The two-hours problem

Every vendor demonstration ends with the same slide: hours saved. The number is usually true and almost always misleading, because it reports half of a subtraction.

Suppose a customer-operations agent produces two hundred and forty responses a week and saves the team ten hours of drafting. Suppose a competent human needs two minutes to confirm each response is acceptable before it goes out, or after, in a sample. Two hundred and forty responses at two minutes is eight hours. The agent has not saved ten hours. It has saved two. If the check takes two and a half minutes, it has saved nothing. If it takes three, the agent is costing the team time while the dashboard reports a gain.

This arithmetic is simple, and it is invisible in most organisations for three reasons. The verification time is spent by existing staff, so it appears nowhere as a cost. It is spread across many people in small amounts, so no one feels it as a job. And the people spending it have no incentive to report it, because reporting it sounds like resisting the technology.

The research literature has begun to name this. Catalini, Hui and Wu, in a 2026 paper on the economics of advanced AI, describe the transition as a race between two cost curves: the cost of automating a task, which falls with computing power, and the cost of verifying the result, which is bounded by human time and expertise. Where the gap between them is wide, they warn of a "hollow economy," high measured output with weak human control. Inside a single firm the hollowing has an exact shape: agents deployed in work where nobody is checking, because checking was never budgeted.

The field experiments that are usually quoted in the agent's favour are consistent with this, read carefully. Brynjolfsson, Li and Raymond found customer-support agents about 14 percent more productive with an AI assistant, and the assistant's suggestions were reviewed by a human before they were sent, which is to say the verification was built into the design and paid for. Cui and colleagues found software developers completing roughly 26 percent more tasks with a coding assistant, inside a workflow where code review already existed. The gains were real because the checking was real. Remove the checking and you have not removed a cost; you have hidden it, and it will reappear as errors.

What follows for the organisation is a measurement rule, and the rule is the reason a Chief Workforce Officer is needed rather than another dashboard. Every agent is evaluated on four things: how much it produces, how good the output is by the standard applied to a person doing the same task, how often it causes an incident, and how many human hours it consumes in verification. The fourth measure decides whether the agent is productive at all, and it is the one nobody currently collects, because collecting it means asking the humans around the agent to account for their time in a way they have never had to.

Only a role with authority over both the people and the agents can ask that question and act on the answer. IT cannot, because the hours belong to another department. The business unit will not, because the answer might embarrass the purchase. Finance does not know to ask. The CWO's first intervention in most companies will be unglamorous: finding out, task family by task family, what the real net gain is, and discovering that for some agents it is large, for some it is zero, and for a few it is negative.

That discovery is not an argument against agents. It is an argument for deploying them where verification is cheap, investing in making verification cheaper where it is not, and refusing to count hours saved until the hours spent are on the same page. The two-hours problem is not a technology problem. It is an accounting problem, and accounting problems are solved by giving someone the job.