The Trust Tax
Last week I asked an AI assistant which lightweight tripod to buy for hiking, and the answer came back in under a minute with a clean comparison table and four cited sources. Then I opened the process logs for the assistant to see what the system actually did while it worked, and there wasn’t a single web lookup in there. Four citations, and not one of them had been ‘found’ on the web before sharing. I wrote that up in detail this week in my LinkedIn newsletter, “Output Isn’t Evidence” , because an answer that lies about its own work is a problem all its own. Here I want to look at what that answer actually cost.
Almost none of that cost lived on the tool side; the tokens the answer burned were fractions of a cent. The real cost was the time I then spent figuring out whether any of it was true, which turned out to be the only part of the whole exchange that took actual work. The machine did the cheap thing in a minute, and I did the expensive thing afterward, by hand. That’s the line item most people are missing for their AI budgets.
Just about every conversation today about AI costs includes the same handful of things: the subscription, the per-token bill, the headcount you can cut, the time-savings. The Economist ran a June piece on companies scrambling to curtail soaring AI costs , reporting that corporate AI bills had jumped thirteenfold in a year. Even the pieces that look past the tool, like Forbes arguing AI costs more than the people it replaced , are still measuring the machine against a salary. Every one of them prices the compute and/or the headcount you swapped out. The machine isn’t where the cost is.
The cost sits in the operating model around the output. Who owns the prompts, who reviews what comes back, who sets the permissions, who keeps the logs, who runs the boring QA loop, etc. When you make output nearly free to produce, that work doesn’t disappear; someone still has to do it. And that someone, the person now deciding whether the free output is worth trusting, is not cheap, and there is no line for them in the deck the vendor handed you.
Your organization’s whole way of checking work was built for a slower world, where a second set of eyes usually rode along for free because output was expensive to make in the first place. I wrote about how that free checking disappeared in The Verification Problem . What I didn’t do in that article was add up the bill.
That bill is the trust tax. You pay it every time a machine hands you something that looks finished, and you pay it in the one currency AI was supposed to give back: skilled human attention.
The reason it belongs on a CFO’s desk and not just an engineer’s is that it grows with scale instead of shrinking. There is an old idea from a nineteenth-century economist named William Stanley Jevons , who noticed that when steam engines got more efficient and coal got cheaper to use, England didn’t burn less coal, it burned far more, because cheap coal meant people kept finding new uses for it. Make a thing cheaper and consumption climbs, which is why the AI world spent this year obsessed with Jevons .
You can see it in law already: as AI drives the cost of document review down, the American Bar Association points out that the effect is not less legal work but more, because reviews that were never worth the old price suddenly are. The per-unit cost fell and the total bill rose.
Cheap output works the same way inside a company. When an analyst can produce ten reports in the time one used to take, they don’t produce one and go home, they produce ten, and all ten still need a human to decide whether they’re real. Production got ten times cheaper and the verification load got ten times heavier at the same time, which is the opposite of the trade everyone assumes they’re making.
The small numbers say it too. AI voice agents are a clean example: the platforms advertise around five cents a minute , but the real production cost runs two to five times that once you stack speech recognition, the language model, text-to-speech, and telephony. One platform, Vapi, charges a flat thousand dollars a month on top of it for HIPAA compliance. The advertised number covers the tool but everything you have to build and watch around it is the actual bill.
The easy objection is that verification is cheap, so just spot-check it, or better, let a second AI check the first one. I understand the appeal but it isn’t that easy (or cheap). Go back to my four fake citations: the fabricated answer and an honest one built from real research came out of the same model, and side by side the fake one looked better, cleaner and more confident, four sources against the honest version’s careful hedging.
You cannot catch a process that never happened by reading the thing it produced, because a lie and the truth can be word for word identical on the page, and the only place the difference lives is the record of what the system actually did. Reading that record is work. Handing the job to a second AI only moves the trust problem down a layer and gives you another machine you also have to verify, which compounds the tax rather than removing it.
I use AI every day, and the cheap execution is real and worth having. But the invoice only ever shows part of the price. Every time you put AI to work, you also take on the cost of deciding whether to trust what it hands back, and that cost is paid in the time of your best people, the ones who can tell a real answer from a confident fake. The more output you generate, the more of that judgment you have to buy, so the savings you booked on the tool quietly come back as a labor bill somewhere else on the org chart.
Most companies aren’t tracking that bill. They compare the price of the AI to the salary it replaced, call the difference savings (until the price far exceeds the salary), and never count the hours their team now spends checking the work.
Before you sign off on the next rollout, ask the question the vendor won’t: who checks this, how long does it take them, and what is their time worth? That number is the trust tax, and you’re already paying it…so put it in the budget.