The bill at the keyboard is the same lesson the boardroom learned from cloud, arriving faster and at a smaller altitude.
In April, Uber’s chief technology officer told The Information that the AI budget he had set for 2026 was already gone. Claude Code usage inside Uber’s engineering teams had outrun every projection the company had made. “The budget I thought I would need is blown away already,” Praveen Neppalli Naga said. The quote was widely covered by PYMNTS and Axios.
That was not an isolated story. A few weeks later, Anthropic completed a pricing shift that had begun the previous November, moving Claude Enterprise customers away from flat-rate seats with bundled tokens and toward usage-based billing at $20 per seat plus consumption. Industry analysts estimated the change could double or triple costs for heavy users.
The conversation around AI is shifting. Premium model tiers. Usage-based pricing. Rate limits tightening. The FinOps Foundation’s 2026 State of FinOps Report found that 98% of FinOps practices now manage AI spend, up from 31% in 2024. A discipline that took cloud the better part of a decade to surface is arriving in roughly two years for AI.
The bill is the symptom. The cause is older than the technology.
I have lived through four abstraction layers. Electronic engineering, where every byte and clock cycle mattered. C++, where memory was managed but performance still had to be earned. C#, where the runtime took over and older engineers muttered about waste. Cloud, where infrastructure became someone else’s problem until the invoices arrived. Each transition felt like freedom. Each one looked, to the previous generation, like laziness. And each one eventually developed a discipline that the previous generation did not need and could not have invented.
AI is the next layer. The bill is arriving. The discipline is not yet here.
The token cost paradox
The most striking thing about the AI cost conversation is that the unit price is falling while total spend is rising.
By any reasonable expectation, AI should be getting cheaper. Token prices have fallen sharply over the last two years, with some industry analyses putting the decline in the hundreds of times for particular model classes and workloads. But that is not what most organisations are experiencing. Enterprise AI spend has risen sharply over the same period, with published estimates pointing to substantial budget increases across 2024, 2025 and 2026.
The exact numbers vary by source, and many of the sharpest figures come from vendors selling cost-management tooling, so the precise multipliers should be read as illustrative rather than definitive. But the direction is consistent across independent reporting from Deloitte, Gartner, the FinOps Foundation, The Information, Axios, PYMNTS, and The Register. Unit price down. Total spend up. Faster than anyone modelled.
Deloitte’s January 2026 analysis put it plainly. AI is now the fastest-growing expense in corporate technology budgets, with some firms reporting that it consumes up to half of their IT spend.
The vendors call this the token cost illusion. Some commentators have gone further and started using labels such as Token Incremental Burn Syndrome, or TIBS, for the pattern where token-based billing rewards verbosity, retries, and engineering workarounds that increase token consumption. The label is glib, but the underlying mechanism is real.
Cheaper units do not automatically create cheaper systems. Cloud taught us that. Cheaper units create more usage, more ambition, more automation, more experimentation, more waste, and eventually a new governance discipline.
AI is moving through the same pattern. Only faster.
The iteration tax
The visible cost of AI is the per-token price. It is the number that turns up on the invoice and the number the vendors compete on.
The invisible cost is round trips.
A vague prompt produces an answer that nearly fits. You clarify. The next answer fits slightly better but misses a constraint you had not stated. You clarify again. The third response is the one you keep. You have just paid three times for one outcome, and you have spent the iteration time yourself.
This is not speculation. Academic work on prompt ambiguity has documented the mechanism directly. A 2024 paper on evolutionary prompt engineering for code noted that “if the developer’s original prompt is ambiguous, the LLM may repeatedly generate incorrect code until the user clarifies the instructions in subsequent prompts. Each clarification adds new token usage.” The same paper measured that some agentic approaches required an additional 234,000 tokens on average to find a correct implementation on benchmark tasks. For a single function.
The pattern compounds at scale. Multiple analyses point to the same shape, even when the exact multipliers differ. One audit of thirty engineering teams running agentic AI in production between March and May 2026 reported that agent workflows cost roughly 3.2x more than direct calls for a simple five-step loop, more than 30x at fifty steps, and over 100x at two hundred steps. The depth of a typical autonomous debugging session. That audit was published by a vendor selling cost-control tooling, so treat the precise figures with appropriate scepticism. But the direction is corroborated elsewhere. Iternal’s 2026 token usage guide puts simple agents at 5,000 to 15,000 tokens per task and complex multi-agent systems anywhere from 200,000 to over a million. Agentic coding workflows, they report, can reach 1 to 3.5 million tokens per task including retries.
The same audit described a single developer running an autonomous refactoring agent over a long weekend who generated $4,200 in API charges before anyone noticed. Anecdotal, vendor-published, but consistent with the broader pattern that the leading AI labs themselves are now responding to commercially. Anthropic separating programmatic agent usage from chat subscriptions from June 2026 is not just a pricing decision. It is a recognition that the iteration tax is real and that the old subscription model could not absorb it.
Multiply across an organisation. Across a thousand prompts a day. Across teams who have not yet noticed that the wrong answer was not the model’s fault. It was the unformed intent behind the question.
Why intent gets lost at the keyboard
Humans under-specify. We have always under-specified, because human communication runs on shared context that we never have to make explicit.
We tell a colleague “make it better” and they often do, because they know us, the brief, the audience, the deadline, the last three things we said about this piece of work. We say “you know what I mean” and they usually do.
Models do not. Every assumption you do not state is a coin flip you are paying for.
The skill most organisations have not yet named is the ability to articulate, before you press send, what you actually want. Not the topic. Not the format. The outcome. The audience. The constraints. The examples of good. The examples of bad. The thing you do not want even though you did not think to mention it.
There is published work that bears this out empirically, though the exact figures vary. One commonly cited analysis from a prompt-optimisation vendor reported that across a sample of tasks, standard prompting cost roughly $0.13 per case with around a 33% success rate, while more elaborate techniques cost five to six times more but reached 74% success. The detail to notice is not the absolute numbers. Those depend on model, task, and methodology, and the source has a commercial interest in the conclusion. The useful point is the shape. Imprecise prompting forces you up the cost curve to compensate. You pay for technique what you did not invest in articulation.
This is not prompt engineering. Prompt engineering is the technical craft of working with the model. This is something more fundamental and more transferable. It is knowing your own intent clearly enough to describe it to something that cannot read your mind.
The discipline moves up the stack
Assembly’s discipline was resource awareness. Every byte counted because there were not many bytes. C++’s discipline was memory hygiene. The bytes were cheap but the lifetimes were yours to manage. C#’s discipline was architectural. The runtime handled memory, so the questions moved to structure and design. Cloud’s discipline was architectural intent at scale. Knowing what you needed before you provisioned it, because the meter was running.
Each era’s discipline sits one layer of abstraction higher than the one before. The discipline keeps climbing toward the human.
AI is the next step up. The discipline is communicative intent. Knowing what you want clearly enough to articulate it, because the meter is running and the model cannot guess what you meant.
This is not coincidence. It is the direction abstraction always travels. The machine handles more, the human handles less of the mechanism and more of the meaning. The questions that used to be technical become questions about purpose. The skill that used to be optional becomes the skill the era is built on.
Where this sits in the architecture
At the organisational level, this is what Human Intent governs. The redesign of structures around the flow of intent rather than the coordination of human work.
At the practitioner level, it is what Intent-Driven Development codifies. The methodologies, the specifications, the fidelity measures, the gates.
This piece sits between them. It is about the individual articulation skill that both depend on. Without it, the frameworks have nothing to capture and the organisations have nothing to govern. Intent has to exist before it can be specified, and it has to be specified before it can be measured, and it has to be measured before it can be promoted.
The bill in your inbox is the cheapest signal you will ever get that this discipline is missing.
We are not lazy. We are early.
It is tempting to read AI overspend as carelessness. It is not. Most people prompting models today were never trained in precision communication with machines. They are writers, analysts, founders, marketers, operators. They are doing what every generation has done at the start of a new abstraction layer. Using the new freedom before the new discipline has been named.
The cloud-native generation did not invent FinOps because they were more thoughtful than the on-prem generation. They invented it because the bills forced them to. The discipline followed the pain, and the organisations that adopted it early pulled ahead of the ones that retrofitted later.
The AI cycle is shorter. The State of FinOps 2026 report identifies FinOps for AI as the top forward-looking priority for FinOps teams, with 98% of practices now managing AI spend compared to 31% two years earlier. The FinOps Foundation’s June 2026 conference in San Diego is themed AI Value: The Era of FinOps for AI, Token Economics, and Agentic FinOps. A discipline is forming in real time, named explicitly, with practitioners already gathering around it. That is not a prediction anymore. It is a count.
The practitioners developing intent literacy now are doing what early FinOps practitioners did, what platform engineers did, what cloud architects did. They are not learning the old discipline in new clothes. They are inventing the one this layer demands.
Cloud taught organisations that infrastructure changes faster than governance.
AI is teaching individuals that execution now moves faster than articulation.
The bill is just how we found out.
Frequently Asked Questions
What is the iteration tax?
The iteration tax is the hidden cost of getting a useful answer from an AI model when your original prompt was imprecise. A vague request produces a near-miss. You clarify. The next response is closer but still wrong in some way you had not anticipated. You clarify again. You have now paid three times for one outcome and spent your own time on the round trips. At individual scale it is annoying. At organisational scale, across thousands of prompts a day, it is the difference between an AI budget that lands and one that triples.
How do I know if I have an iteration tax problem?
The simplest test is to count round trips. For a sample of your AI interactions, ask how many exchanges it took to get to a result you were willing to keep. If the typical answer is three or more, you have an iteration tax problem. A second signal is the shape of your follow-ups. If most of them are corrections, constraints you forgot to mention, or context you assumed the model knew, the issue is intent articulation rather than model capability. A third signal is your invoice. AI spend rising faster than AI usage is almost always the iteration tax showing up in aggregate.
Is this just prompt engineering by another name?
No. Prompt engineering is the technical craft of working with a specific model. It includes things like chain-of-thought structure, few-shot examples, system prompt design, and model-specific quirks. Communicative intent sits underneath all of that. It is the skill of knowing what you actually want clearly enough to describe it to anything, human or machine. Prompt engineering optimises the wrapper. Communicative intent supplies what goes inside it. You can be a skilled prompt engineer and still pay the iteration tax if the intent underneath the technique is unformed.
How is this different from FinOps for AI?
FinOps for AI is the organisational discipline of governing AI spend. Budgets, allocations, tagging, showback, optimisation, the operating model that makes cost visible and accountable. Communicative intent is the individual skill that determines how much spend gets generated in the first place. FinOps governs the meter. Intent literacy governs what flows through it. The two are complementary. Organisations that invest in one without the other tend to either build governance around waste that could have been prevented, or develop sharp individual habits that get drowned by ungoverned team-level usage.
What does communicative intent actually look like in practice?
It looks like stating the outcome before the task. Naming the audience. Declaring the constraints. Showing an example of what good looks like. Saying what you do not want. Surfacing the assumption you would normally leave implicit because a colleague would have inferred it. In day-to-day use it adds perhaps thirty seconds to a prompt and removes two or three round trips from the conversation that follows. The investment compounds. The articulation you build for one prompt becomes the scaffolding for the next ten, because most of the context you need to surface is stable across a body of work.
What is the one thing I should do first?
Pick the AI workflow you use most often and write down, once, the context that you currently assume but never state. Who the audience is. What the success criteria are. What the constraints are. What the failure modes look like. Keep it somewhere you can paste it into the top of a prompt. That single document, applied consistently, will remove more iteration tax than any prompt engineering technique you can layer on top of it. The discipline is not clever phrasing. It is making the implicit explicit, once, in writing, where you can reuse it.
Does this matter for individuals or only for organisations?
Both, but the individual case is the leading indicator. Organisational AI spend is the aggregate of individual interactions, and the patterns visible at scale are forming today at the keyboard. Individuals who develop intent literacy now will be more productive, more accurate, and more legible to the AI tools they work with. They will also be the ones who can lead the organisational conversation when it arrives. The cloud-era equivalent was the engineer who understood cost implications of architectural choices before FinOps was a discipline. That engineer became extremely valuable extremely quickly. The same window is open now.
How does this connect to your other writing?
This piece is the practitioner-altitude companion to The Cloud Said It First, AI Is Saying It Again, which makes the same argument at the organisational level. The structural response at the organisational level is Human Intent. The methodology at the engineering level is Intent-Driven Development. Communicative intent is the individual skill that all three rely on. Without it, the frameworks have nothing to capture and the governance has nothing to govern.
0 Comments