The productivity case for AI usually counts the time saved producing something and ignores the time spent confirming it is right. In many workflows the second number is larger.
Why checking is expensive
Reviewing work is cognitively harder than producing it. Finding an error in something that reads correctly requires holding the whole thing in mind and testing it against knowledge you have to retrieve.
Machine output makes this worse because it is uniformly fluent. Human error usually comes with signals – hedging, awkwardness, a sentence that does not quite parse. Machine error arrives in the same confident register as everything else.
Where the calculation works
Tasks where verification is cheap or automatic. Code that has tests. Data extraction that can be validated against a schema. Translation into a language you can read. Anything where being wrong is immediately obvious.
Where it does not
Tasks where checking requires the same expertise as doing. If you must be an expert to spot the error, having a non-expert generate the draft has saved nothing and added risk.
This is why AI has landed well in software – tests verify – and awkwardly in fields where correctness requires judgement.
How to structure around it
Use it where verification is built in. Build verification where it is not – a checklist, a second pass, a defined review step.
Never let unverified output reach a customer, a regulator, or a decision. The failure cases that make the news are all instances of that rule not existing.
The honest accounting
Real productivity gains exist and are smaller than headline claims, because headline claims measure generation and ignore review. Any organisation measuring properly should count both.