AI spend per employee: why the falling metric misleads

September 9, 2026 · 5 min · Free, with full transcript

Enterprise AI spending per employee dropped at top firms in August 2026. Does that mean companies got efficient, or that adoption quietly stalled? The position here is that the ratio itself is the problem: token prices at the frontier labs have fallen by more than half since early 2025, and headcount grew at the heaviest adopters, so the per head figure falls even when absolute budgets rise. You get a way to tell a real efficiency gain from a dying pilot, using weekly active users divided by paid licenses per team, the Google AI 30 percent shipping claim checked against O'Reilly Radar survey data, and the Klarna customer service reversal as a warning about cutting cost before fixing the workflow.

Listen to the podcast

4 min

Chapters

Key takeaways

  • Split spend per employee into its parts before reading it: falling token prices and rising headcount both push the ratio down.
  • Track weekly active users divided by paid licenses per team over time instead of dollars per head.
  • Treat unused licenses as a stall, not a saving: 120 weekly logins against 500 seats means 380 people abandoned the tool.
  • Read a healthy drop as rising usage with flat or rising total spend and falling unit cost.
  • Redesign the workflow before cutting cost, as Klarna's customer service rehiring showed.
Read the full transcript

Host:Welcome to the Leaders Insights Podcast. Today's episode AI spend per employee is falling. Efficiency win or adoption stall? Enterprise AI spend per employee dropped in August 2026. Everyone's calling it either the efficiency dividend or the beginning of the crash. You think both camps are wrong? Convince me.

Expert:They're both looking at a ratio and pretending it's a story. Spend per employee is spend divided by headcount. If either number moves, the ratio moves. And right now, both are moving

Host:at once. So the number went down and told you almost nothing. Break it apart then. What's actually happening underneath?

Expert:Two things. First, prices collapsed. The cost per million tokens, a token being roughly three quarters of a word the model reads or writes, has fallen by more than half at the frontier labs since early 2025. So you can do the same work for less money. That's not a slowdown. That's deflation. And the second thing? Headcount grew at the firms doing the most AI. When your denominator, the number of employees, swells because you're hiring. Spend per person drops even if total spend climbs. The top quartile of adopters increased absolute AI budgets in August. The per head figure fell anyway. It's an arithmetic mirage in…

Host:Google AI put out numbers suggesting productivity gains are driving the efficiency. You buy that?

Expert:Google AI says teams using their tool ship features around 30% faster. But they sell the models, so that figure has a thumb on the scale. Cross-check it against O'Reilly Radar, which surveys practitioners who don't have anything to sell you. O'Reilly's read is messier. Real gains in coding and drafting. Near zero gains in workflows, nobody bothered to redesign. The tool got cheaper. The habits didn't change.

Host:So the falling number could be hiding a genuine adoption stall in some places.

Expert:Exactly. And that's the part both camps miss. Averages lie. If 10 teams cut spend because they gave up quietly, and 10 teams cut spend because they got ruthlessly efficient, the average looks identical. Same number, opposite meaning. You have to look at usage, not dollars.

Host:How does a manager tell the difference from inside their own company?

Expert:Look at active seats versus paid seats. If you bought 500 licenses and 120 people log in weekly, your spend per employee looks lean and healthy because 380 people abandoned the thing. That's the stall wearing efficiency's clothing. Cheap because unused is not a win.

Host:That's the uncomfortable version. Give me a company example that's real.

Expert:Klarna made noise in 2024 about replacing customer service headcount with AI, then quietly rehired humans when quality slipped. The lesson wasn't AI failed. It was that they cut cost before they fixed the workflow. So the savings evaporated the moment customers complained. Cost per employee looked brilliant for about two quarters. Then it didn't.

Host:So what does a smart spend per employee drop actually look like?

Expert:It looks boring. Total spend flatter up, usage up, and the per head number falling because the price per unit of work dropped and people are using more of it, not less. Rising volume, falling unit cost. That's a company eating the deflation and putting it to work.

Host:And the dangerous drop?

Expert:Spend down, usage down, and someone in finance calling it discipline. That's not discipline. That's a pilot project dying in a spreadsheet where nobody has to watch it.

Host:Give me the dry version of that.

Expert:A dead subscription is the most cost efficient subscription you'll ever own. Zero usage, predictable bill, terrible outcome.

Host:So the headline metric is basically useless.

Expert:It's worse than useless because it's confident. It gives aboard a single tidy number that feels like insight and hides whether your investment is compounding or rotting. If you only track dollars per head, you've optimized for the one thing that can't tell you if the strategy works.

Host:And what should someone track instead starting tomorrow?

Expert:One number. Weekly active users divided by paid licenses per team tracked over time. If that ratio climbs while your bill holds steady, you're winning on efficiency. If the ratio falls, you're not saving money. You're abandoning the tool in slow motion. And the falling spend is the receipt.

Host:Usage over seats, not dollars over heads. That's the takeaway.

Expert:That's the whole game.

Host:Sources for today's episode, the decoder, Google AI vendor, AI lab, TechCrunch AI, K nuggets, ours, Technica AI, O'Reilly radar. That's your five. The full AI track, structure, no fluff, is at MBA-training.com.

Read and go deeper