’s law enforcement agency) projected that 90 per cent of online content may be synthetically generated by 2026. Not bad content. Just content. Indistinguishable in volume, similar in texture, arriving from every brand simultaneously.

AI is a homogeneity engine. It produces what is statistically most likely. When every marketing team uses the same tools trained on the same data, output converges. The feeds fill up.

Having the pipeline is not the competitive advantage. It is the floor.

There is a second problem most brands running this pipeline have not yet found. It is quieter than the slop problem. It is, in some ways, more dangerous.

You Own Nothing

Most brands running an AI pipeline have not read the licence clause that defines who owns what the AI creates and what legal rights the company keeps.

In January 2025, the United States Copyright Office published the second part of its Artificial Intelligence and Copyright Report. The finding: writing a prompt into a generative AI tool does not make you the author of the output. Prompts are instructions, not expressions of creativity. The Supreme Court declined to hear the appeal in Thaler v. Perlmutter in 2026, leaving the lower court ruling in place. The rule stands: only human beings can hold copyright authorship in the United States.

This has a direct and serious consequence for every brand running an AI creative pipeline.

If a marketing team generates a campaign image in Midjourney, a video in Runway, and a voiceover in ElevenLabs , and a human simply reviewed and approved the output without substantially editing or modifying it those assets are not copyrightable. They enter the public domain. A competitor can use them. A stock agency can sell them. Anyone can copy them freely, and legally.

The tools themselves offer very different levels of protection.

Midjourney tells paying subscribers they “own” what they create commercially. But intellectual property lawyers have identified the gap: if the law does not recognise AI-generated output as copyrightable, Midjourney cannot grant a copyright it does not have. What it is actually offering is a promise not to assert its own rights against the user. There is no legal protection standing between a Midjourney-generated brand asset and a copyright claim.

Adobe Firefly (Adobe’s AI creative toolset, part of Creative Cloud) is the meaningful exception. Its training data was sourced from Adobe Stock (a library of commercially licensed images), alongside public domain content. Adobe explicitly guarantees commercial safety on Firefly outputs , and includes IP protection, meaning if someone sues you claiming your Firefly output infringes their copyright, Adobe will cover your legal defence costs up to $10,000 per claim for paid subscribers. It is the only major AI image tool with contractual legal protection built in.

Suno (the AI music generation platform) technically retains authorship even for paid subscribers. Subscribers receive a commercial use licence , not ownership. The music in last quarter’s campaign belongs, in a legal sense, to Suno.

The practical exposure is real. Brands are building creative libraries , social assets, campaign images, advertising videos, product photography, using AI tools without documenting the human creative decisions that would establish copyright. Those libraries may contain no legally protected intellectual property at all.

Reinvent IP (an intellectual property law advisory firm), in a March 2026 briefing for agencies, identified this directly: “the content created with AI tools might not actually belong to the organisation.”

The remedy is not to stop using AI. It is to document the process. The Copyright Office’s guidance: copyright protection applies where a human makes specific, substantial creative choices, editing, selecting, arranging, modifying in ways that express genuine creative judgment. Brands need to build what practitioners now call a Creative Audit Trail: a documented record of the human decisions that converted AI-generated output into a protected brand asset. Which elements were generated. Which were edited or restructured by a human. What creative choices were made during that process.

Disney, Universal, and DreamWorks filed a federal lawsuit against Midjourney in June 2025, alleging the platform functions as a “virtual vending machine, generating endless unauthorised copies” of copyrighted characters. Midjourney filed a fair use defence, arguing that training an AI on existing images is comparable to a human artist learning from existing work, not directly copying it. The case is ongoing. Its outcome will shape what brands are permitted to build with AI creative tools, and what they legally own when they do.

Any brand building significant creative output with AI tools should be having this conversation with legal counsel now, not next quarter.

The legal question and the competitive question are not separate. Both reduce to the same thing: whether any human being in the building made a real decision about what the brand is saying. A Creative Audit Trail protects the asset legally. A clear brief protects it commercially. The work is the same work.

The Taste Gap

The pipeline is shared. The legal exposure is shared. The tools are identical. What is not shared is the judgment about what to make.

When production costs approach zero and the tools are available to every competitor, the resource that becomes scarce is the one the tools cannot supply: judgment about what to make, for whom, and why.

The VC Corner (a venture capital and technology newsletter), in an April 2026 analysis of what differentiates companies in the AI era, put it plainly: “When everyone has the same jet engine, speed is no longer a moat.” Rex Woodbury (a technology analyst and writer) called the current moment the “Costco era of software”, mass-produced, instantly forgettable output that looks like everything else because it came from the same engine as everything else.

Production skill still matters. The brands getting the best results from Runway and Veo 3 are not the ones who opened the tool for the first time this week. There is real expertise in building consistent characters, writing prompts that hold brand voice, and understanding which model does what. That expertise has value. But it has a ceiling. Because the expertise is learnable, and the tools are available to anyone willing to learn them. Production mastery is not the final differentiator; it is a cost of entry.

The brands solving this are not using different tools. They are making different decisions before they open the tools.

Heinz (the food brand known globally for its ketchup) ran a campaign using DALL-E 2 (an AI image generator from OpenAI) prompted with phrases like “ketchup in outer space.” Then it invited consumers to generate and submit their own AI ketchup artwork, displayed in a digital gallery and on physical advertising such as billboards and posters. The campaign generated over two billion impressions and 25,000 personalised videos. Heinz did not use AI to make an advertisement. It used AI to make an invitation. The tool was the same as every other brand’s. The judgment about what to do with the tool was not.

Coca-Cola’s Islam ElDessouky, global vice president for creative strategy and content, described the internal reality after the 2025 AI holiday campaign: “The masses, the audiences, do not necessarily look behind the technology. They just look at the story that they’re receiving.” By Coca-Cola’s own testing metrics, the AI-generated advertisement was one of the best-performing in the brand’s history. The online backlash came from people who noticed how it was made. General audiences responded to what it said.

The gap is not between brands with AI and brands without. It is between brands that know what they are saying and brands that are simply producing.

What taste actually is, how it is built, and what it means for creative strategy gets its full treatment in a separate piece. Here, the argument is narrower: direction precedes production. The brief is the competitive asset. The tools are not.

The Brief Is the Product

The diagnosis is clear. What the new job looks like is different from the old one in every layer that matters.

The brief , the specific human argument about what a piece of content should make an audience feel, and why, is where the work is now. Not in execution. Execution is automated. The brief is where competitive advantage begins, where intellectual property can be established through documented human creative decisions, and where brand identity is either held or surrendered to the algorithm.

The New York Times, in a June 2025 analysis of creative work in the AI era, described the shift: writers becoming “article designers,” “story designers” emerging as a distinct role in film and television, “world designers” shaping marketing and gaming. The paper summarised: “More people will be tasked with making creative and taste decisions, steering the AI where they want it to go.”

That is the new job. Not making. Steering.

APR named the three roles a brand needs at the top of its AI creative operation. An AI Workflow Producer designs and maintains the pipeline, how the tools are connected, which workflow runs for which content type, and how governance works across the system. A Prompt Engineer embeds brand voice, audience context, and quality standards directly into the instructions the AI receives, turning the brief into a repeatable, brand-consistent specification. An AI Supervisor holds the standard across all output: the person who reads every finished asset against the brand before it publishes, and removes what does not meet it.

At the level of strategy, the question is: what does the brand stand for, and is the brief specific enough for an AI tool to reproduce that accurately? At the level of tactics: which workflow track is running (Craft, Maker, Content Engine), and does each have a human responsible for the brief? At the level of daily operations: is there a documented Creative Audit Trail for every asset produced, establishing the human creative decisions that make it protectable?

All three questions need answers before the tools are opened.

AI may be a multiplier. It is not a shortcut. A weak idea at scale is still a weak idea. The only thing a brand truly owns is its point of view , and the brief is how that point of view survives contact with the machines.

That shift from tool-access to creative judgment has a direct analogue in an earlier industry.

Music Production Got Cheap. Taste Won.

This has happened before.

In 1991, a Bay Area company called Digidesign launched Pro Tools: a digital audio workstation (software that replaced physical recording equipment, letting music be recorded, edited, and produced entirely on a computer) that destroyed the economics of professional recording studios within a decade. Artists including Billie Eilish and Clairo built careers recording in spare bedrooms. Production that required a £5,000-a-day studio could be done for the cost of a laptop and a software subscription.

What happened to the recording engineers?

The great ones became more valuable than ever. The ears that could distinguish between a bedroom recording that sounded genuinely professional and one that merely sounded decent were suddenly rare , because the tools no longer revealed the distinction automatically. The middle ground of competent-but-interchangeable engineers, who relied on expensive equipment to provide their edge, was hollowed out. The floor dropped. The ceiling rose.

A 2025 ScienceDirect (a major academic research publishing platform) analysis of technology’s impact on content production found the same pattern repeating across every creative technology wave: “each episode shares a common logic , the automation of a critical human task at scale, and the democratisation of that capability to the wider population.” The difference with AI is that it “simulates the expert labour itself,” not just enabling more people to participate, but potentially automating the creative functions that once defined professional advantage.

Pixar (the American animation studio behind Toy Story, Up, and WALL-E) has always had better rendering technology than most competitors. Pixar won because it never mistook the technology for the art. Its defining films were stories about longing, loss, and belonging. The technology served the story. The story came from people who had something to say.

The brands in the most exposed position right now are those in the middle, outsourcing creative direction to agencies whose primary value was production capacity, now watching those agencies restructure and contract. The brands have not yet built the internal creative direction capability to replace what those agencies provided.

The window to build it is now. Not when the tools are more advanced. Not when the legal framework is clearer. Now, while the advantage is still available to those who move.

Where to Start

On 31 October 2025, Amazon sent a cease-and-desist letter to Perplexity AI. The accusation was not that Perplexity had built a competing product. It was that Perplexity’s Comet browser was helping people use Amazon’s own product, in a way Amazon had not designed for and had not authorised.

Comet is a shopping agent. You tell it what you want. It navigates the web, finds the product, and completes the purchase on your behalf. The person using Comet gave it explicit permission to act for them. Amazon, the platform Comet visited, had not agreed to that arrangement.

On 9 March 2026, a federal judge in California granted Amazon a preliminary injunction. US District Judge Maxine Chesney wrote that Amazon had provided strong evidence that Comet had accessed its website “at the user’s direction, but without authorisation” from the platform itself.

The legal proceedings may run for years. The distinction Judge Chesney drew is not a legal technicality. It is the structural problem every product now faces, whether they have been to court about it or not.

That gap, between what the user authorised and what the platform was built for, is the interface problem of 2026.

The scale of that problem became visible in Q1 2026, when the traffic data arrived.

The Traffic You Did Not Send For

Adobe Analytics, which covers more than one trillion visits to US retail sites annually, measured AI-sourced traffic through the first quarter of 2026. It grew 393% year over year. During the 2025 holiday season, the growth had been 693%. The rate is not levelling off.

Salesforce’s analysis of the same period found that AI and agents drove 20% of all retail sales and influenced $262B of global online spend over November and December 2025. Autonomous agent actions increased by 142% over the same window.

The most striking figure is the conversion reversal. In March 2025, AI-sourced traffic converted 38% worse than human traffic. In March 2026, it converted 42% better. Revenue per visit was 37% higher. Time on site was 48% longer.

The product did not change. The agent got better at finding what the human actually wanted, and sending only the sessions most likely to convert.

Two Users, One Interface

The human who uses your product and the agent acting on their behalf want the same outcome: the right thing at the right price, quickly. The paths they take are entirely different.

The human navigates. They land on the homepage, scan the headline, watch the video, read the pricing comparison, hesitate, close the tab, come back. The interface earns their trust across several visits. The copy matters. The personality matters. The design matters.

The agent parses. It arrives at a URL and extracts: price, specifications, availability, a countable proof signal. It compares across three competitors in the time the human was reading the first paragraph. It executes when the criteria match.

The agent does not watch the video. It does not respond to brand warmth. It does not care about the scroll animation. What it needs is information that is structured and labelled (organised so software can find and extract specific values), and readable by a machine.

What the Machine Cannot Read

Adobe’s 2026 analysis of US retail sites found that the average homepage is 75% machine-readable (structured so software can understand and extract it). The average product page: 66%.

The 25 to 34% that is invisible to an agent is, in most cases, the differentiating content: the main promotional video at the top of the page, JavaScript-rendered pricing comparisons (content that only appears after the page loads fully in a browser, invisible to anything reading the underlying document), CSS star ratings, interactive demos, scroll-triggered copy. Every design pattern that builds trust with a human reader is, to an agent, a void.

The 34% invisible on a product page frequently contains the specific differentiators that separate the product from its competitors: the copy written to justify the price, the proof that addresses the main objection, the specification that wins the decision. An agent that cannot read those sections defaults to price alone.

Early analysis of Shopify merchant data by Presta in 2026 found that stores optimised for agentic discovery saw 28% higher conversion from AI-driven traffic compared to stores that had not made those changes. These results come from a single study; real-world variation by category is expected. But the direction is consistent with Adobe’s broader data: the structured surface converts better for agents and, in most cases, better for humans too.

What the Protocols Require

MCP: The Interaction Layer

The Model Context Protocol (MCP), published by Anthropic in November 2024, is the agreed standard that lets AI agents interact with products without scraping webpages. In plain terms: rather than guessing what the price is from the HTML, the agent has a direct, structured conversation with the product itself. Within ten months, OpenAI, Google, and Microsoft had adopted it. By March 2026, over 10,000 active MCP servers were running in production, with 97 million monthly downloads of the SDK. MCP is infrastructure, not experiment.

ACP: The Transactional Layer

The Agentic Commerce Protocol (ACP), co-developed by OpenAI and Stripe, published as an open standard in September 2025, is the checkout flow for agents. If MCP is the door, ACP is the checkout. It handles how an agent identifies itself to a merchant, how payment credentials are handled securely, and how a purchase completes without a human clicking through a single screen. Salesforce, Etsy, and major Shopify merchants have adopted it. US ChatGPT users can already complete purchases from Etsy sellers inside the chat interface, without visiting Etsy’s website. The specification is open-sourced on GitHub.

The Amazon case turns on the Computer Fraud and Abuse Act (the US federal law making it illegal to access a computer system without the owner’s permission). The distinction: the user’s authorisation versus the platform’s authorisation. The Agentic Commerce Protocol is precisely what is designed to resolve that gap before it reaches a court.

The legal picture is still settling. The commercial case for acting now is not.

One Fix, Two Better Users

The changes that make a product agent-readable also make it a better human product. Optimising for agents is not a concession to a secondary user at the expense of the primary one. The improvements compound.

Clear information hierarchy (price stated plainly, specifications structured, social proof as text with a countable number, a single unambiguous action) is what an agent needs to parse and act. It is also precisely what reduces friction for the human buyer. The page that gets out of the way of the decision converts better for both.

The agent’s requirements are a forcing function for the human-readable improvements most product teams should have made already. The plays that follow serve both users.

Three Plays

Serve the Second User on Your Terms

The Number in the Layoff Memo

On 15 April 2026, Snap fired about a thousand people. Sixteen per cent of the company. Gone in a memo.

The memo did something unusual. It told the truth.

Most layoff notes hide behind weather words. Headwinds. Realignment. A challenging macro environment. Snap’s chief executive skipped all of that and named the cause out loud: AI agents now generate more than 65% of the company’s new code, and small squads using AI tools can do the work that used to need larger engineering teams (CNBC, 2026).

Then the part nobody could have scripted. The stock went up. Around eight per cent.

Read that sequence again. A company says machines now write most of its software, says it therefore needs far fewer humans, and the market rewards the confession. A sceptic will say the stock went up because Snap has been burning cash for years and the market simply applauded any sign of discipline. That reading is not wrong. But it misses what was actually being priced: not the headcount reduction, but the named cause attached to it. The layoff was not the bad news. The layoff was the proof of concept. (

Worth noting: the roles Snap actually cut were in product and partnerships, not engineering directly. Spiegel named engineering efficiency as the reason; the redundancies landed elsewhere. That gap between stated cause and actual cuts is precisely why the sceptics in this piece deserve a hearing.

The question the memo raises is the right one: if AI writes the code, what does the engineer write?

To answer it properly, it helps to look at what AI can actually do for an engineer right now, in 2026.

That is the entire story. Everything else in this piece is evidence for it, or implication from it.

The Deliverable Moves Up the Stack

The code was never the valuable part, not for most teams. The code was the toll you paid to find out whether your judgement was right. Now the toll is cheap.

This is not the first time a craft's tooling got cheap. It happened to photography when digital arrived; to music production when software replaced the studio; to design when templates made layout trivial. Each time, the commodity layer collapsed and the layer above it, the judgement, the eye, the taste, became the thing that separated people who were genuinely good at it. Software is the latest version of that pattern.

For thirty years, the scarce and defining act of building software was the writing of it. The code was the work. You hired for it, you queued for it, you protected the people who could do it.

That bottleneck is dissolving. Google says 75% of its new code is now AI-generated and approved by engineers; Microsoft puts its figure at 20 to 30%; Snap is past 65 (Google; CNBC, 2025-2026). Roughly nine in ten developers now reach for an AI tool as a matter of course. The act that used to be the job is becoming the cheap part of the job.

Most teams are not at 65 or 75 per cent yet. But the direction is not in question, and preparing now, before the transition arrives at your organisation, is the entire point.

So where did the value go?

Most people get this wrong in the same direction. They assume that when the cost of code collapses, the value collapses with it. The opposite is happening. The value did not vanish. It moved, up the stack, to the things the model cannot do for you: deciding what to build, knowing whether it is any good, specifying it precisely enough that an agent can execute it, and getting it in front of the right people.

What is left exposed is the engineering. Items eleven through fifteen, the understanding, the decisions, the reliable systems, the communication with humans, were always the job. Now they are the only part of the job that compounds in value as the tools get cheaper.

IBM CEO Arvind Krishna has described what he calls the $8 trillion math problem: reaching 100 gigawatts of global AI compute, at roughly $80 billion per gigawatt, requires $8 trillion of capital expenditure, committed entirely to making items 1 through 10 faster and cheaper. Not one watt of those data centres makes items 11 through 15 easier to do (Futurism / IBM, 2025).

The return data confirms the gap. Goldman Sachs calculates that sustaining current investor return expectations requires AI companies to generate more than $1 trillion in annual profit from their AI infrastructure. The 2026 consensus estimate: around $450 billion, roughly $0.45 back for every $1 of capital expenditure (capex) (Goldman Sachs, 2024-2026). The report named the problem plainly: “too much spend, too little benefit.” The DORA 2026 developer research programme found the missing variable: strong engineering foundations are what actually drive AI return on investment. The organisations seeing real returns are the ones with mature practices in the things AI cannot do, requirements, decisions, reliable systems.

Cheap to Write Is Not Cheap to Own

Those numbers describe where the value actually lives , and it is not in the production layer. Cheap to write is not the same as cheap to own. This is not a theoretical concern, the evidence is in, and it is not flattering. The cost of code did not disappear when the typing got fast. It relocated, downstream, to review, to maintenance, to the slow tax of running software nobody on the team fully understands.

CodeRabbit (a code review analytics platform) found that AI-co-authored pull requests, bundled sets of code changes submitted for review, carried about 1.7 times more issues than human-only code, with security problems up to 2.7 times worse (CodeRabbit, 2025). The code arrives faster, and arrives carrying more of the kind of problem you do not see until later.

Then the part that should unsettle you. A controlled trial, sixteen experienced developers, 246 tasks, randomised assignment, put developers on code they knew well. With AI, they were about 19% slower (METR, an AI evaluation safety research organisation, 2025). Not faster. Slower. And they believed they were faster the whole time. The tool did not just cost them time. It cost them the ability to notice they were losing it, which is a different kind of problem entirely.

A large industry study found the pattern underneath: AI raises throughput (the rate at which code ships) and worsens delivery stability (DORA, the annual industry developer research report, 2025). It does not fix your team. It amplifies whatever your team already is. Disciplined shops get a multiplier. Sloppy ones get a faster way to ship the mess. There is a name worth keeping for the bill that comes due here: comprehension debt, the accumulating cost of shipping code nobody fully understands. You take it on quietly, at speed. You repay it all at once, in production, on the worst possible day.

The Apprenticeship Needs a New Curriculum

The cost moved. The work moved to higher ground. What about the people?

Stanford found that employment for early-career developers, the ones aged 22 to 25, is down about 20% since AI went standard (Stanford, 2025). Not redistributed. Down. The roles that contracted fastest were the ones closest to items 1 through 10, take a clear ticket, write the obvious implementation, hand it back. That is exactly what an agent now does for nothing.

But look at what is expanding. The roles companies cannot fill fast enough are the ones requiring items 11 through 15: the technical lead who can run a requirements session, the senior engineer who can debug production without a map, the architect who makes good decisions under real ambiguity.

The apprenticeship is not being automated away. It is being rebuilt around a different curriculum. The path forward for early-career engineers is not to compete with AI on items 1 through 10. It is to use those tools to compress the routine work and invest the hours saved into items 11 through 15. The engineers who do that now, before the transition arrives everywhere, are not behind the curve. They are the curve.

Three Roles Worth Understanding Now

Technology Changed. Engineering Did Not.

When Effort Stopped Signalling Quality

Slide nine of Adam Mosseri’s New Year’s memo is easy to miss. He is not, on the surface, talking about Instagram. He is talking about cameras.

“The camera companies are betting on the wrong aesthetic. They’re competing to make everyone look like a pro photographer from 2015. But in a world where AI can generate flawless imagery, the professional look becomes the tell.”

For twenty years, the logic of creative credibility ran on a single premise: visual quality was expensive. A polished carousel took hours of real work. A well-lit photograph required equipment, training, or both. A written piece with genuine rhythm required someone who knew what they were doing. The polish itself was the proof. If something looked considered, it was because someone had considered it.

AI dissolved that logic in approximately eighteen months.

It is not simply that anyone can now generate something that looks like effort. It is that the effort signal itself, the cue audiences used to calibrate trust, has been severed from the product. The carousel that took four hours of genuine thought looks identical to the one assembled in eleven seconds. The written piece with real rhythm is, in form, indistinguishable from the piece a machine produced by pattern-matching against a million others.

When a signal breaks, the market searches for a new one. Here is what makes this moment unusual: the new signal is something most working creative people have been building for years without knowing it had a name.

Mosseri named it in slide twelve. After explaining that AI will soon replicate even the raw, imperfect aesthetic that some creators are leaning into as proof of humanity, he wrote: “At that point we’ll need to shift our focus to who says something instead of what is being said.”

This is not a platform strategy note. It is a description of what authenticity has always been, once everything else it used to cosplay as has been stripped away.

What Taste Actually Is

Start with what it is not.

Taste is not aesthetic preference. It is not having opinions about fonts. It is not knowing what looks good. All of those things can be faked, borrowed, or generated from a reference library. They are outputs of taste, not taste itself.

Taste is the operating system that runs at every decision point in a creative process. It evaluates options not on the question “is this competent?” but on the question “is this right for this?” The two questions look similar. They are not.

AI is very good at “could.” Ask it for options and it will produce fifty of them, all technically competent. What it cannot do is tell you which of those fifty is right for this audience, this moment, this brand, this specific argument. That knowing, which is not reasoning from first principles but feels like recognising something you have seen before in a different form, is taste.

The radio producer Ira Glass (host of the long-running American public radio programme This American Life) described the beginner’s version of this gap in a 2009 interview that has circulated among creative professionals ever since. Beginners get into creative work, he said, because their taste is already strong. The problem is that their taste exceeds their ability. They can sense the work is wrong but cannot fix it yet. The gap closes slowly, through volume.

AI inverts this exactly. The machine’s output capability now exceeds most people’s capacity to judge it. The new gap is not “can I execute?” It is “can I sense which execution is right?”

That is good news. The sensing, built slowly through years of having opinions, defending them, being wrong about them, and having them again, is what no tool generates from a prompt. You have been building it for longer than you think.

Every Democratisation Is a Taste Test

This has happened before. It just tends to take longer.

In 1985, PageMaker and the Apple LaserWriter put a print shop inside every office. Typesetting, which had required training, equipment, and the quiet expertise of people who understood leading and kerning and when not to mix typefaces, was suddenly available to anyone with a Macintosh and an afternoon. The first generation of desktop-published newsletters used approximately every font available simultaneously. By 1989, offices were drowning in justified text, clip art, and drop shadows rendered at 72 dots per inch.

The typographers and designers who survived that flood were not the ones who learned PageMaker fastest. They were the ones with enough taste to know which of its capabilities to leave unused.

They did not just survive. They became more valuable.

In 2010, the iPhone 4 camera marked the moment when photography stopped being a technical bottleneck for most people. The image-making floor rose sharply. Professional photographers were widely declared redundant. What actually happened was that photographers who had built genuine visual judgment over years, who could see something worth capturing before lifting the camera, became more valuable, not less.

The technical floor rose. The taste ceiling rose further.

The pattern holds across every access event: the work that survives is the work that was never just about technical access. Desktop publishing, affordable studio equipment, the smartphone camera, and now generative AI, each one raises the floor. Each one makes the question of what is worth making more important than the question of whether you can make it.

The difference this time is speed and scope. AI has raised the floor not on one skill but on dozens simultaneously: writing, design, music, code, photography, video, strategy documents, analysis. The taste test is happening across every discipline at once. And the people who had been quietly building judgment in their field are, right now, sitting on the most valuable professional asset in the room.

The Three New Signals

After the old signal breaks, what does the market use instead?

Mosseri named three in his memo, though not in those terms. Specificity. Voice. The visible presence of a real person’s perspective.

Specificity is the observation that only someone in that room could make. Not “AI is changing creative work” but “on the Thursday morning after our studio read the Mosseri memo, we retired our standard carousel template for the first time in two years.” The machine can approximate general. It cannot approximate yours. It does not know what your Thursday morning felt like, what specific decision preceded this one, what two things were rejected before you landed here. Specific detail is a proof of origin. It costs nothing to add and is impossible to replicate.

Voice is not consistent tone across a brand. It is the grain of how one person processes a problem: the rhythm, the things they notice before anyone else notices them, the comparisons they reach for that no one else would reach for. Voice takes years to develop because it is the residue of thousands of choices made with conviction. You cannot synthesise residue.

The visible presence is sometimes the smallest thing. The Instagram educator Rishi Shine (creator of the widely followed @therishishine carousel format) signs every cover with “by Rishi” in handwriting. One handwritten line in an otherwise designed system. The machine perfection breaks exactly once, and the reader’s nervous system registers it: a person made this.

Not as a disclaimer. As a presence.

This logic extends far beyond Instagram. The same signal applies to written pieces, product decisions, design systems, and the work of anyone who communicates publicly. The cost of the signal is approximately zero. A script font signature on a cover. The admission in a piece of writing that you changed your mind mid-argument. The one specific detail no one else would have included.

For Patient Comet, the equivalent of “by Rishi” is “by Nadim” in a script font on every cover. Not because it looks good. Because it is the one element in an otherwise pixel-perfect system that a machine, left to its own devices, would never think to add.

Taste Gets Stronger With Every Decision

What makes taste unusual as an asset is that it appreciates through use rather than depreciating.

Every creative choice you make , which headline, which case study, where exactly to end a paragraph builds a judgment model that becomes the filter for the next choice. The reps are not visible on the outside. The accumulation is.

This is the most hopeful part of the picture. Every creative professional who has ever stopped at a bad layout and asked themselves why it is wrong, everyone who has ever rewritten a headline three times to find the sharper version, everyone who has ever cut a paragraph they loved because it was not serving the argument, has been doing compound-interest work without calling it that.

This is why experienced creative directors are not faster processors of options. They are more refined refusers of options. The ability to scan forty competent layouts and immediately identify the three worth discussing is not speed. It is a dense, compressed library of pattern-and-reject, assembled through sustained, opinionated exposure to work that was trying and mostly failing to be excellent.

In an AI-accelerated workflow, this matters more than it ever has. When a tool generates forty competent options in the time it once took to produce one, the bottleneck shifts entirely to the human doing the selecting. The judgment has to be faster, sharper, and more accurate than the tool is prolific. The person with a compounding taste library just became the most valuable person in that room.

The Cannes Lions international festival of creativity (the industry’s benchmark awards event, running since 1954) made this concrete in 2026. For the first time in the festival’s history, all submissions are subject to integrity standards that evaluate provenance and human authorship alongside craft. What Cannes is acknowledging, in its characteristically deliberate institutional way, is that the creative industry has moved from valuing who can make it to valuing who can judge it.

The question, then, is not whether to build taste. It is how.

How It Gets Built

Taste is not received. It is built slowly, through a specific kind of practice.

The mechanism is not consuming more. It is having more opinions about what you consume. The designer who can identify a wrong spacing decision from across a room developed that ability not by reading about spacing but by stopping at bad spacing, naming what is wrong, articulating exactly why, and filing the specific quality of its wrongness somewhere she will access automatically for the next twenty years.

Three practices matter more than any tool or tutorial.

The first is cross-domain exposure. The sharpest taste in any field tends to be built on fluency in adjacent ones. The designer who reads novels. The strategist who studies architecture. The engineer who has spent real time with music. Not because the domains share technique but because they share the underlying logic of what it means to decide what to include and what to leave out. Paul Rand (the American graphic designer, creator of the IBM and ABC logos) read widely outside graphic design. Dieter Rams (the German industrial designer whose work became the direct inspiration for Apple’s design language) drew from music and architecture as reference disciplines. Taste built in one room tends to come from inputs gathered in several.

The second is defended positions. You do not have taste unless you have opinions you will argue for. The ability to say “this is right and this is wrong,” precisely, in the face of technically competent alternatives, and with a clear account of why, is the practice that sharpens taste into something you can actually use when fifty options are sitting in front of you.

The third is the mistake archive. Every call you got wrong is tuition paid. The headline that underperformed. The design direction you championed that confused the audience. The feature you shipped that nobody used. Taste only compounds if you stay in the room after the wrong call and understand specifically what you missed. Not “it didn’t land” but “I misjudged this because I prioritised x over y.” The specificity is the compound interest.

Who Becomes Most Valuable Now

The work that AI cannot do well is not creative work in general. It is the specific work of choosing.

A creative tool can produce a hundred options. It cannot tell you which of those options is right for this brand, this audience, this specific tension in the market at this specific moment. That knowing lives in accumulated human judgment, and it does not transfer to a prompt.

This changes what you should be building in yourself and in the people around you. In a pre-AI talent model, the expensive thing to develop was technical skill: how to produce something at the required standard, how to operate the tools, how to meet the spec. In the model we are now in, the expensive thing to develop is judgment. How to survey a space of options and know, immediately and defensibly, which three are worth discussing. How to say “not that” in a way that moves the work forward. How to recognise the version that is almost right and to fix it, rather than accepting the version that is merely technically complete.

This also changes who you want in the room. The person who produces the most options fastest is not the most valuable person at this stage. The person who can look at fifty options and immediately identify which three deserve attention, and can explain why in terms that others can use, is worth considerably more than they were five years ago.

Mosseri ended his memo with this: “In a world of infinite abundance and infinite doubt, the creators who can maintain trust and signal authenticity, by being real, transparent, and consistent, will stand out.”

Strip out the platform-speak. What he is describing is judgment: applied publicly, maintained over time, and expressed with enough specificity to be recognisably yours.

The Asset You Already Have