AI Agents for Teachers: What They Can Take Off Your Plate in 2026

Search this topic and every result lands in the same place. Generate a lesson plan. Generate a worksheet. Generate a quiz, three differentiated versions, a starter activity. All of it useful, all of it aimed at the planning half of the job.

The other half is admin and messages, and almost nobody writes about it. Chasing the four permission slips that never came back. The email to the parent who wants to know why the grade moved. Turning a week of scrawled behaviour notes into two paragraphs a form tutor can actually use. Pulling three assessment spreadsheets into one sheet for a meeting on Thursday.

That is the agent-shaped list. It repeats, it follows a rule, it has a recipient, and it needs your signature more than your judgement.

I read Gallup's two 2026 teacher surveys on 27 September 2026 while putting this together. Six in 10 teachers already use AI tools for their work, and three in 10 use it at least weekly. Gallup's second release finds that most teachers receive no formal guidance on AI use.

So here is the split as I would draw it: what an agent is shaped to take, what it must never touch, and what changes when your school's IT lead is in the room.

Key takeaways

• The under-served half of the week is administrative. Permission slips, parent messages, behaviour summaries, cover admin, data collation. Those repeat on a rule, which is what an agent is for.
• An assistant answers; an agent runs. If you open the same chat every Friday and type a near-identical prompt, you have found a job to set once.
• Gallup's 5.9 hours is a self-estimate. Weekly AI users estimate that saving by the half hour, task by task. Gallup puts it at the equivalent of six weeks across a 37.4-week school year.
• Four things stay human: the grade that goes on a record, anything safeguarding-adjacent, SEN and IEP judgements, and anything a parent could formally challenge.
• Settle the account and the data before the tool. Gallup found most teachers get no formal guidance on AI use, so the ask has to come from you.

What an AI agent is, for a teacher who has only used ChatGPT

You already know the chat window. You type, it answers, you close the tab. That is an assistant, and it is genuinely good at the things you ask it once: rephrase this for a Year 8 reading level, give me four ways into osmosis, tighten this email.

An agent is the same underlying model pointed at a job instead of a question. You describe the outcome once, connect the accounts it needs, and it runs on a trigger or a schedule. Nobody re-types anything. The output arrives where you asked for it.

The distinction shows up in how the two fail. An assistant fails silently, every week you are too busy to open it. An agent that is set up properly fails where you can see it, in your inbox, which is a failure you can act on.

Three tells that a job is agent work rather than chat work:

• It repeats on a calendar. Every Friday, every half term, the week before reports.
• It has a rule you could write down. "Everyone with two or more missing pieces, grouped by tutor group, one paragraph each."
• It ends in a recipient. A parent, a head of year, a spreadsheet somebody else opens.

If a job has none of those, keep using the chat window. It is faster and you will not have to configure anything.

"Agentic AI in education" is the sector-press name for this shift from answering to doing, and the school-wide version is in the piece on AI agents across education. For the category boundaries without the education framing, read what an AI agent actually is and AI agent versus workflow automation.

The jobs worth automating

Seven jobs, and the column I care about is the third one: what still needs you after the agent has run. Capability tables usually stop at column two. That third column is the honest part.

Job shapes, not product claims. I have not tested a tool against this list, and whether a given agent can reach a given school system depends on what your school allows it to connect to.

The pattern in the right-hand column is the whole point. The agent gathers, formats, drafts and chases. You decide. That division holds across all seven rows.

Gallup's qualitative data point the same way on where the recovered time goes: more nuanced student feedback, individualised lessons, writing emails to parents, and getting home at a more reasonable hour. A modest, believable list. Nothing in it about becoming a different teacher.

On the headline number, read the footnote. Gallup says weekly AI users estimate they save 5.9 hours a week, and that the estimate was made by the half hour on a task-by-task basis, then summed across non-overlapping tasks. Over a 37.4-week school year Gallup calls that the equivalent of six weeks, and names the effect the "AI dividend". A self-report, not a measurement. I read it as evidence that teachers feel a difference rather than proof of its size.

The jobs to keep human

This is the part of the page I would not compromise on, so it sits above the FAQ rather than tucked under it.

Four categories of teacher work should never be delegated to an agent, however good the drafting gets. The reason is the same in all four: you are the person who has to stand behind the decision, in a room, to someone who disagrees with it.

• Any grade that goes on a record. Coursework marks, report grades, predicted grades, anything feeding a decision about a set or an entry. An agent can pre-sort and draft. The number is a professional judgement and it carries your name.
• Anything safeguarding-adjacent. A disclosure, a pattern that worries you, a change in a child you cannot explain. This goes through your school's safeguarding process and a named human being, immediately, and nothing about it belongs in a tool. Not as a summary, not as a draft, not as a note to yourself.
• SEN and IEP decisions. Provision, access arrangements, targets, and the judgement that a strategy is not working. These depend on knowing a specific child over time. Summarising evidence is fine. Recommending provision is not.
• Anything a parent could formally challenge. Sanctions, exclusion paperwork, an appeal against a mark, a complaint response. If there is a plausible future where you read the message back in a meeting, you write it.

There is a simpler version of that rule and it travels well. An agent belongs on the work around the judgement, never on the judgement itself.

One more line, because it is the failure I would expect to see first. A drafted message you send without reading is now your message, with your name on it, and "something else wrote it" is not a defence anybody accepts. Approval is not a formality here. It is the workflow.

The same logic runs on the other side of the desk, worth knowing when a pupil asks: the guide to AI agents for students draws the integrity line in the same shape, around the thinking rather than through it.

Schools and districts: what changes when IT is involved

Everything above assumes you are choosing for yourself. The moment a school is choosing, the questions change, and they stop being about features.

Start with the one that has no workaround. Children's personal data does not go near any tool until your school has said it can, in writing, with someone named as the owner of that decision. Class lists, behaviour records, assessment data, SEN information, anything identifying a child. I will not name a regulation here, because the rules that bind you depend on where you work and they move. Read your school's own data-protection position, and ask the person who wrote it.

The move while you wait for that answer: pick the jobs that need no identifiable pupil data at all. Lesson-prep research, resource drafting, your own rota and planning admin, a parent-email template with no names in it. Real time in that list, and none of it turns on pupil data, which is the part that has to wait for an answer.

When the conversation does happen, these are the questions I would put on the table:

• Who owns the account? A tool bought on a teacher's personal login walks out of the building when the teacher does, and takes its history with it.
• Where does the data go, and for how long is it kept? Ask for the answer in a sentence you could repeat to a parent.
• What does the audit trail show after a run? If nobody can reconstruct what the agent read and what it sent, the tool is not ready for pupil data.
• Does a human approve anything that leaves the school? This should be a setting, not a promise.
• What happens on the day it gets something wrong? Who notices, who tells the parent, who fixes the record.

Gallup's second 2026 release is the reason this section exists at all. Its finding is that most teachers receive no formal guidance on AI use. The same release shows the form of the guidance matters: 69% of teachers who receive formal guidance on using AI to make student materials say it encourages them to use AI, against 51% of those who get informal guidance. Eighteen points between a stated position and a corridor conversation.

So the bottleneck is not teacher willingness. Six in 10 are already using these tools. The missing piece is an institutional answer, and somebody has to go and ask for one. For the vendor-facing version of that meeting, take the AI agent security checklist; what actually stays private covers setups that keep data on your side of the line.

Setting one up in practice

Here is how I would sequence it if I were starting on a Sunday evening with one job in mind. Four steps, and the first one is the only one people get wrong.

Pick the job you resent most

Not the biggest job. The one that annoys you out of proportion to its size, because that is almost always a job with a rule in it, and a rule is what an agent needs. The job people describe that way is usually a chasing job. Slips, forms, missing work, replies.

Then check it against the pupil-data question above. If the job needs names, it waits for your school's answer. If it needs a class list, ask whether a tutor-group label would do instead.

Write the instruction like a cover note

You are not configuring software. You are writing the note you would leave for someone competent who does not know your class: what to look at, what rule to apply, what the output looks like, who it goes to.

The shape I would use: "Every Friday at 4pm, check my missing-work sheet, find anyone with two or more missing pieces, and draft one short email per pupil to the address on file. Neutral tone, name the pieces, offer next Tuesday lunchtime. Put them in my drafts. Do not send."

Those last three words do a lot of work.

Keep yourself in the approval seat

Set it so nothing reaches a parent without you reading it. Every time, not just for the first fortnight. Reviewing a batch of drafts is quick next to writing them, and it is the difference between a tool that helps and a tool that writes cheques you cash in a meeting.

Approval also teaches you the instruction. The first batch shows which part of the brief you wrote badly, and you fix the sentence instead of the output.

Put it on the day it is needed

A job that runs Friday afternoon lands while you can still act on it. The same job at 9am Monday is one more thing to read during registration. On Gravity the schedule is part of the sentence you wrote, and the agent keeps running it without you opening anything.

Cost, so you can size it before applying: the first agent is free and does not ask for a card. Pro is $5 a month during the alpha. If you run out, you buy more usage. Current figures live on the pricing page at /pricing.

Start with one. One reliable agent on the job you resent most beats four subscriptions and a Sunday spent chasing by hand.

FAQ

Sources