AI Changes Work Design, Not Just Productivity

A supplier risk assessment gets drafted by a model on Monday morning. A category manager checks the facts in it. Legal reads the liability language and clears that. Procurement sees two sets of initials and reasonably concludes the recommendation has been approved. By the time it reaches the person who signs, three people have reviewed the document and nobody has reviewed the judgment.

No individual failed. Everyone did the part they understood to be theirs. The failure sat between the assignments.

This is the change that matters, and it is not the one AI gets sold on. Where automation takes over routine execution, the human work that remains tends to concentrate around judgment, validation, exception handling, and deciding what happens when the normal flow breaks. Not always — sometimes the work fragments instead, or becomes harder to see, or quietly disappears. But the common pattern is that the leftover work is the difficult kind, and it lands on the people with the least room in their week.

In many places the constraint is no longer how fast work can be produced. It is how much of it can be properly looked at. How many outputs still need a human decision. How many exceptions need interpretation. How often something plausible has to be caught and rerouted before it spreads downstream. The work did not disappear. It got concentrated.

And it exposes whatever was already fragile.

Vague ownership becomes untraceable ownership. Weak handovers get faster without getting better. Informal escalation rules meet a volume they were never designed for. A review threshold nobody wrote down stays survivable at low volume. At ten times the throughput, the ambiguity stops being theoretical and becomes operational.

Automation does not create these problems. It inherits them. And then it scales them.

The symptoms are recognisable. Decisions reopen. Work comes back slightly altered. Review can expand rather than contract, as people begin checking one another's output without a shared rule for what each review is meant to establish. Escalation becomes the default way to resolve uncertainty. Something introduced as a productivity tool starts generating coordination problems that either did not exist before or were quiet enough to absorb.

What looked like a tool problem turns out to be a workflow problem.

Which makes the useful question a different one. Not what can be automated, but what has to be made explicit before it is. Who owns the decision. What gets checked, by whom, at what threshold. Which exceptions need a human. Where authority sits when the normal process stops applying.

These are design questions, and too many organisations reach them after go-live rather than before.

It is worth being precise about what the missing review was. Factual review asks whether what the document states is accurate. Legal review asks whether the language creates exposure the organisation will not accept. A judgment review asks something neither of those touches: whether the recommendation should be acted on at all, given the costs, risks and priorities the organisation is actually trading off this quarter. Three people can clear a document on the first two and leave the third unasked, and nothing in the file will show it.

Go back to the supplier assessment. It did not need better technology. It needed the workflow to distinguish factual verification from commercial judgment, so that initials against the first could not imply approval of the second, with one role accountable for deciding whether the recommendation should be acted on.

One explicit rule, embedded in the workflow before the model was switched on, would have made the missing judgment visible. Writing it down is only the start; the work is making the workflow enforce it, which usually means system states, permissions, a labelled approval step and somewhere for exceptions to go.

That is slower and less visible than the tooling. It also decides whether any of the speed survives contact with the organisation.