There is a month-two scene I keep hearing described, in almost the same words each time. The pilot went well; the pilot budget rounded to zero. Then the agent workflow went to production, the first full invoice landed, and someone in finance is now holding a number nobody warned them about, asking the engineers a question they find they cannot answer: what exactly did we buy? Set that scene against the market backdrop and it looks impossible. Blended enterprise token cost — tokens being the metered unit of AI text, a few characters each, priced per million — fell 67 percent year on year, from $18.40 to $6.07 per million, measured across an analysis of 2.4 billion API calls, driven by open-weight competition and multi-model routing. Prices collapsed. The bills rose anyway.
Nobody in that story was cheated. The unit price genuinely collapsed; the vendor's celebration slides are accurate. What happened is that the quantity consumed is governed by an architecture decision nobody priced, the default version of that decision is the expensive one, and the pilot was structurally incapable of revealing it. The gap between falling prices and rising bills is not a market failure. It is a cost model that was never shown to the finance function, because the people who understood it did not think of it as a financial document.
This piece is that cost model, written out. The one-sentence version, worth keeping: with a context-stuffed architecture, your AI bill scales with the size of your archive; with a governed one, it scales with the size of the task. Falling token prices discount both. They cannot fix the multiplier — a 67 percent discount on an archive-shaped bill is still an archive-shaped bill, and archives only grow.
The concession first: the price collapse is real, and the objection it funds deserves a full answer
The strongest objection to everything that follows is macroeconomic, and it should be stated at full strength rather than discovered in the comments: unit prices are falling so fast that architecture discipline is a distraction — whatever you waste this year costs half as much next year, so build sloppy and let the market fix your bill.
Take the premise seriously, because it is better than it sounds. The 67 percent decline is measured, not vendored; it reflects genuine price cuts plus a rational mix shift toward cheaper models that are good enough for more of the workload. Independent measurement of the price of a fixed capability level — what it costs to buy last year's benchmark score this year — falls faster still. Anyone modelling agent economics on frozen prices is building a straw man, and I am not going to.
Here is why the objection still fails. It assumes the quantity term is fixed while the price term falls. But the quantity term is the product of run count, step count and context size — and all three grow with programme success. Run count grows because that is what scaling a programme means. Step count grows because each capability generation makes longer chains viable, and reasoning-mode models consume thinking tokens on top. And context size grows because the default architecture loads the archive, and the archive compounds. A declining price multiplied by a compounding quantity is a race, not a rescue — and the quantity is compounding on purpose, because every one of its growth terms is something the programme calls progress. The bill does not fall when prices fall. It falls when someone changes the shape of the runs, which no market does on your behalf.
A seat is bounded by human hours. A run is bounded by nothing but the architecture.
The deeper error, underneath the pricing one, is a category transfer from the software the enterprise already knows how to buy. Two decades of SaaS trained procurement to think in seats, and a seat has a property so convenient nobody noticed it was load-bearing: it is self-limiting. A licensed human works bounded hours, gets tired, goes home. Per-seat cost is therefore roughly linear in headcount, and finance can model it with a spreadsheet and a hiring plan.
An agent run has no such bound. An agentic workflow is not one model call; it is ten, twenty, forty steps — plan, retrieve, call a tool, read the result, decide, act, check. At each step, the system re-reads its working context — the briefing material loaded into the model's reading window: everything loaded rides along on every call, and is paid for again. Consumption is therefore proportional to steps multiplied by context, and neither term answers to headcount. The correct mental model is not a licence; it is a metered industrial input, like power — except that where a factory's power draw is bounded by its machinery, an agent's draw is bounded only by what the architecture permits it to re-read. Procurement models the licence. Nobody models the meter.
This is also why the two departments that should have caught it each missed half. Engineering understood the re-read mechanics but thought of them as a latency and quality problem. Finance understood consumption pricing but was told the unit price was collapsing, which sounded like the problem solving itself. The multiplier lives exactly in the seam between them.
And the seam is widening, not closing, because each new capability makes the quantity term bigger. Reasoning modes add thinking tokens the user never sees. Multi-agent patterns multiply the number of contexts being re-read in parallel. Longer windows invite longer loads, because the capacity exists and nothing in the default architecture charges anyone for using it. Every one of these is sold — accurately — as a capability improvement. Every one of them is also, unannounced, a consumption multiplier, and the organisation that meters by invoice rather than by run will discover each of them the same way: two months late, in aggregate, unattributable.
The arithmetic, worked in full
Put numbers on it, at mid-tier prices of roughly $3 per million input tokens. The default architecture loads everything that might help — the document store, the case history, the wiki — because context windows are huge now and stuffing them is the path of least engineering. Half a million tokens of just-in-case, re-read across a twenty-step run, is about ten million input tokens: roughly $30 per run before the model writes a single word of output. At a thousand runs a day, that is $30,000 a day — eleven million dollars a year — for one workflow, at prices everyone agrees are historically cheap.
Now run the same workflow through a governed context assembly: the eight thousand tokens this task class actually needs and is entitled to — the case record, the applicable policy, the tool results — re-read across the same twenty steps. About 160,000 input tokens. Roughly $0.48 a run; $480 a day for the same thousand runs. Same task, same model, same price sheet. The sixty-times multiplier is not a pricing fact. It is an architecture decision, made by default in week two by whoever wired the retrieval pipeline, and never surfaced to anyone who owns a budget.
Two caveats keep this honest, and both are on the figure. Provider pricing spans $0.14 to $10 per full million-token call — a seventy-one-fold spread — so your constants will differ. And prompt caching genuinely discounts re-reads on unchanged prefixes. Neither changes the structure: caching discounts the archive-shaped bill without making it task-shaped, and a discounted multiplier is still a multiplier. The decision that matters is what gets loaded, not what it costs to reload.
Why the pilot never predicts the bill
The month-two surprise has a mechanism, and it is worth stating precisely because it will otherwise repeat at the next capability generation.
Pilots are run by engineers on curated inputs: short context, clean documents, demo scenarios. Cost per run rounds to nothing, so nobody instruments it. The extrapolation to production is then done on run count — a thousand runs a day instead of fifty — while the quantity that actually explodes is run shape. Production context carries the accumulated sediment of everything anyone thought might help, and sediment only accretes: every incident adds a document to the pile, and no incident ever removes one. The pilot measured a system that does not exist in production.
The second mechanism is quieter and more expensive. The research on long-context degradation — Chroma's context-rot results across eighteen frontier models are the cleanest public version — shows accuracy falling well before the advertised window fills; GPT-4.1 lost roughly a quarter of its accuracy at about a tenth of its rated window in their released results. Which means the bloated runs are not just costly. They are worse. And worse outputs produce retries, escalations and human clean-up — costs that are real, caused by the same architecture decision, and never appear under the AI line item at all. They surface as analyst hours on someone else's budget line, which is why nobody attributes them. The archive-shaped architecture is billed twice, and the second bill is filed under a different department.
The third mechanism is organisational rather than technical: nobody owns the meter during the window when owning it is cheap. The provider invoice arrives monthly and aggregated, engineering's telemetry tracks latency and error rates rather than tokens per outcome, and the workflow tags that would join the two do not exist because nobody asked for them at build time. By the time the month-two invoice makes the problem undeniable, the architecture has been in production for six weeks, the archive-shaped pattern is load-bearing, and the conversation has become a remediation instead of a design review. The same fix costs a sprint at design time and a quarter afterwards.
The five-times variable
If the mechanics above are right, you would expect the programmes that can see their consumption to be the programmes that can prove their value — and that is exactly what the best available survey data shows.
KPMG's Global AI Pulse for the second quarter of 2026 — 2,145 senior leaders across twenty markets — found that only 26 percent report full, real-time visibility into what their AI systems cost to operate, while planned AI spend runs to a weighted average of $202 million over the following year. Sit with that pair for a moment: three-quarters of large enterprises are planning nine-figure spend on a consumption-priced input they cannot see in real time. In no other category of industrial input would that sentence survive an audit committee.
The finding that matters most sits one cross-tabulation deeper: leaders reporting strong cost visibility are five times more likely to report established ROI — fifteen percent against three. Visibility is not bookkeeping hygiene; on this data it is the single strongest observable correlate of programmes that can prove they work. The mechanism is not mysterious. The instrumentation that produces cost-per-outcome is the same instrumentation that produces outcome measurement at all — you cannot know what a resolved case costs without first knowing what a resolved case is and counting them. The five-times gap is what it looks like when one cohort runs the economics per outcome and the other divides the monthly invoice by enthusiasm.
The market is already running this referendum
If the internal version of this discipline feels optional, the last week of July was a preview of how it stops being optional, administered at the largest possible scale.
In one earnings week, the three most-watched AI spenders guided roughly half a trillion dollars of capital expenditure. Alphabet, on 22 July, raised 2026 capex guidance to $195–205 billion — its third raise of the year, having started at $175–185 — on a quarter where revenue beat and cloud grew 82 percent; the shares fell about five percent anyway. Meta, on 29 July, narrowed its 2026 guide to $130–145 billion, raising the floor by $5 billion, disclosed $31.1 billion of quarterly capex and finance leases, watched the analyst call converge on free cash flow, and dropped roughly nine percent after hours. Microsoft, the same day, closed its fiscal year with $41 billion of fourth-quarter capex — up seventy percent — and an expectation of roughly $175 billion for fiscal 2027, restated from about $190 billion by a useful-life accounting change that is a reclassification rather than a cut; its shares rose, because the guidance held and the demand signal behind it was legible.
Read the three reactions together and the pattern is unmistakable: the market has stopped paying for AI capital intensity as narrative and started demanding its unit economics. Beats no longer buy immunity; legibility does. And what public markets do to trillion-dollar companies in one after-hours session, boards and audit committees do to internal programmes over one budget cycle, with a lag of about two quarters. The CFO asking what a resolved case costs is not being difficult. The CFO is applying, internally, exactly the test the tape applied to Alphabet and Meta in the last week of July — and the programmes that cannot answer it will meet the internal version of that nine-percent repricing at planning season.
Where the pressure lands first — the three-market read
The same discipline arrives through three different doors in the three markets this practice covers.
In North America, the door is the earnings cycle itself. The capex referendum above is a US story first, and it propagates: listed enterprises inherit the analyst's question at their own earnings calls, and their boards inherit it from there. The supervisory side pushes the same way by stepping back — the revised interagency model risk guidance of 17 April (Fed SR 26-2 / OCC Bulletin 2026-13) supersedes SR 11-7 and expressly places generative and agentic AI outside its scope, a deferral that leaves cost and control questions with the institution rather than the examiner. Out of scope is not out of risk. Nobody in Washington is going to specify your agent economics. Your audit committee will.
In India, the door is the regulator, and it is already open. The RBI's draft model risk guidance of 24 June puts every model a bank relies on in scope — third-party and foundation models included — with autonomy as a tiering axis and monitoring obligations that presuppose exactly the per-outcome instrumentation this piece describes. A bank that cannot attribute agent consumption and outcomes per workflow will struggle to evidence the monitoring the draft contemplates. Separately, Mission-subsidised compute at ₹65–150 per GPU-hour is quietly changing the build-versus-rent arithmetic for steady-state volume workloads — which makes cost-per-outcome the deciding instrument, since it is the only number that survives a change of serving substrate.
In the Gulf, the door is abundance itself. Sovereign programmes are making capacity plentiful and cheap domestically, so the binding constraint on an enterprise programme is not access to compute — it is proving per-outcome value and assurance to conservative supervisors and sovereign shareholders. When capacity is the subsidised default, the differentiating artefact is the unit-economics record, because it is the one thing the programme cannot inherit from the national buildout.
Three doors, one room: whatever your jurisdiction, the instrument set that answers the question is identical, which is the argument for building it once, now, rather than per-stakeholder under deadline.
Honest limits — what this argument does not establish
Four weaknesses, stated plainly.
The worked example is an illustration. The $30 and $0.48 are constructed from representative parameters, not measured from a named production system — I cannot cite one without breaching confidences, so treat the structure as the claim and your own metering as the test. The 71-fold provider price spread means your constants will differ; the multiplier's existence, not its exact value, is the argument.
The 67 percent is a blended figure. It reflects genuine price cuts and a mix shift toward cheaper models, and it is the right number for claims about the commoditising layer — but it does not mean any specific model got 67 percent cheaper, and it should not be quoted that way.
The KPMG finding is survey self-report, and the causality can run both ways. “Established ROI” is a claim by the programme's owner, not an audit, and programmes that work may get instrumented as a consequence of working rather than the reverse. I read the five-times gap as mechanism plus selection, in unknown proportions — the practical implication survives either reading, since the instrumentation is a precondition of proving value whichever direction the arrow points.
And the capex parallel is a frame, not an equivalence. Hyperscaler capital expenditure is depreciated infrastructure investment; enterprise token spend is operating cost. The guides differ in definition — Microsoft's figure is fiscal-2027 and reflects a lease-accounting change, Meta's includes finance leases — and one day's after-hours move is a reaction, not a verdict. What transfers is the direction of scrutiny, not the accounting.
What a CFO instruments on Monday
None of the following requires technical depth to demand, and none of it requires a platform purchase to supply. It requires three numbers, one tagging discipline, and one gate.
First: cost per outcome, trended. Not spend per month — the fully loaded cost of a resolved case, a processed claim, a closed ticket: model consumption plus retries plus human touches, divided by outcomes that completed. The supply side is ordinary engineering: every model call carries a workflow and action-class tag, billing exports join to those tags, and rework hours get attributed. If the programme cannot produce this number within a sprint, that inability is itself the finding — it means nobody is watching the multiplier.
Second: context budget per action class. How many tokens does a limit adjustment need? A claims summary? A well-governed programme sets these the way manufacturing sets a bill of materials — deliberately, with an alert on breach — because the discipline that bounds cost is the same discipline that bounds risk: assembling only the context this task is entitled to is one control with two beneficiaries. Finance and the risk function are natural allies here, and in most organisations neither has noticed.
Third: retry and escalation rate, attributed. The hidden line item. When context bloat degrades accuracy, the failure returns as rework — billed once in tokens and again in analyst hours on another budget line. Until it is counted against the workflow that caused it, the archive-shaped architecture will keep looking cheaper than it is, and the clean-up will keep looking like a staffing problem.
Then the gate: no workflow scales past pilot until its three numbers exist and are trended. Not until they are good — until they exist. The reversibility is the encouraging part, and it is the part I would put in front of a board: unlike most cost problems, this one is an architecture decision, and programmes that move from context-stuffing to per-task assembly routinely cut consumption by an order of magnitude while accuracy improves, because the model stops drowning in its own archive. In my experience that is the rare cost initiative that reduces risk instead of adding it — though I would hold it as practitioner observation rather than a cited study, and let your own first workflow's numbers make the case.
The unit price will keep falling; the vendor will keep celebrating; and neither motion touches the multiplier, because the multiplier lives in your architecture, not their price sheet. The bill stops being shaped like the archive on the day someone in the building asks what the runs are shaped like — and that someone, on the evidence of this July, had better get there before the market does. If your agent bill stopped making sense in month two and nobody can say why, this decomposition is a short piece of work with a long payback — vikramjha.work is where to find me.