AI’s useful autumn test: a better task, with a boundary leaders can defend

01 / Choose a task whose quality can be assessed

Choose a task whose quality can be assessed

Begin with one low-risk task and a clear purpose. A generic draft of an internal agenda may be easier to evaluate than a personalised pupil decision. The school should decide what information may be used, who checks the result and what would make the trial unsuccessful. Avoid selecting a task merely because the technology can perform a visible demonstration.

Set a comparison with the existing process. What does a good result look like? How long does the task normally take? Which errors would matter? Keep the evaluation proportionate, but include the work after generation. If a colleague spends substantial time correcting the output, that belongs in the account. The school should not describe the time to produce a first answer as the time to complete the task. That distinction is central to an honest claim about usefulness.

02 / Establish the data boundary first

Establish the data boundary first

Before staff use a tool, the school needs to understand its approved arrangements and the information involved. DfE’s guidance points leaders towards the data protection and technical checks required around AI use. Staff should not upload identifiable pupil, family or colleague information simply because it would make a prompt more convenient. Removing a name may not remove identifiability when the remaining detail is distinctive.1

Write the boundary in practical terms. Which information is permitted? Which is excluded? Who can answer a question about a proposed use? What should staff do if information has already been entered incorrectly? These instructions must be usable in the moment, not buried in a lengthy policy. Where the school does not yet understand a tool’s processing or retention arrangements, treat that as an unresolved approval question. Access to a service is not the same as permission to use it with school data.

03 / Make checking part of the task

Make checking part of the task

Verification should match the consequence of error. A generic draft may require factual and tone checks; a document referring to policy needs the current source checked directly. Staff should know that fluent wording does not establish accuracy. The school must retain responsibility for material it issues or decisions it makes, even when AI helped prepare the draft.

Define who checks the output and what they must check. If the task requires expertise the reviewer does not have, the trial is poorly designed. Do not use AI to make that expertise appear unnecessary. Where sources are suggested, open and verify them rather than reproduce an invented citation. Where information is uncertain, the final document should preserve the uncertainty. A confident answer that conceals a gap can be more damaging than an incomplete answer that prompts a useful question.

04 / Keep consequential decisions with accountable people

Keep consequential decisions with accountable people

A school should be particularly careful where an output may influence safeguarding, assessment, access to support or employment. A general-purpose system’s suggestion is not an adequate basis for a decision about a person. Leaders need the relevant evidence, professional judgement and applicable process. The trial proposed here is deliberately about bounded assistance, not transferring accountability.

Also consider how staff and pupils experience the use. If a communication reads as personalised advice, the recipient may reasonably assume a professional has considered the circumstances. The school should ensure that assumption is true. AI can assist preparation, but the responsible adult must understand and approve the result. The test is not whether a disclosure label is added at the end. It is whether the work has received the scrutiny its consequence requires.

05 / Evaluate the workload that remains

Evaluate the workload that remains

Ask participants to record the completed task, time spent drafting, time spent checking and the changes required. Include an account of errors and any uncertainty about data or approval. Use the findings to decide whether the activity is useful in the school’s context. A small trial does not prove a general effect across teaching or leadership.

Look for hidden transfers of work. Does one colleague gain drafting time while another must verify unfamiliar material? Does a manager now receive more documents to review? Does output volume rise without improving decisions? These are important practical consequences. If the trial adds work or lowers quality, stop or redesign it. That is a successful use of evaluation, not a failure to embrace technology. The school should be willing to decline an application whose benefit does not survive the complete task.

Give the trial an end date and an exit decision. If the school stops the use, clarify what happens to working material and access under the approved arrangements. Do not let a temporary experiment become permanent merely because nobody revisits it. If the use continues, retain the review trigger for changes in the task, information or tool. Permission should remain connected to the conditions that made the original decision defensible, rather than become an unrestricted endorsement.

06 / Decide before expanding

Decide before expanding

At the end of the trial, bring a concise account to the responsible leadership team: the task, approved boundary, quality findings, total time and unresolved risks. Decide whether to continue, modify or stop. If continuing, state the permitted use clearly and provide staff with a route for questions. Do not turn one promising result into blanket permission for unrelated activities.

Before half term, leaders can identify one task and run this bounded review with the data protection and digital leads involved from the start. Keep the evidence and decision in the existing governance record. The autumn standard should be neither excitement nor prohibition for its own sake. It should be a use that improves a defined task, preserves the school’s responsibilities and remains open to challenge when the evidence changes. That gives leaders a practical basis for progress without relying on supplier claims or assumed productivity gains.

Leadership questions

Questions that turn the development into a governing conversation.

What we are monitoring next

Publication is the beginning of the watch.

• Updated DfE guidance, local trial evidence and whether quality and total workload justify continued use.

Continue the evidence trail

How this develops our earlier coverage

• The wider AI evidence test ↗