AI Sellers Can't Prove the Product, So They're Buying the Labor

Hundreds of contract physicians grade ChatGPT's medical answers. They work across 49 languages — Kenya, Nepal and Brazil among them — and they have reviewed more than 700,000 responses since a misdiagnosis lawsuit forced the question of what the model is allowed to say about a body ^[1]. The program belongs to OpenAI. The people in it are checking the output of the company that pays them. That is not a safety program. It is one purchase inside a pattern. The AI industry has spent two years selling a workplace product it cannot price and cannot prove, and since spring its sellers have converged on the same two answers: stop selling the tool and own the outcome, priced against what labor costs — and buy the human checkers the tool turns out to need, at the cheapest professional rates on earth. The diagnosis first, and fast. MIT found 95% of enterprise pilots never touch the profit and loss, and the agents that do ship "fail for the same reasons new hires do" — the work they were meant to absorb was never written down ^[2]. One company's Anthropic bill was projected to jump from $400,000 to $1.4 million the moment it crossed a 150-seat threshold ^[3]. Long-term cost models, by the admission of one industry executive, "lack a logical basis" ^[4]. Nvidia's own vice president put the math bluntly.

For my team, the cost of compute is far beyond the costs of the employees. — Bryan Catanzaro

Still the spending climbs — up 110% — while the maturity score that says whether any of it is ready sits at 51 out of 100 ^[5]. So the sellers changed what they sell. Consulting is walking away from the billable hour, because AI's efficiency has made the hour obsolete, and from the fixed fee, because clients no longer trust a fee to show them a return — the consultants now take the financial risk on results ^[6]. The pitch has a number attached: labor budgets, roughly 13% of GDP. Sequoia's pick for the next trillion-dollar company is, in effect, a software firm dressed as a services firm ^[7]. Anthropic sells the capability — Claude embedded in Excel, PowerPoint and Slack — while positioning the buyer's own business knowledge, expertise and trusted relationships as the thing that completes it ^[8]. The model is the product. The human expertise is explicitly outside it. Then they changed what they buy. The physicians return here, one instance of a class. The product is probabilistic, so it needs a human anchor — the people running these companies keep saying so. Xoriant's chief executive is blunter about it.

The models themselves keep getting better, but the value companies actually derive barely moves. — Rohit Kedia

And the anchor is not free. The consultancy reporting the figure puts the savings from keeping humans in the loop at up to 70% of invoice-processing costs ^[9]. The physician who runs OpenAI's contract-checking program describes the job this way ^[1].

We are going to continue to march on towards improvement, as far as I am aware, indefinitely. — Rebecca Soskin Hicks

That checking work is real and growing — and it is being bought at deflated rates from the world's lower-wage professional rungs, the same rungs AI competition is now squeezing. India's IT services are getting cheaper as AI arrives: TCS's revenue down to $28.6 billion, Wipro to $9.9 billion, Infosys guiding just 1.5 to 3 percent growth ^[10]. China's urban youth unemployment sits near 19%, and after employers told officials AI could erase a third of existing roles, the state quietly pushed firms not to cite AI in layoff notices — the cuts go on, just unlabeled ^[11]. Across the white-collar economy, output is up 16% since late 2022 while employment has fallen about 2%, and the economist tracking the divergence describes software and IT firms that get far more output per worker while barely adding headcount ^[12]. The gains are real where they land — the Bank of England found productivity gains concentrated in the software and consulting firms that adopted AI ^[13], and Google's overall revenue rose 24% on enterprise AI demand ^[14]. The checking is real work that really cuts costs. What the record shows is overlap and deflated rates, not anyone saying they hire from the deflating pools because the pools deflate. One company puts the whole thing on a single org chart. TCS, the Indian outsourcer, is absorbing the employees of its own clients — Porsche and Best Buy among them — into its workforce as billable "AI-first" staff, the people who will turn those clients into AI-first firms while AI-led work stays around a tenth of TCS's own revenue ^[15]. The displaced become the billable. The rung the buildout cuts is the rung it hires from.