The fuel gauge and the flight plan — why AI spending is easy to see and AI value is not

1 The bill went up while the price went down

Two years ago, a million tokens of frontier-model output cost roughly sixty dollars. Today it costs a small fraction of that — model prices have fallen about 280-fold in twenty-four months. Over the same two years, enterprise spending on AI rose about 320%. The price of the thing collapsed, and the bill went up five-fold anyway.

That is not a contradiction. It is the whole story.

In April, Uber disclosed it had burned through its annual budget for AI coding tools in four months — thousands of engineers on tools like Claude Code and Cursor, and a leadership team that couldn't connect the rising token consumption to a matching rise in business value. One healthcare enterprise ran up more than six million dollars in unplanned inference costs against a single trillion-token workload. In a recent industry survey, 83% of enterprise IT leaders named AI cost unpredictability as a top concern.

The reason is structural, and it has a name now: the agentic loop. A chatbot answers once. An agent reasons, calls a tool, checks the result, corrects itself, and calls another — ten or twenty model invocations to finish one request. Industry measurements put agentic workflows at five to thirty times the token consumption of a simple chat. So even as the price per token falls, the number of tokens per unit of work explodes, and the bill outruns the price cut.

That's the part everyone can see, because it arrives on an invoice. Here is the part that doesn't.

2 One problem wearing two coats

MIT's NANDA initiative reviewed more than 300 publicly disclosed enterprise AI deployments and found that 95% produced zero measurable return. Not a small return — zero that anyone could measure. S&P Global's 2025 enterprise survey found 42% of companies had abandoned most of their AI initiatives, up from 17% a year earlier. By one accounting, of the roughly $684 billion enterprises spent on AI in 2025, over $547 billion produced no measurable result.

Read those two problems side by side and the industry's instinct is to treat them as separate. One is a cost problem — the bill is too big and too jumpy. The other is a value problem — nobody can prove the thing works. A cottage industry is forming around each. On the cost side: prompt caching, model routing, compression, cheaper models. On the value side: attribution frameworks, ROI calculators, "AI value management" as a new discipline. At FinOps X 2026, the industry's cost-governance community spent much of its main stage on exactly this split and left repeating one phrase: value per token.

But these are not two problems. They are one problem wearing two coats.

You cannot compute a cost-benefit ratio when both the cost and the benefit are invisible. Tokens are volatile and hard to forecast — the numerator jumps around. And the value delivered by any single AI action is almost never measured — the denominator is usually blank. Most of the market is racing to shrink the numerator. But cutting the cost of an initiative whose value nobody has measured doesn't fix the initiative. It just means you lose money slightly more slowly. If 95% of deployments show no measurable return, a 30%-cheaper token is a 30% discount on a zero.

The expensive mistake of the last two years was not that AI cost too much. It was that almost nobody built the instrument that shows cost against value in the same frame — so nobody could tell which of their AI spend was working and which was just warm. And in a market where 95% of deployments show no measurable return, the single most valuable discipline a vendor can offer is a rigorous, visible line between the two. Anyone can show you a big green number. The question worth paying for is: is that number measured, or is it modeled?

3 What we're building: bounded spend, and a value figure that admits what it is

I'll tell you plainly what we've built, what is doctrine, and what is still ahead — because a company whose entire pitch is honest measurement cannot fudge its own.

The cost side is the fuel gauge, and it's enforced, not suggested. In OpsATC.AI, every tenant carries a token budget and every user a daily cap. When a request would cross the line, it is stopped — with an atomic, race-safe cutoff, so two requests arriving at once cannot both slip past a nearly-exhausted budget, and every block is written to an append-only record. Your AI spend is bounded by a number you set, not by how the month happens to go. That is shipped and enforced in code today.

The value side is where we refuse to draw a flight plan on the fuel gauge. The Captain is read-only against your systems: she reads, reasons, cites her sources, and recommends — and a human decides. Every recommendation she surfaces is logged the moment it appears, with the value it modeled attached and the decision a human took — accepted, rejected, actioned — recorded against it. That's the spine of real attribution: not a black box that emits a monthly savings figure, but a record of discrete, cited, human-decided actions, each carrying the number the AI claimed for it.

The Captain RECOMMENDATION

4 min ago · Buffer rebalance · Distribution · Logbook entry #L-2213

⋯

Two SKUs are carrying three weeks of excess safety stock while a third is about to stock out. Rebalancing the buffer covers the gap without a rush order.

Modeled value Avoided rush freight + carrying cost ~$8,200  ·  MODELED — not yet measured

Basis Lane rate + on-hand + demand signal Projection from connected data

Disposition Awaiting your decision Logged the moment this surfaced

The operator decides:

Accept Modify Reject

◷ View 3 sources · ERP, WMS, planning suite

ILLUSTRATIVE · IN-APP EXAMPLE

And here is the line we hold, which is the whole point of the essay: the value a recommendation modeled is labeled modeled — not measured, not realized, not captured. Until a human confirms what actually happened, or an outcome is measured against the systems of record, it is a projection, and we mark it as one. A vendor who paints a modeled figure as a realized one isn't being optimistic; they're rebuilding the exact problem you're trying to escape.

What's ahead — and I would rather name it than imply it's done — is the automated close of that loop: measuring, from your connected systems, what a recommendation actually delivered, and setting that measured outcome against what was modeled, so the projection and the result sit side by side. That is on the roadmap, not in your hands today. When it ships, this newsletter will say so, on the day it does.

The bet underneath all of it is simple. The AI market spent two years and hundreds of billions of dollars optimizing a ratio whose bottom half nobody was measuring. The companies that come out ahead won't be the ones with the cheapest tokens. They'll be the ones who can point at a specific decision and say, with a straight face, this one paid for itself, and here is the measurement — and who can be trusted, because they never once dressed a projection up as a fact.