AI will not give you your evenings back on its own.

AI reduces teacher workload only where the task is high-volume, low-judgement and easy to check. Where any of those three fails, the supervision cost eats the saving, sometimes entirely. That is an unglamorous rule and it predicts outcomes better than any product page.

Which tasks give back the most time?

The ones where you already know the answer and the labour is in expressing it. You have read the books and you know what the class needs; writing that down twenty-eight times is the cost. You know what the trip letter has to say; the cost is producing it in the school's format again. You know the worksheet works; the cost is making three more versions of it.

• Drafting feedback from marking you have already done. High volume, low judgement, trivially checkable.
• Producing alternative versions of an existing resource. The thinking happened when you made the original.
• Routine correspondence with a fixed shape. Trip letters, reminders, standard notices.
• Turning notes into a document. Observations into a report, minutes into actions, a plan into a staff guide.
• First drafts of anything you were dreading starting. A bad draft you can fix beats a blank page you cannot.

Which tasks cost more time than they save?

Anything where verifying the output takes as long as producing it would have. This is the trap, and it is counterintuitive, because these tasks often look like the most impressive use of the technology.

• Producing subject content you then have to fact-check line by line. Checking a plausible-sounding history summary is slower than writing a true one.
• Anything requiring knowledge of a specific child. The tool has not met them, so you supply the context, and supplying it is the work.
• High-stakes documents. Guidance rules out AI for summative judgements and statutory plans anyway, and the review burden would be total.
• Tasks where you cannot tell whether the output is right. If you would not notice an error, you are not supervising, you are hoping.

Does AI actually reduce teacher workload?

It can, and there are figures circulating that claim to say by how much. Treat them carefully. Most come from short pilots, self-reported estimates, or vendors measuring their own product, and the ones with real methodology behind them are rarer than the citations suggest. We are not going to add a number of our own until we have run a proper study in real schools.

What is not in doubt is the underlying problem. Detailed written marking of every piece of work is not the most effective form of feedback, and it is consistently one of the largest contributors to workload. That mismatch existed long before AI and will not be solved by a tool alone.

What has to change alongside the tool?

The policy that generated the work. A marking policy requiring deep written comment on every piece is not fixed by drafting the comments faster, it is fixed by changing the policy. A school that adds AI on top of its existing expectations will find the time going somewhere else within a term, because nothing structural moved.

This is the least commercially attractive thing to say, which is roughly why it is worth saying. The tool is the smaller half of the intervention. Aevum Teacher Manager is in development, and it is built on the assumption that a school will also change what it asks of people.