An AI assistant can be genuinely helpful.
It can find the missing document, draft the first version, explain the obscure spreadsheet, prepare someone for a difficult conversation, or remove the little bits of administrative friction that make a job feel like it has been designed by people who dislike humans.
It can also record what people ask, how long they take, where they hesitate, which suggestions they accept, what they revise, and how many outputs they produce.
Those are not the same offer.
The first says: here is more capability.
The second says: here is a finer-grained way to measure you.
Most workplace AI discussions glide past this distinction because “data-driven” has become a polite phrase. But anyone who has worked under a bad metric knows that measurement changes the work long before it improves it.
The useful tool can become the boss
Think about what happens when a system knows not only the final report but the route someone took to produce it.
It can see how many drafts they made. Whether they checked the source. How often they asked for help. Which task took longer than average. Whether they used the recommended workflow. It can make a manager’s dashboard look wonderfully complete.
It can also reward the person who produces the easiest-to-count version of the work.
The messy conversation that prevents a mistake may disappear from view. The extra hour spent teaching a new colleague may look inefficient. The person who slows down because they notice that the model is confidently wrong may appear less productive than the person who sends the answer on.
That is how a support tool becomes an algorithmic manager: not when it gives an instruction, but when the system around it starts treating its trace of the work as the work itself.
The ILO has warned that AI at work can create psychosocial risks through intrusive monitoring, work intensification, reduced autonomy, and uncertainty about data use. This is not a reason to ban useful tools. It is a reason to stop pretending that implementation is only a technical question. ILO
The promise of time is easy to steal
Every AI business case has a version of the same sentence: this will save people time.
Fine. Then what happens to the time?
There are at least three possibilities.
It can become better work: more care, better decisions, more contact with customers, more learning, less repetitive administration.
It can become more work: the same person is now expected to produce more reports, answer more messages, manage more cases, and be more available because the system has made them “efficient.”
Or it can become measurement: every reclaimed minute becomes evidence that the next target should be tighter.
The model cannot decide which of these happens. Management does.
That is why people are often less enthusiastic about “productivity” than the people selling the tool expect. They have met productivity before. It was not always a gift.
A simple test
Before adding AI telemetry to a workflow, ask four ordinary questions.
- What information is being collected about the person, not just the task?
- Who can see it, and for what decision?
- What work that matters will this data fail to capture?
- What does the employee get in return for being more legible to the company?
If the answer to the last question is “nothing, but the dashboard will be excellent,” the design is already telling you what it is for.
Good AI should make people more capable without making them feel permanently observed. It should give them room to exercise judgment, not a more sophisticated way of proving that they did not take too long to have any.
Otherwise we have not built an assistant.
We have built a stopwatch with a chat interface.