Smart technology at work: what the evidence actually says

It helps, sometimes a lot

The gains are real. In one of the first large field studies, 5,179 customer support agents given a smart technology assistant resolved 14 percent more issues per hour on average (Brynjolfsson, Li and Raymond, NBER and Quarterly Journal of Economics). Writing tasks show a similar pattern, with productivity up and weaker writers gaining the most (Noy and Zhang, Science, 2023).

The gains are uneven

Look closer at that support study. Novice and less skilled agents improved by 34 percent. Experienced agents barely changed (NBER). The tool seemed to spread the know-how of the best people to everyone else.

That is good news for onboarding. It also raises a question most business cases skip. If the tool is learning from your experienced people, what happens to them, and to the next generation who would have become them?

There is an edge, and you cannot always see it

A study of 758 consultants at Boston Consulting Group found that on tasks the technology was good at, people completed 12.2 percent more work, 25.1 percent faster, at noticeably higher quality. On a task just outside what it could do well, people using it were 19 percent less likely to get the right answer than those working without it (Dell'Acqua and colleagues, Harvard Business School).

The researchers called this a jagged frontier. The tool is brilliant at some things, weak at others, and the line between them is not obvious. People tend to trust it most precisely where they should be checking it.

Feeling faster is not being faster

This one surprised us. In a 2025 trial, 16 experienced software developers worked on real tasks with and without smart technology tools. Before starting, they expected to be 24 percent faster. Afterwards, they believed they had been 20 percent faster. In fact they took 19 percent longer (METR, 2025).

It is one small study in one kind of work, so we would not build a strategy on it. But it is a useful warning about relying on how people feel about a tool. The International Labour Organization makes a similar point at scale: workers report saving a few percent of their hours, but those savings have not yet shown up in measured output, earnings or employment (ILO, 2026).

The jobs picture: quiet, not dramatic

So far there is little sign of large-scale job loss (ILO). The change is showing up somewhere quieter. Payroll data from Stanford show that employment of 22 to 25 year olds in the most exposed occupations is about 19 percent below where it would otherwise be, with no similar gap for experienced workers. It is happening mainly through less hiring, not more firing (Brynjolfsson, Chandar and Chen, Stanford Digital Economy Lab, 2026).

For leaders, that is a talent pipeline question. The entry-level roles where people used to learn the job are the ones thinning out first.

Who is in control

Smart technology is also used to direct and monitor work, not just to help with it. A review of more than 1,100 studies set out how this happens: systems that restrict, recommend, record, rate, replace and reward (Kellogg, Valentine and Christin, Academy of Management Annals, 2020). The review also describes how workers resist, individually and together, through tactics the authors call algoactivism. That links straight to the health evidence on change, where losing control over your own work is one of the main routes to harm.

What we suggest instead

1. Start with the work, not the tool.

   Pick a specific task, ask who does it, what good looks like and where it goes wrong. Then ask whether the technology helps with that. Not the other way round.

2. Map the frontier for your own tasks.

   Run a small test on real work. Find where the tool is reliably good and where it is not, and write it down. Then tell people.

3. Measure output, not opinion.

   Time the task. Check the quality. Compare with and without. Self-reported time savings are a starting point, not a result.

4. Keep the experts in the loop.

   Use experienced people to check, coach and improve what the tool produces, and recognise that contribution. Their judgement is what keeps the quality up at the edges.

5. Protect the learning path.

   If entry-level tasks are being automated, decide on purpose how new people will build the skills that those tasks used to teach.

6. Be clear about monitoring.

   If a system records or rates work, tell people what it tracks, why and who sees it. Give them a say in how it is used.