AI shows zero payoff outside Big Tech. What the companies it pays off for do differently
Apollo's chief economist says AI has shown no financial payoff for most companies, while Anthropic's own AI use now outweighs its payroll at retail prices. Why the gap exists, and a five-step test to run before you pay for an AI project.
The short answer
- Apollo chief economist Torsten Slok wrote on October 24, 2026 that revenue per worker and profit margins are rising for the Magnificent 7 tech companies and flat for the other 493 companies in the S&P 500.
- Slok's note argues that 'the AI story will be decided in the S&P 493, not the Mag 7', and finds no sign yet of AI lifting productivity growth.
- Business Insider reported in November 2026 that Anthropic's internal AI use, valued at retail prices, now exceeds its employee pay. That is a valuation at list prices, not cash spent.
- AI pays off when it is wired into one piece of work with a number attached, such as briefs collected, hours saved or orders handled, and that number is checked after 30 days.
- On tanujrajput.com, rebuilding the project intake on Claude Haiku 5.5 cut a full brief from about 12 chat messages to four or five taps, with a median reply time of about 1.8 seconds.
Want this done rather than explained? see my published prices
Apollo's chief economist, Torsten Slok, wrote on October 24, 2026 that AI has shown no financial payoff for S&P 500 companies outside the biggest tech firms. Revenue per worker and profit margins are rising for the Magnificent 7 and flat for everyone else. A week later, Business Insider reported that Anthropic's own AI use, valued at retail prices, now exceeds what it pays its staff. Both stories describe the same gap. AI spending is easy to see. The payoff only shows up when AI is wired into one piece of work that earns or saves money, and someone checks the number afterwards.
I build AI tools and web apps for founders, and I run several AI products of my own. This post is the test I use before I'd tell anyone, including myself, to pay for an AI project.
What did the Apollo economist actually say?
Torsten Slok, partner and chief economist at Apollo Global Management, published "No Signs of AI Payoff Outside the Tech Sector" on October 24, 2026. His evidence:
- Revenue per worker is rising for the Magnificent 7 and flat for the S&P 493.
- Profit margins follow the same split.
- There are no signs yet of AI lifting productivity growth.
- Mainstream companies are starting to treat AI token spending as a cost that weighs on margins.
- AI spending is concentrated in a small number of firms.
His conclusion: "the AI story will be decided in the S&P 493, not the Mag 7." TechCrunch's write-up ran with the phrase "zero financial payoff", and the story drew hundreds of comments on Hacker News. Slok is describing the data so far. He isn't saying AI can't pay off.
What does Anthropic's AI bill have to do with it?
Business Insider reported that Anthropic's internal AI use, priced at retail rates, now exceeds its employee pay. The article also described tech companies starting to treat AI tokens as a "fourth component of employee compensation."
Hacker News readers were quick to point out that this isn't cash out the door. As one commenter, hungryhobbit, put it: "It's valuing its internal AI consumption at retail prices." Anthropic runs its own models, so the number is what that usage would cost a customer.
That is still the useful part. The company that makes the models puts a retail price on its own usage. Most small businesses I talk to have never put a price on theirs, so they can't say whether it paid off.
Why do most companies see no payoff from AI?
From the projects I see, three patterns explain most of it.
- AI sits beside the work instead of inside it. Staff copy text into a chat tab, paste the answer back and lose the time they saved. I wrote about this pattern in stop switching to ChatGPT tabs.
- Nobody picked a number before the build. If "better" was the goal, there is nothing to measure afterwards.
- The first version was too big. A six-month AI platform delays any result by six months, and the budget runs out before the number moves.
What do companies that profit from AI do differently?
They pick one job, attach a number to it and measure it. Here are two from my own work.
My project intake. The old intake chat on this site needed about 12 messages to collect a project brief. I rebuilt it on Claude Haiku 5.5 with tap-to-answer options, and it now reaches a complete brief in four or five taps. The median reply comes back in about 1.8 seconds. The number I watch is briefs completed, because collecting briefs is the only job it has.
A weekly founder brief for a D2C brand. I built a private dashboard on the brand's Shopify data that emails the founders a one-screen summary each Monday: sales against last week, stock running low, dead stock and margin from real cost prices. It does one job a week, and the result is easy to check: the founders know where the week stands without opening Shopify.
Neither project is large. Both do one thing whose result you can count.
How do you test an AI project before paying for it?
Run these five steps before you sign off a budget, whether I'm building it or someone else is.
- Name the number. Hours per week on a task, leads collected, tickets closed, refunds handled. One number.
- Price the work per unit at list prices. Work out the AI cost per brief, per ticket or per report, the way Anthropic prices its own usage. If the cost per unit is close to what a person costs, stop here.
- Build the smallest version that moves the number. Cut every feature that doesn't. Enough, my scope-cutting skill for Claude, exists for this step.
- Get the prompts right once. A vague prompt means retries, and retries are paid tokens. PrePrompt rewrites a prompt with the context and format added before it's sent.
- Measure for 30 days, then decide. Compare against the 30 days before launch. Keep it, fix it or switch it off.
Where does AI pay off fastest for a small business?
| Job | What AI does | Number to measure |
|---|---|---|
| Lead intake | Asks the right questions and writes up a brief | Briefs completed per week |
| Support triage | Sorts and drafts replies to common messages | Time to first reply |
| Weekly reporting | Turns data you already have into a summary | Hours saved per week |
| Product data for AI search | Cleans titles, attributes and descriptions | Products eligible in AI channels |
These are the jobs where the output is clear and the before-and-after is easy to count. That is why they pay off before the big platform projects do.
If you have one job like this in mind, I build MVPs and AI-powered web apps from $2,400 to $15,000, and I finish and harden apps started in Lovable or Bolt from $2,000, on my published pricing. Describe the job and I'll tell you whether AI is worth it for that job.
FAQ
Frequently asked questions
Is AI paying off for most companies?
Not yet, according to Apollo chief economist Torsten Slok. His October 24, 2026 note found revenue per worker and profit margins flat for S&P 500 companies outside the Magnificent 7, and no sign of AI lifting productivity growth so far.Does Anthropic really spend more on AI than on salaries?
Business Insider reported that Anthropic's internal AI use, valued at retail prices, now exceeds its employee pay. Anthropic runs its own models, so this is a valuation at list prices rather than cash paid to a supplier.How do I measure the ROI of an AI project?
Pick one number before you build, such as hours spent on a task, leads collected or support tickets closed. Work out the AI cost per unit of that work at the provider's list price. Measure the number for 30 days after launch and compare it with the 30 days before.Where does AI pay off fastest for a small business?
In repetitive work with a clear output: collecting lead details, sorting support messages, writing weekly reports from data you already have, and cleaning up product data so AI search can read it.
Last updated November 07, 2026
