AI tools for teachers: what actually saves time in a real week

Almost every list of AI tools for teachers opens with lesson-plan generation. It is the wrong place to start. Lesson planning is the part of the job most teachers are fastest at and most protective of, and a generated plan usually needs so much rewriting that the time saved is negative.

The hours actually disappear somewhere less glamorous: producing three versions of the same worksheet for three reading levels, writing the twenty-eighth comment on the same misconception, chasing a resource that exists somewhere in a shared drive, and converting material between formats. Those are the tasks worth automating.

1. Differentiation — the highest-value use

Take one worksheet and produce a scaffolded version, an extension version and a version with the reading level dropped two years. This is mechanical, high-volume, and the output needs only a quick check rather than a rewrite. It is the single clearest win.

A local model handles this well because the source document never leaves your machine — which matters when the worksheet contains named student examples or material you do not have redistribution rights to.

2. Feedback banks, not finished feedback

Generated comments on a specific child’s work tend to be bland and occasionally wrong, and students notice. What works better is generating a bank of feedback for the misconceptions you know appear in a topic, then selecting and adapting.

You keep the judgement, the AI removes the typing. Marking a set of thirty then becomes selection plus personalisation rather than composition from scratch.

3. Format conversion

None of these are impressive demos. All of them are ten to twenty minutes each, several times a week.

• A photographed page of a textbook into editable text you can adapt.
• A set of notes into slides, or into a printable revision sheet.
• A recorded lesson into a transcript, then into a summary for absent students.
• A mark scheme into a set of practice questions that target the same criteria.

4. Quiz and retrieval-practice generation

Generating retrieval questions from material you have already taught is low-risk, because you can verify every answer against the source in seconds. Generating questions from material the model invented is high-risk. Always feed it your own text.

What not to hand over

• Anything summative that goes on a report without you reading every word.
• Safeguarding or pastoral notes — those need a human and an audit trail.
• Assessment of a student you have not taught, where you cannot sanity-check the output.
• Anything where a hallucinated fact would reach students unchallenged.

The data question

The reason many schools stall on AI is not scepticism about usefulness, it is that pasting a child’s work into a cloud tool means transmitting it to a third party. That is a genuine data-protection question, not paranoia.

On-device tools remove the question rather than answering it. If the model runs on the teacher’s laptop, there is no transfer to disclose, no retention period to negotiate and no training-data clause to read. That is why Grout runs its models locally by default.