Is there an AI payoff outside tech?

On 24 October 2026 Torsten Slok, chief economist at Apollo, published five charts on where AI is paying off. Two of them carry the argument. Revenue per worker, adjusted for inflation to the fourth quarter of 2019, is rising for the Magnificent 7 and flat for the other 493 companies in the S&P 500. Profit margins split the same way: up for the seven, flat for everyone else.

The remaining charts point in the same direction. Productivity growth, measured as utilisation-adjusted total factor productivity from the San Francisco Fed and the Bureau of Labor Statistics, shows no lift from AI. Spending on AI, in data from Ramp, is concentrated in a small number of firms.

Slok states the conclusion himself: “The AI story will be decided in the S&P 493, not the Mag 7.” There is nothing here to rebut. If your company is in the 493, the chart describes it. If it is private, or listed somewhere else, it would sit among the 493 on any version of the chart drawn with your numbers. Expect the note to arrive from a board member this week with a one-line question about your own AI budget.

What can revenue per worker see?

Revenue per worker is a blunt measure, and that is why it helps here. It rises when the same people produce more. AI can only move it by doing work a person would otherwise do, on live work, every day, at the volume the business actually runs.

A tool that only a trial group uses can be excellent and still leave the number untouched, because none of the company's daily volume passes through it. Slok makes the same point at the scale of the economy: as long as agentic AI and token spending stay in a small part of it, he expects AI's wider effect to stay very limited.

A pilot that summarises documents for twenty people in a sandbox cannot show up in revenue per worker, however good it is.

That turns a macro chart into a question you can answer from your own books. How much of your AI spend runs on production work today, on volumes you already count, and how much is still in trials?

Where does the AI budget go instead?

The survey evidence says a large share never leaves trials. In S&P Global Market Intelligence's 2025 survey of more than 1,000 companies in North America and Europe, reported by CFO Dive in March 2025, 42% of companies had abandoned most of their AI initiatives, up from 17% in the previous survey. The average organisation scrapped 46% of its AI proofs of concept before they reached production.

IBM's 2025 study of 2,000 chief executives, published in May 2025, shows the same shape from the top. CEOs reported that 25% of their AI initiatives had delivered the return they expected over the last few years, and 16% had scaled across the enterprise.

Companies abandoning most of their AI initiatives, 2025 survey vs the year before

Share of surveyed companies, percent. More than 1,000 respondents, North America and Europe.

• Previous yearS&P Global survey
  17%
• 2025S&P Global survey
  42%
• 020%40%60%

• 46%

  of AI proofs of concept the average organisation scrapped before production.

  Source: CFO Dive reporting S&P Global, Mar 2025
• 25%

  of AI initiatives delivered the ROI expected, CEOs reported.

  Source: IBM Institute for Business Value, May 2025 (2,000 CEOs)
• 16%

  of AI initiatives scaled across the enterprise, CEOs reported.

  Source: IBM, May 2025

The surveys did not sample the S&P 493, but they describe companies like them. Across both, a large share of AI work stopped before production, and production is the only place it could have changed how much each person produces. Add that up across 493 companies and a flat line is what you would expect to see.

What does it take to move AI pilots to production?

The respondents named their own obstacles. Those in S&P Global's 2025 survey put cost, data privacy and security at the top. Half of IBM's CEOs said the pace of recent investment had left them with “disconnected, piecemeal technology”. Those complaints point to four things a pilot can skip and production cannot: access to the company's own systems, a way to test before anything goes live, permissions, and a price for every run. A newer model solves none of them.

What production asks of an agent that a pilot never does

The first is access. A pilot runs on sample data that someone copied in. In production the agent needs the company's own systems, and the procedures its experts follow when a case is unusual, or it handles the easy cases and hands the rest back to a person.

The second is a way to test before anything changes. A pilot never writes to a live system, so nobody has to ask what happens when it is wrong. Production needs somewhere to run the agent against real systems first, with its writes recorded rather than executed, so the people responsible can see what it would have done.

Permissions come third. In a trial everyone sees everything. A production agent needs an identity of its own and a defined list of what it may read and what it may change, which is exactly what a privacy or security review will ask to see.

Fourth, a price per run. A trial budget absorbs cost without anyone itemising it. Once an agent handles thousands of cases a month, finance needs the cost of each one, by model and by provider, to decide whether the workflow pays for itself.

What should a CFO fund next? Fund one, measure it, pause the rest

Before the next budget cycle, ask for one list: every AI initiative in the company, with two columns beside it. Is it running on live work today? What does one run cost? Then make three decisions from it.

1. Fund one. Pay for the path to production for the initiative with the clearest owner and a volume you already count: claims handled, invoices matched, tickets closed.
2. Measure it. Track output per person on that workflow against last quarter, and the cost of each run, from the first month.
3. Pause the rest. Stop the pilots that cannot answer either column, and revisit them once the first workflow has numbers.

The two-column AI inventory

Example. Illustrative rows, not data: replace them with your own initiatives.

In this example the first row gets the money. Its owner already reports a monthly count, so output per person has a baseline before anything is spent, and the result is a chart of output per person for one team, the workflow-level cousin of Apollo's revenue per worker, that you can put in front of the board.

For the pilots you pause, the Agent Maturity Index gives the COO and CIO a common maturity scale to rank them on before you revisit them.

Where does SynOS fit?

SynOS sits between a company's agents and its own systems and supplies what production asks for and a pilot skips. Agents built on it get context from the company's systems and from its experts' procedures, and rehearse against real systems with writes recorded rather than executed. Each one has its own identity, and every run is priced per model and per provider, so for those agents the cost column of the inventory is already filled in. SynOS is self-hosted by default, so the data stays on your infrastructure. Bring one workflow, the one with the clearest owner, and we will start there with you.

Questions CFOs ask about AI ROI