Everyone Has an AI Copilot. Does Everyone Need One?

At some point, every workplace tool seemed to acquire an AI copilot. Email, meetings, spreadsheets, sales software, coding. If it has a screen and a subscription fee, there is a decent chance somebody has added AI to it. The promise is difficult to argue with: less admin, faster work, better decisions. And unlike some technology promises, this one is supported by growing evidence. AI can make people more productive in the right circumstances.

A large study published in The Quarterly Journal of Economics followed more than 5,000 customer-service workers and found that those using generative AI resolved around 15% more customer issues per hour. Less experienced employees improved the most, suggesting that AI can be particularly powerful when people are dealing with recurring questions, searching for information or learning how to perform a task well.

That is meaningful. It is also where a perfectly reasonable conclusion has been stretched into a much bigger one. “AI can make certain tasks faster” has rather quietly become “AI makes work faster.” The first statement is supported by evidence. The second assumes that producing something quickly and completing useful work quickly are the same thing.

They are not.

AI is extraordinarily good at making the beginning of work disappear. A blank page becomes a draft in seconds. A long document becomes five bullet points. An email you did not want to write suddenly exists. There is something deeply satisfying about watching twenty minutes of visible effort collapse into a prompt and an answer.

The output may appear instantly, but the work is not necessarily finished. Someone still has to check whether it is accurate, useful and worth acting on. Sometimes that takes less time than starting from scratch. Sometimes it takes more. AI has reduced the effort it takes to produce something. It has not reduced the effort required to decide whether that something is good.

That distinction became unusually visible in a 2025 experiment by research organisation METR. Experienced software developers were asked to complete work on projects they already knew well, sometimes with access to AI tools and sometimes without them. The developers expected AI to make them faster. Instead, they took 19% longer when they used it. Even more strikingly, after completing the experiment they still believed AI had made them about 20% faster.

The study was deliberately narrow, and it would be a mistake to conclude that AI generally makes developers slower. The technology has also improved significantly since the experiment. What makes the finding interesting is the gap between feeling productive and being productive. AI delivers output so quickly that it can create a powerful sense of acceleration, even when the total task takes longer.

And that points to a larger problem with the way workplace AI is often discussed. We count the minutes it removes much more enthusiastically than the minutes it creates.

Workday's 2026 research found that 85% of employees using AI said it saved them between one and seven hours per week. That sounds like an extraordinary productivity story until the rest of the calculation appears: around 37% of the time saved was being lost again to checking, correcting, rewriting and clarifying AI-generated work.

The work does not always return to the person who created it either. Sometimes it simply moves.

Researchers from BetterUp and Stanford have called low-quality AI-generated work passed between colleagues “workslop”: documents, summaries, presentations or other outputs that look finished but leave the recipient to do the thinking that should have happened before they received them. Their research found that workers spent close to two hours dealing with each instance.

This is where the usual idea of “time saved by AI” becomes much less straightforward. If AI saves one employee fifteen minutes drafting something but creates thirty minutes of checking, clarification and correction for somebody else, the employee became faster. The organisation did not.

The work simply changed desks.

That may be the productivity metric companies are missing. The unit that matters is not the prompt, the draft or even the individual employee. It is the finished piece of work from beginning to end.

This sounds painfully obvious, but it changes the question companies should be asking. Ten thousand employees having access to AI is not an outcome. Neither is one million prompts submitted in a quarter. Those numbers tell us that a technology is being used, not whether it is doing anything particularly useful.

Did the work take less time from start to finish? Did the quality improve? Did customers get a better result? Did employees spend less time on work nobody wanted to do? Did the company actually save money?

Those questions are less exciting than announcing an AI rollout, but they are considerably closer to productivity.

None of this is an argument against AI. There are already areas where the gains are substantial, and there will almost certainly be more as the technology improves. The mistake is assuming that because AI is powerful, more AI must automatically mean more progress.

It does not need to be useful everywhere to be transformative. It only needs to be very useful where it belongs.

Sometimes that may mean putting a copilot into the workflow. But sometimes it may also mean leaving the perfectly functional spreadsheet alone.