AI Hasn't Had Its ROAS Moment

In the summer of 2025, MIT's Media Lab put a number on something everyone already felt. After $30 to $40 billion in enterprise spending on generative AI, 95% of organizations were seeing no measurable return. The report — The GenAI Divide: State of AI in Business 2025 — went off like a bomb. To the skeptics it confirmed a bubble. To the believers it was an embarrassment to be explained away.

Both camps misread it, because both were asking the same broken question: why isn't AI returning on our investment?

I want to offer a different answer. The 95% aren't failing to get a return. They're measuring the wrong thing — and the right way to measure it is something one corner of the business world has been doing for twenty years.

The misdiagnosis

Read the explanations for the 95% and everyone reaches for a familiar villain. The models aren't good enough. The regulation is too heavy. The talent is too scarce. The data isn't ready.

MIT's own researchers found something quieter and more damning. The barrier wasn't model quality, infrastructure, regulation, or talent. It was that the systems didn't integrate into a workflow, didn't adapt, and didn't learn over time. The deployments that failed were brittle — capability dropped into an operation with nothing around it to convert that capability into a result. The deployments that worked were built into the process, judged by business outcomes instead of demos, and most often bought from specialized vendors rather than assembled in-house.

The model was never the problem. What was missing was the machine around it.

AI is fuel

Follow the revenue in this entire wave and it pools, overwhelmingly, in one place: the layer that takes raw electricity and data and refines them into a finished commodity everyone downstream consumes. That is not what an engine does. That is what a refinery does. The only other industry that ever made its money in exactly that shape is oil and gas.

AI is not the engine. It's the fuel — abundant, refined, and inert until something converts it into directed work. That single reclassification is the whole essay, because fuel and the machines that burn it are valued in completely different ways.

How we already account for fuel

Here's the objection, and it's a good one: firms compute returns on operating expense all the time. Marketing spend, sales spend, training, consulting — all operating expense, all routinely measured for return. So any flat claim that "you can't compute a return on a consumable" is simply wrong, and a CFO would say so in the first thirty seconds.

But look at how finance does it when it does it well.

The cleanest example is ROAS: return on ad spend. A marketing team does not ask what return the ad account intrinsically produces. They ask what ROAS the system is generating. And ROAS is not a capital ROI — it's a ratio, revenue produced per dollar of spend. A conversion coefficient. Everyone in the room understands that the spend itself produces nothing; the targeting, the creative, the funnel, the attribution stack produce the result. Two companies spend identically on ads and get wildly different ROAS, and the entire difference is the machinery around the spend.

And ROAS isn't alone. Finance already runs on conversion coefficients — customer acquisition cost, revenue per employee, asset turnover, manufacturing yield. Each measures how efficiently a resource becomes output, tracked over time and benchmarked across peers. The coefficient I'm proposing for AI isn't a novel intuition you have to accept from scratch. It's the newest member of a family finance already trusts.

That is the argument of this essay, already operating in one corner of the business for two decades. Marketing had its reckoning long ago — it stopped demanding a return from the spend and started measuring the efficiency of conversion. AI hasn't had that reckoning yet. It's still being judged the way marketing was judged before ROAS: as raw expenditure expected to return on its own.

The conversion coefficient

Generalize what ROAS does for advertising and you get a single framework for AI spend. Call the variable k.

• E — AI operating expense: the fuel. Inference, tokens, access. A flow.
• O — the value-output attributable to it, in dollars: revenue gained or cost avoided. A flow.
• k = O / E — the conversion coefficient. Dollars of output per dollar of AI spend. ROAS, generalized.
• k₀ — the baseline coefficient: what you converted before.
• C — the engine: the conversion mechanism that directs the fuel — workflow, integration, the causal layer. A durable system.
• T — the period over which the lift holds.

The framework is short. Measure the fuel by its coefficient, k = O / E, never by an intrinsic return. Recognize value only to the extent the lift is real and measured — Δk = k − k₀ > 0 — so spend that produces no measured lift recognizes no value. And put the return question where it belongs: on the engine.

Note what gets recognized: the lift over baseline, not the gross. AI takes credit only for Δk — never for the output you'd have produced anyway. That incrementality is the first thing a serious finance reader will check, so it's built in.

A worked example

Two companies. Same industry, same models, same fuel bill.

Company A spends E = $500,000 a year on AI for support and back-office work. It drops the capability in and hopes. Drafts get reworked, escalations barely move, the productivity is real but diffuse and never reaches the P&L. Measured honestly, the attributable value is about what they spent. k ≈ 1.0, Δk ≈ 0, recognized value ≈ $0. Company A is the 95% — not because the AI failed, but because nothing converted it.

Company B spends the same $500,000 on fuel, but first invests C = $300,000 in the engine: the integration, the workflow, the causal layer that decides what the model should do and verifies it did it. Now the same fuel produces O = $2,000,000 in measured cost avoided and retained revenue. k = 4.0.

The fuel spend never changed. The entire difference between the 5% and the 95% was a $300,000 engine that moved the coefficient from 1 to 4. The return didn't live in the AI. It lived in the machine around it.

This has happened four times before

If that feels like a special claim about AI, it isn't. It's the most ordinary pattern in the history of enterprise technology.

ERP systems did not create value because SAP licenses existed — the value came from the firms that re-engineered their operations around the software. CRM did not create value because a contact database existed; it came from rebuilding the sales motion around it. Cloud did not create value because servers changed location; it came from re-architecting how software gets built and shipped. Industrial automation did not create value because robots were installed; it came from redesigning the line.

In every case the capability commoditized and the return pooled in the conversion layer — the operational redesign, the integration, the new operating model. AI is not the exception. It's the latest instance of a recurring economic phenomenon. And marketing, with ROAS, is the case that already named the coefficient. AI just hasn't caught up to its own precedent.

The question that turns productivity into value

Now the part that matters most, because it's where this stops being a productivity discussion and becomes a valuation one.

The question is not did k rise. It's will it stay risen.

A workflow that lifts k for six months is useful. A workflow that lifts k for five years is an asset. Investors have always known the difference — nobody pays a multiple for a temporary efficiency gain; they pay it for a durable mechanism that keeps producing the gain. Enterprise value, in the end, is the durability of Δk.

This is why the framework carries an impairment rule, and why that rule isn't pedantry. MIT found the failed deployments were brittle — they didn't adapt, didn't learn, didn't hold. In coefficient terms, their k decayed back to baseline. A conversion mechanism whose lift evaporates is an impaired asset and should be written down like any other. The deployments that worked had the opposite trait: their mechanisms compounded, because they were built to learn and stay integrated. Durable Δk. That is the asset. Everything else is a feature wearing an asset's clothes.

Where the value actually lives

Put it together and the field reorganizes. If AI consumption is a coefficient on operating expense, then the only durable, return-bearing asset in the entire AI stack is the engine — the conversion mechanism that raises k and holds it there. Not the model. Not the tokens. The machine around them.

That is the single place a real return can exist, which means it's the single place durable enterprise value can accrue. Most of the market is still selling fuel and calling it transformation. The companies worth building — and the companies worth backing — are the ones building engines: the mechanisms that turn potential into directed work, move the coefficient, and keep it moved.

We don't have a return problem in AI. We have a measurement problem. Ask the question marketing already learned to ask — not what does the spend return, but how efficiently does the machine convert it, and will it keep converting — and the fog clears.