Teaching with AI: the strategies that actually work in schools
10-second summary
Adopt AI against a specific task, not as a platform. The rollouts that stick start with one weekly job every teacher hates — differentiating a worksheet, writing a quiz, drafting a parent email — and prove the time saving before anything else is introduced.
Most school AI rollouts fail in the same order. A tool is chosen over the summer. An INSET day covers it in ninety minutes. Enthusiasm is high in September. By half term, three teachers are still logging in and the rest have quietly returned to what they did before.
The failure is rarely the software. It is that the rollout asked teachers to adopt a platform when what they needed was help with a task.
Start with one task, not one tool
Pick a job that happens every week, that every teacher dislikes, and where the output is visible enough that success is obvious. Differentiating a worksheet for three reading levels is close to ideal. So is turning a text into a quiz, or writing the first draft of a parent email about a difficult conversation.
The reason this works is that it makes the benefit legible in one sitting. A teacher who watches a colleague produce three levelled versions of tomorrow's worksheet in two minutes does not need convincing about AI in education. They need the link.
Put staff before students
Staff-side use is the easier half of the problem and the better place to build competence. Nobody needs a safeguarding review for a teacher generating a comprehension exercise. Nobody needs to write to parents about it. The school accumulates real experience of what these tools are good and bad at before it has to answer harder questions about student use.
Schools that invert this order usually spend their first term writing policy instead of teaching, and arrive at student use with no internal expertise to draw on.
Feed it your curriculum, not a topic
The most common complaint about AI-generated material is that it is generic. That is almost always an input problem. Output quality tracks input specificity closely, and the inputs that matter are unglamorous:
- The exact standard or objective, not the topic name.
- The reading level you actually need, which is often below the year group.
- The misconceptions you expect, so the questions probe them rather than avoiding them.
- The format — printable, one page, no answer key on the same sheet.
A tool that already holds your national curriculum standards removes most of this typing, which is the difference between a teacher using it weekly and using it twice.
Treat every output as a first draft
The teachers who get the most out of these tools are not the ones who trust them most. They are the ones who read the output quickly, keep two thirds, and rewrite the rest. Framing AI as a first-draft machine rather than an answer machine sets the right expectation and removes the disappointment that follows an imperfect result.
The goal is not a perfect worksheet with no teacher involved. It is a decent worksheet in two minutes instead of forty.
Make one person the answer to “who do I ask?”
Adoption stalls at the first small problem nobody can solve. Naming one member of staff — not necessarily the most senior, usually the most curious — as the person to ask removes that stall. Give them a period a fortnight for it and it will pay for itself in the department.
Measure hours, not logins
Usage dashboards tell you a tool was opened. They cannot tell you it helped. A better measure is to ask six teachers what a specific weekly task cost them before and after, and whether the output was good enough to use without heavy editing. Two honest answers to that are worth more than a term of analytics.
Where to begin this term
Choose one department, one task, and six weeks. If the department is still using it in week six without being asked, expand. If it is not, the tool was wrong or the task was. Most schools find their footing with lesson planning or quiz generation because both produce something a teacher can hold up in a meeting.
Frequently asked questions
What is the best way to start using AI in a school?
Pick one recurring task that every teacher already dislikes and solve only that. Differentiating a worksheet for three reading levels is a good candidate because the pain is weekly, the output is visible, and success is easy to judge. Broad platform rollouts fail because they ask teachers to change many habits at once for a benefit they cannot yet see.
Should students or teachers use AI first?
Teachers, almost always. Staff-side use has a clear time saving, no safeguarding exposure, and no parent communication burden, so it builds confidence and internal expertise before the harder student-facing questions arrive. Schools that start with students tend to spend their first term on policy rather than practice.
How do we stop AI from producing generic, off-curriculum material?
Give it your curriculum rather than a topic. Output quality tracks input specificity almost exactly: the year group, the standard or learning objective, the reading level, the misconceptions you expect, and the format you need. A tool that already holds your national curriculum removes most of this work.
How much training do teachers actually need?
Less than most schools plan for, if the tool is task-shaped. A teacher who sees a colleague generate a differentiated worksheet in two minutes needs a link, not a training day. Long INSET sessions on AI in general tend to produce interest without adoption.
How do we know whether it is working?
Measure planning hours, not logins. Ask a small group of teachers what they spent on a specific weekly task before and after, and whether the output was good enough to use unedited. Usage dashboards tell you a tool was opened; they do not tell you it helped.