AI was supposed to be cheaper than a person

The pitch is always the same shape, and it always fits on a napkin. A person costs you $120k a year. The agent costs a few thousand in tokens. Swap one for the other and pocket the difference. Every "AI replaces X" business case is really that one subtraction.

The subtraction is wrong, and 2026 is where the bills proved it. Forbes ran the quiet-part-out-loud headline — AI costs more than the people it replaced — while the companies furthest ahead on adoption slammed on the brakes: Uber reportedly burned through its entire 2026 AI budget in about four months and capped engineers at $1,500 a month; Meta put hard caps on internal AI spend as consumption tracked toward billions.

You didn't remove the work. You moved it.

The napkin math assumes the job was "produce the output." It wasn't. The job was "produce the output and be accountable for it being right." The agent does the first half and hands the second half back to you. So the work doesn't vanish — it moves, and usually upstream, onto someone more expensive:

• The token bill is not fixed, and it's not small. A salary is a known number. Metered inference is a variable you don't fully control, and it spikes exactly when the agent is working hardest — retrying, re-reading, "thinking." Ask Uber's finance team.
• Someone senior now reviews everything. The junior you didn't hire was cheap and took ownership. The agent's output still needs a human to catch what it got confidently wrong — and that review is where the time actually goes now, on the calendar of your most expensive people.
• The failures that slip through are the priciest line. A person who ships a bad decision owns it. When the agent works 57% of the time, the other 43% reaches a customer, and cleanup — refunds, trust, the incident — dwarfs any token saving.

Automation doesn't delete labor. It converts salaried, accountable work into a metered bill plus senior oversight — and then hopes the sum is smaller. Often it isn't.

The efficiency turn is this realization, industry-wide

The whole "tokenmaxxing is over" mood of mid-2026 is just a lot of companies running the fully-loaded number for the first time. The winners aren't the ones who spent the most or the least — they're the ones who route work to the cheapest model that's actually good enough, meter every dollar against an outcome, and keep a human in the loop only where a mistake is expensive. That's not a retreat from AI. It's just accounting finally showing up to the party.

The bottom line

AI is genuinely cheaper than a person for the parts of a job that are cheap to get wrong. It is often more expensive for the parts that are expensive to get wrong — because those are exactly the parts that come back to a human. The napkin only ever counted one side.

Before you replace a role, price the whole thing: tokens at peak, senior review time, and the cost of the failures that reach a customer. If that number still beats the salary, automate with confidence. If you didn't run it, you're not saving money — you're just moving the bill somewhere it hasn't arrived yet.