Why AI still cannot hold a job

I was very surprised to learn that nine in ten of nearly 6,000 executives say AI has not changed their firm's productivity in three years (NBER, 2026). The problem is not that AI does not work. For example, when one company gave more than 5,000 support agents an AI assistant, they resolved 15 percent more issues per hour (Quarterly Journal of Economics, 2025).

So where does the benefit go? A job is more than a list of tasks: someone must decide what comes next, give the result to a colleague, know when it is done and answer for it when it is wrong. An insurer does not pay a claims handler to read the vet bill. It pays her to settle the claim and close the file.

AI gets the reading. The settling, where the money is, stays with a person, because the risk team will not hand that decision to an AI.

Where does the benefit go?

The gap

In March 2026 Anthropic published a study of what AI could do at work and what it actually does (Labor market impacts of AI, Anthropic, March 2026). It started from an earlier estimate of which tasks a language model could in theory do at least twice as fast: 94 percent of the tasks in computer and maths jobs and 90 percent in office and administrative jobs. Then it counted what people really use Claude for, and in computer and maths jobs the share was 33 percent.

The same paper looked for the damage. Since late 2022 it found no rise in unemployment among the workers most exposed to AI. A forecast of what AI could do and a count of what it has done sit in one paper, from an AI company, and they do not match.

The money

The gap shows up in company results too. Nine in ten executives in the US, the UK, Germany and Australia say AI has changed neither their firm's employment nor its productivity in three years (Firm Data on AI, NBER, February 2026). The survey asked nearly 6,000 of them between November 2025 and January 2026, and it was run by the Atlanta Fed, the Bank of England, the Bundesbank and Macquarie University.

Workers tell the same story. In Denmark, two years after people in AI-exposed jobs started to use chatbots, neither their pay nor their working hours had changed by more than 2 percent (Still Waters, Rapid Currents, NBER, 2026).

The national accounts tell the same story. Over the past year, investment in AI hardware and software added 0.70 percentage points to US growth of 2.1 percent. Once the imported hardware is subtracted, it added about 0.14 points (How much does US growth depend on AI investment?, Flossbach von Storch Research Institute, August 2026). Both numbers measure money spent on buying AI. Neither measures what AI does once it is switched on.

Is AI taking the jobs instead?

US employers have blamed AI for 120,136 job cuts this year, but those are announcements, not people who have already left (Challenger, Gray & Christmas, October 2026). Is the AI now doing the work those people did, or did the budget move to AI before the work did? Andy Challenger, who counts the cuts, said in April that whether or not AI replaces the jobs, it replaces the money for them.

I will leave the verdict to you. The research has not found AI pushing employed knowledge workers out of their jobs (Brynjolfsson, Chandar and Chen, August 2026). What it finds is fewer young people hired into them.

Who is accountable when AI decides?

A job is a series of decisions. Every decision that moves money, changes a contract or answers a customer carries a risk, and that risk needs an owner. The owner is never the AI.

When Air Canada argued that its chatbot answered for itself, the tribunal made the airline pay. Insurers draw the same line from the other side. In November 2025 WR Berkley proposed to exclude any actual or alleged use of AI from its corporate policies, and other large insurers asked for similar exclusions (Financial Times, November 2025).

The law draws it too: from 20 November 2026, a borrower in the EU can ask a person to review a credit check done by a machine (Consumer Credit Directive, Article 18).

So the company carries the risk alone, and its risk team does the sensible thing: it caps what the AI may decide. The agent stays read-only, or a person approves every action. AI keeps the safe tasks, and the decisions that carry the money stay out of its reach. The cap is safe, and so is the job, because you pay for the automation and keep the queue.

Go back to the Anthropic paper. Its authors asked why real use trails what AI could do, and their answer was not only the models. Some tasks may spread slowly because of legal constraints and human verification steps.

Their example is a task rated as fully within reach of a language model: authorising drug refills and passing prescription information to pharmacies. They had not seen Claude do it once.

Reading the prescription is the easy part. The refill needs someone who answers for it if it goes wrong.

(open full screen)

Are companies only chatting with it?

Most companies still use AI as a chat window, not as a worker with a named task. The executives in that survey who used AI themselves did so for about 1.5 hours a week. A chat window makes one person faster at one task, and the rest of the job stays where it was.

Does the saved time go anywhere?

The hours AI saves often have nowhere to go. The European Central Bank found that the gain depends on whether the employer can put the extra time to use (ECB Blog, August 2026). Much of the work also waits in queues between the steps, which no model speeds up, as I wrote in Domain-driven AI transformation.

Is it too early?

New technology takes years to reach the numbers, because companies first have to rebuild how the work runs. That may be part of it, but it does not explain why tasks that AI could already do are not given to it.

The new hire

Of these causes, only responsibility has a known fix. None of this is new, because companies solved it for people long ago. A new hire in finance is not trusted with the company's money on her first day. She gets a limit, an approver above it and a record of everything she does, and the limit goes up as she proves herself.

We never trusted people either. We trusted the limits around them, as I wrote in Your newest employee skipped onboarding.

One task, end to end

Firms that hired AI skills faster than their peers grew their productivity about 1 percentage point a year faster (Canaries in the Gold Mine, NBER, August 2026). The authors trace the gain to lasting know-how that the firms built up by learning on the job. The gain came from what the firms learned, not from the tool alone.

So start with one task and give it to the AI whole. Take a refund. The agent reads the customer's message and the order, and proposes 140 euros.

Before any money moves, a separate set of rules checks the proposal. One refund may be up to 500 euros, and all refunds in a day up to 5,000 euros, whoever issues them: the agent, a vendor's agent or a person in the back office. Inside those limits the refund pays, and nobody looks at it. Above them it goes to a named person, with the message, the order and the proposal already attached.

Every decision, including every refusal, leaves a record that neither the agent nor its team can change. The risk team can sign this off, because the largest loss in a day is a number it chose.

Where the rules live

A check that runs after the money moves is a report, not a control. The refund works only if the rules sit outside the agent and check each refund before it pays.

A limit in the agent's prompt is a request, and a model can misread it or a customer can talk around it. A limit inside each system counts only what passes through that system, so the refund from the helpdesk and the refund from the vendor's agent never meet. An agent platform limits each agent, and a person issuing the same refund from the back office is outside it. A person on every action meets the rules on paper and brings back the queue.

The scorecard is from 5 ways to limit AI agents that move money, where each row is explained. Only a separate layer, which every actor must ask before it acts, holds all six. It is the one option where the agent works without a queue and the risk team can still sign.

Buy the agent. Own the policy.

The person who approved 300 refunds a day now decides the 10 that cross a limit, and she owns the limits. Her job did not disappear. It moved to the part of the work that needs someone to answer for it.

Back to the 33 percent

Anthropic says it will repeat its measurement. When that 33 percent rises, the interesting question will not be what AI learned to do. It will be what companies finally learned to let it do.

Because the bottleneck was never just capability. It was accountability.

---

This is what I build: stategram.io

Br, Silver