Last week a CIO I have known for fifteen years sent me a screenshot on WhatsApp. A receipt: one developer, one weekend, 5,300 € in coding-agent tokens. No comment: he just sent the screenshot. I sure as hell knew what the comment would have been: language that is frowned upon in most of the civilized world.
Cory Doctorow has a word for what is about to happen to enterprise AI, and the word is very ugly on purpose. Enshittification (so sorry: it’s a brainworm… it will stay glued in your neurological system for a while). His 2025 book carries it as a title (Enshittification: Why Everything Suddenly Got Worse and What to Do About It), although the word had already done its damage in the public conversation long before the book arrived, waddling into circulation with all the grace of a drunken swearword that had been waiting years for its proper historical moment. Doctorow uses it to describe the decay of digital platforms. First they are useful to the user, then they become indispensable, then the value gets quietly rerouted toward the platform owner… once everyone is locked in and switching costs are a wall.
That is the playbook. Sounds familiar, no? AI is (obviously) the next chapter. It began as magic (O Arthur C Clarke!). Then habit, then infrastructure. Once it sits inside writing, coding, research, sales, service, reporting, legal review, executive decision support and the small rituals of corporate life, the pricing stops behaving like software. It starts behaving like a wildfire craving oxygen, with a smiling token-meter attached.
SaaS was the rehearsal, and CIOs still have the scars
Remember the early SaaS pitch? No servers, no upgrades, no chilled basement humming like an angry Soviet submarine. Just clean subscriptions, automatic updates, predictable cost, and the faint possibility that finance might finally stop treating IT as a cash-fuelled bonfire with login credentials.
Then reality arrived with a grim and unapologetic renewal notice. Per-seat pricing multiplied. Storage moved into premium tiers. APIs were pushed behind enterprise plans. Security features became upsells, which remains one of the most magnificent little insults in modern software (rather like buying a car and discovering the brakes require an add-on). Admin controls cost more, audit logs cost more, SSO cost more. Data export cost more, integrations cost more, support cost more… and most of that was in no way initially budgeted for. The Bessemer cloud market index EMCLOUD grew 524% between 2013 and 2023, outpacing the NASDAQ, the S&P 500 and the Dow Jones combined. Per-employee SaaS spend hit $8,700 in 2025, up 27% in two years. SaaS price inflation is running at roughly five times general market inflation. Nobody signed off on that line item. Everybody is paying it.
Departments bought their own tools. Marketing had a stack. Sales had a stack. HR had a stack. Finance had a stack. Legal found something with a dashboard and a reassuring font. Every tool promised productivity. Every tool created data gravity. Every vendor had a roadmap. Every roadmap eventually discovered an enterprise tier with exactly the features governance should have had on day one. CIOs survived that first heart attack because CIOs are a hardy species. Like cockroaches, but with architecture diagrams and a recurring sense of moral exhaustion.
AI arrived, and everyone forgot the trauma.
Phase one: indecently generous
The first AI moment felt like a gift. Free access, magical demos, a chatbot that could summarize board papers, write Python, explain semiconductor geopolitics, draft a resignation letter, produce ten campaign angles and politely apologize for hallucinating like a junior consultant in a windowless room. For free. People tried it for fun, then for work, then for the work they did not want to admit they were giving it.
Phase two looked harmless. Twenty dollars a month. A rounding error. One mediocre airport hamburger. Suddenly every consultant, strategist, developer, analyst, lawyer, marketer and ambitious intern had a personal cognitive exoskeleton for the price of a Spotify subscription with delusions of grandeur.
Then the ladder started climbing. The market is already normalizing personal AI subscriptions at $100 and $200 a month. Anthropic offers Claude Max plans at $100 and $200. Google launched AI Ultra in the United States at $249.99 per month. Perplexity launched its Max tier at $200 per month in July 2025, joining the club. OpenAI’s business pricing combines seats and credits and usage logic across products, which is where the old SaaS pain starts learning new tricks.
Now the beautiful little monster enters our nightmare: credits and tokens.
The shift from seats to weather systems
Seats were painful, political and often absurd, but seats were easy. A user had a license. A department had a count. Procurement could argue. Finance could model. Legal could complain. The CIO could pretend the whole thing was under control until someone discovered 700 inactive licenses still billing quietly in the weeds.
AI does not stop at seats. AI turns a user into a weather system of variable cost. Every prompt, model choice, document upload, context window, reasoning run, coding session, research task, image generation, API call, agent workflow, retrieval step and automated action becomes part of the consumption layer. The bill does not follow headcount. It follows behavior. It follows ambition. It follows curiosity. It follows panic. It follows the consultant who has discovered that asking for “twenty more scenarios” is easier than thinking carefully about the first three.
That is where the second heart attack begins. Because the better AI gets, the more people use it. The more they use it, the more work they move into it. The more work moves in, the more they want larger context, better models, richer integrations, longer sessions, higher rate limits, agentic workflows and premium reasoning. The $20 miracle becomes a starter kit. The $200 plan becomes normal for “power users.” The $300 SuperProMaxWhatever plan arrives with a launch video, an elegant gradient, and a pricing page that looks calm enough to be legally suspicious.
The tech bros will not help. Some of them are already warming up the sermon. A good consultant, they suggest with the sacred confidence of men who have never had to defend a budget in front of a CFO with acid reflux, should burn through tokens as proof of seriousness. Low AI spend becomes a moral weakness. Cheap prompts become intellectual underinvestment. If your best people are not incinerating compute at a rate normally associated with aluminium smelting, are they even augmented?
You can see the keynote slide forming. “The $250,000 token-a-year AI-native consultant.” White background. Gradient orb. One sentence about scale (theirs, not yours). A founder in black trainers explaining that talent is now measured in token velocity. Behind the circus sits a real problem, and it is not about tokens.
The usage layer is a different animal
AI cost behaves like a hybrid of software, cloud, labor augmentation and financial cocaine. The fixed costs are familiar territory. Seats, admin controls, security, compliance, enterprise workspaces, connectors, retention policies, audit tooling. Procurement can wrestle with that beast. It has wrestled worse. It has stared into Oracle and Microsoft contracts and walked out with only minor facial twitching.
The usage layer is the new wild animal. A junior analyst discovers deep research and does in one afternoon what used to take three days, which is wonderful until everyone discovers deep research. A developer lets coding agents run overnight, which may be brilliant until the token bill resembles a ransom note (see screenshot, page one). A proposal team generates ten strategic routes instead of two. A legal team feeds every contract into the machine. A marketing team creates forty variants because marginal cost feels invisible. A manager asks for “just one more simulation” until the compute meter starts glowing like Chernobyl tourism.
Usage-based pricing always begins as fairness. Pay for what you use. Lovely, democratic, almost Scandinavian. Then workflows get built around it. Then the best features move upmarket. Then “unlimited” grows an asterisk, a footnote, a fair-use policy, and the emotional warmth of a parking fine. Then credits arrive, and credits are where pricing goes to put on a fake moustache.
Nobody experiences credits as money. That is their entire genius. They create just enough abstraction to make pain feel technical. Tokens, compute units, reasoning credits, premium requests, agent runs, video seconds, context boosts, secure retrieval calls, workflow actions. The vocabulary becomes a fog machine around the invoice.
The CFO will ask a simple question: “How much does this cost?” The honest answer will be: “That depends how intelligently the company behaves.” Good luck putting that in a budget cycle. AI-FinOps anyone?
Why AI is tailor-made for Doctorow’s pattern
Enshittification is a story about dependency followed by extraction. First the platform gives value. Then it becomes part of life. Then switching becomes painful. Then the platform starts reclaiming the value, because it can.
AI is the cleanest fit for that pattern the industry has ever produced, because employees do not experience AI as software. They experience it as personal capability. Take away someone’s project management tool and they complain. Take away someone’s AI assistant after they have learned to think with it, draft with it, code with it, research with it, argue with it and use it as a second brain with fewer coffee stains, and they look at you as if you removed part of their skull. The lock-in becomes cognitive before it becomes contractual.
That is the nasty part. By the time procurement notices the bill, the organization has already changed shape around the tool. People build private prompt libraries. They build workflows. They build half-secret methods. They trust a model’s rhythm. They learn its strengths, forgive its lies, and route more of their work through it. Teams start relying on AI-assisted throughput. Managers start expecting the new speed. Clients start receiving work that quietly assumes AI acceleration. The board hears productivity stories. The vendor sees dependency forming, as vendors tend to do, with the tender affection of a crocodile watching a river crossing.
The platform does not need to own your files if it owns the layer through which your people interact with those files. It does not need to own your processes if it owns the way those processes are interpreted, accelerated, summarized and partially automated. It does not need to own your employees if it becomes the place where their work begins.
Very profitable. Very quiet. The first act is magic, the second act is habit. The third act is extraction with a dashboard.
The answer is doctrine, not abstinence
The answer is not to ban AI and become the Department of No, guarding the past with a lanyard and a risk register. That would be pathetic. AI is useful. Sometimes brutally useful. In the hands of a good professional, a strong model compresses days into hours, improves thinking, exposes blind spots, automates sludge, accelerates delivery and makes knowledge work feel less like drowning in wet cardboard.
The answer is also not an unlimited AI buffet for every department because the CEO read a breathless article about agents on a flight to Singapore. The answer is doctrine. A company needs to decide what kind of AI consumption it wants, before the invoice decides on the company’s behalf. Which work deserves premium models? Which work can run on cheaper ones? Which tasks require deterministic automation rather than generative fireworks? Which data can go where? Which agents may act, which may suggest, which should stay in a sandbox with crayons and adult supervision? Which outputs need human review? Which usage should be logged, capped, benchmarked, challenged, or killed with fire?
CIOs need to bring FinOps discipline into AI before AI spend becomes the new cloud bill with a poetry degree. Value has to be measured with unpleasant honesty. Paying $200 a month for a brilliant employee who genuinely saves ten hours of high-value work is a bargain. Paying $20 a month for 4,000 employees so they can produce longer emails, prettier mediocrity, and seventeen alternative meeting summaries is digital obesity with a productivity badge.
Do not measure prompts. Measure cycle time. Measure error reduction. Measure margin impact. Measure faster delivery. Measure fewer escalations. Measure customer response time. Measure legal review throughput. Measure software defects. Measure proposal win rates. Measure work improved in the real world, where the furniture has corners and nobody cares about your innovation theatre: if you fall or fail, you bleed. Measure f*cking real outcomes, or measure nothing.
The bait is competence
The danger is that AI enshittification will hide inside real value. Bad tools are easy to remove. Useful tools are much harder. AI will create productivity before it extracts margin. It will make people better before it makes them dependent. It will solve enough problems to earn the right to create new ones. That is the elegant brutality of the model.
Doctorow’s warning lands here, hard. Enshittification is value reallocated after lock-in. AI has better bait than social media ever had. The bait is competence. The boring things, done early, are the things that matter. Separate AI access from AI strategy. Giving everyone a tool is distribution, not transformation. Build an AI cost model that includes seats, tokens, credits, API calls, storage, connectors, agent runs, governance, security, training, and exit costs. Protect the right to leave: data portability, prompt portability, model plurality, open standards, clean logging, internal knowledge architecture, contract discipline.
The right of exit sounds boring until you no longer have one. The AI industry will keep selling magic, because magic has better margins than plumbing. The demos will sparkle. The agents will wave their tiny digital hands. The subscription tiers will climb. The credits will multiply. The tokens will become budget dust. The consultant class will discover that burning compute feels like strategy when the invoice lands somewhere else. Somewhere inside the building, a CIO will remember SaaS and feel a familiar tightness in the chest.
Ouch.
