Apollo’s chief economist Torsten Slok ran the numbers in October and found nothing. Revenue per worker is climbing for the Magnificent 7 and flat for the S&P 493. Profit margins, same story. Productivity growth, no movement at all. His conclusion was blunt: AI has delivered zero financial payoff to corporate America outside of tech. If you run quality or operations at a manufacturer, that data probably matches what you have seen after a year of pilots.

But the charts measure spending on models, not changes to how work gets done. Those are different things, and only one of them produces margin. This post breaks down why AI ROI outside tech keeps stalling, where the value actually sits in a plant or a quality department, and what a payback case looks like when you redesign the workflow instead of buying tokens.

Flat Margins, Flat Productivity: The Data Nobody in AI Wants to Quote

Slok’s sharper point gets less attention than the charts. Token spending is concentrated in VC-backed companies, while mainstream corporate America “does not need the latest model and is unwilling to pay for it.” Finance teams have stopped treating that spend as growth and started booking it as a cost centre that drags on margins.

The bottom line is that the AI story will be decided in the S&P 493, not the Mag 7, and so far, there is no evidence that revenue growth or margins are improving outside the tech sector.

So you are being asked to fund AI in precisely the segment where the payoff has not shown up yet. That is not an argument for waiting. It is an argument for being ruthless about where you point it.

What Slok Actually Measured, and What It Does Not Prove

Three measures sit behind the headline: revenue per worker, profit margins, and total factor productivity. The TFP series is the utilization-adjusted one from the San Francisco Fed (Fernald, 2014), with the revenue figures inflation-adjusted to 2019 Q4. These are the right tools for answering a macro question. They are the wrong tools for answering whether your scrap rate moved.

Why aggregate margin data lags operational change by years

Margin data is quarterly, consolidated, and reported after the fact. A workflow change made on a production line in March shows up in a segment margin maybe four quarters later, if nothing else in the cost base moves in the meantime. Commodity prices, labour agreements, and FX swings all move faster and harder than any internal efficiency gain.

Total factor productivity is worse on this front. It is a residual, calculated across the whole economy, smoothed for cyclical variation in capacity use. By design it cannot see a single plant. Historically, the productivity effects of electrification and enterprise software took a decade to register in national statistics, and nobody argues those technologies did nothing.

The gap between a working pilot and a visible financial result

Say a deployment removes two FTE-equivalents of document review from a quality department. That is real, measurable, and defensible in a budget review. Inside a company with billions in revenue, it rounds to zero on the margin line. A hundred of those, running in parallel across sites, would be visible. Most manufacturers have one or two, sitting in a pilot that never crossed into production.

So both things are true at once. Slok is right that there is no evidence of improving revenue growth or margins outside the tech sector. That is a statement about the average, and the average company bought model access and changed nothing about how work flows. The question of AI ROI outside tech is not settled by that data. It is settled plant by plant, and most plants have not started.

Token Spending as a Cost Center: Why Mainstream Firms Are Pulling Back

The Ramp AI Index data behind Slok’s charts shows the split clearly. VC-backed companies keep pushing token spend higher because burning capital on frontier models is part of their story. A plant with a 6% EBIT margin has no such luxury. When your AI bill arrives every month and your scrap rate has not moved, finance is right to call it a cost.

Most of what manufacturers actually need AI for is unglamorous. Pulling values off a supplier CoA. Routing a deviation to the right reviewer. Drafting the first version of a CAPA. Comparing a spec revision against the previous one. None of that requires reasoning at the edge of what the field can do.

Match the model to the task, not to the headline benchmark

Frontier models are priced for frontier problems. Document extraction, classification, and templated drafting are commodity work, and small or mid-tier models handle them at a fraction of the cost per call. Running all four through the same premium endpoint because it scored well on a public benchmark is how a useful tool turns into a margin drag.

Build a tiering rule and apply it per workflow. Cheap model for extraction and triage. Mid-tier for summarisation and drafting. Frontier only where the output feeds a judgment call a human cannot easily verify. Test the cheap option first and move up only when accuracy on your own documents fails the bar you set.

Measure cost per completed task, not cost per seat

Seat licences hide everything that matters. A €40 per user per month tool tells you nothing about whether the deviation got closed faster or whether anyone opened the thing twice. Cost per completed task does: tokens plus review time, divided by the number of items that actually cleared the workflow.

Track that figure against the manual baseline from day one and it stays honest. If a CAPA draft costs €0.80 in tokens and nine minutes of reviewer time against forty minutes before, you have a number your CFO can test. That is the discipline separating AI ROI outside tech from AI spend outside tech.

Judgment, Accountability and Trust: The Limits Slok Names Correctly

Slok makes one argument that deserves more attention than his charts. He writes that AI “does not understand accountability, trust, culture, unwritten rules, ethical trade-offs or fairness, all of which shape reputational risk and ultimately drive revenue growth and profit margins.” He is right, and in regulated manufacturing that limit is not philosophical. It is written into your quality system.

AI may be a great calculator, but in business, intelligence is mostly judgment. Judgment depends on context, and the answer varies with which human or which model you ask.

A model can draft a deviation report. It cannot sign it. An auditor opening your batch record does not ask which model summarised the investigation, they ask who reviewed it, who approved it, and whether that person was qualified to do so. Accountability has a name attached. Software does not have one.

This is the cleanest predictor of where AI pays off and where it stalls. Work that is high-volume, rule-bound and evidence-based returns value fast, because the output can be checked against a source document in seconds. Work that is a judgment call with reputational or regulatory consequence returns almost nothing, because the review takes as long as doing it yourself.

So the real design decision is where you place the human signature. Put it too early and you have automated nothing, because a person still reads every line. Put it too late and you have an unsigned record and an audit finding waiting to happen.

Get that placement right and AI handles the volume while your people spend their time on the calls that actually need a name behind them. That is where AI ROI outside tech starts showing up.

Where the S&P 493 Can Actually Find Its Payoff in 2026

Slok asks whether mainstream corporate America can turn “increasingly cheap intelligence into higher revenue growth and higher profit margins.” That is the right question, and the answer depends almost entirely on whether operations teams stop buying capability and start redesigning process. A licence purchased changes nothing. A workflow rebuilt around a model changes cost per cycle.

Which means the next budget cycle should look different. Pick two high-volume workflows, not eight. Cap model spend at a number finance signs off on before deployment, not after the invoices start arriving. Then hold the line on measurement.

The three metrics to baseline before your next deployment

You cannot prove a payoff you never measured a starting point for. Most manufacturing pilots fail this on day one, which is exactly why they produce anecdotes instead of evidence.

• Hours per cycle: How long does one document review, one deviation routing, one CoA entry take today, measured across a real month rather than estimated in a meeting.
• Defect or rework rate: The current percentage on the specific process you are touching, not the plant-wide figure.
• Fully loaded cost per cycle: Labour, rework, and delay cost combined, so the model spend has something honest to sit against.

Capture all three before anything is deployed. Two weeks of baselining saves you a year of arguing with your CFO about whether the spend did anything.

A six-month decision gate that kills pilots that aren’t paying

Set the gate at the start and write down the threshold. If the workflow has not moved hours per cycle or defect rate by a defined margin in six months, it gets switched off. No extensions, no “the next model version will fix it.”

That discipline is what separates a cost centre from an investment. Kill fast on the ones that do not land, and redirect the budget to the one that did. The macro charts will stay flat until enough operations teams run this unglamorous loop, and that is how AI ROI outside tech eventually becomes visible at all.

Source: apollo.com