Who this guide is for
This guide is for principals, heads of department, training managers and anyone deciding whether and how to bring AI tools into a school, college or training program. It turns the published evidence into a sequence of steps. The steps and checklist are our own suggestions, built on what the studies report. They are not a standard and not legal advice.
The evidence behind it is summarised in our ebook, AI in Education: What the Trials and Surveys Show. The short version: results depend on how a tool is designed and supervised, and none of the trials we cite took place in Nepal.
Step 1. Write down the problem before choosing a tool
Name one specific problem in a sentence, for example “teachers spend too long preparing worksheets” or “students in a mixed-ability class cannot get help when stuck.” If you cannot name the problem, you cannot judge the result.
Then record a baseline: hours spent per week on the task, current test results, or the number of questions students ask. You will need these numbers in Step 8.
Step 2. Choose the lowest-risk use first
The evidence points to different levels of risk for different uses. This table is our reading of the sources, not a ranking from them.
| Use | What the sources report | Risk to learning |
|---|---|---|
| Teacher drafts materials, then edits | Gallup: weekly users estimate 5.9 hours saved per week; quality judged at least somewhat better by 57% to 74% | Lower, if teachers review the output |
| Answering routine administrative questions | Our own consultancy case study reports faster responses; see our inquiry automation guide | Lower, if answers come from maintained information |
| Student tutor with hints and structure | Harvard trial: gains on one short topic with a carefully designed tutor | Medium; depends on design and supervision |
| Open chatbot access for students | PNAS study: 17% lower later scores after practising with an unguarded chatbot | Higher |
If this is your first project, our own suggestion is to start in the first two rows.
Step 3. Set rules in plain language
Write a one-page policy that staff and students can read in five minutes. The UK Department for Education guidance is a useful example of what a government says to schools: safety should be the top priority, use should be carefully considered and assessed, and pupils should use these tools only with safeguards such as close supervision and filtering and monitoring. It also recommends that personal data is not used in generative AI tools. It is UK guidance, so check what applies in your country.
- Which tasks AI may be used for, and which it may not (for example, graded assessments).
- Whether students must say when they used AI.
- Who approves new tools.
- What staff must check before sharing AI-drafted material.
The OECD reports that seven in ten teachers on average believe AI could let students present others’ work as their own, so be explicit about assessment rules.
Step 4. Check the data and the supplier
An AI tool works on whatever is typed into it. Before any student or parent data goes in, ask the supplier the questions below. They draw on the UNESCO Recommendation on the Ethics of AI principles of privacy protection, transparency and human oversight, which include that AI systems “do not displace ultimate human responsibility and accountability.”
- What data does the tool collect, and does it need student personal data at all?
- Where is the data stored, and who else can see it?
- What is the minimum age, and how is it enforced?
- Can staff switch the tool off, and what happens to stored data?
- How was it tested, and against what comparison?
Our AI governance checklist covers more of these questions.
Step 5. Choose tools that teach, not just answer
For any student-facing tool, ask what it does when a student is stuck. The Harvard tutor used pre-written step-by-step solutions and a structure that guided students through each part of a problem. The PNAS authors describe unrestricted GPT-4 access as a “crutch.” A tool that gives hints and asks questions is closer to the first than the second.
Ask for a live demonstration with a real problem from your own curriculum, in the language your students use.
Step 6. Train staff before launch
Hold a short session so that teachers know what the tool can and cannot do. The OECD names professional development as part of the infrastructure needed for fair access, and Gallup reports that benefit grows with regular use. Our own suggestion is to give teachers time to try the tool on their real tasks before students see it.
Teach one habit above all: AI output is a first draft. A person checks it before it is used, because generative systems can produce confident but wrong content. NIST describes this as confabulation, “confidently stated but erroneous or false content” in its generative AI profile.
Step 7. Run a small pilot with a teacher in the loop
- Pick one course, one group of teachers or one class.
- Run it for a fixed period, such as one term.
- Keep a teacher present when students use any AI tutor.
- Collect problems as they happen: wrong answers, student confusion, privacy concerns.
- Agree in advance what would make you stop.
A small pilot is also the right response to the evidence gap. Without trials in Nepal, only your own data can tell you what works for your students.
Step 8. Measure learning, not only activity
Compare against the baseline from Step 1. Include at least one check that students take without the tool. In the Turkish trial, practice performance rose with AI access, but scores fell once it was taken away. Practice scores while using the tool do not show that learning happened.
- Teacher time on the task, compared with your baseline.
- Student results on a test taken without AI.
- Number and type of errors found in AI output.
- Student and teacher feedback in their own words.
Then decide: stop, adjust, or expand to one more group. Avoid expanding because it felt good. Expand because the numbers moved.
Checklist before you launch
- A single written problem and a recorded baseline
- A first use that is teacher-facing or tightly supervised
- A one-page policy covering allowed uses, assessment and disclosure
- A named person who can pause the pilot
- Supplier answers on data, age limits, storage and switch-off, in writing
- No personal data entered unless you have checked it is permitted and necessary
- Staff trained, and a review step for AI-drafted material
- A tool that gives hints and steps instead of answers, if students use it
- A way to test students without the tool
- A date to review the results and decide
If you would like help designing a pilot or a custom system for your institution, see our custom AI agent development or schedule a call.
Sources
- Kestin et al., AI tutoring outperforms in-class active learning, Scientific Reports, June 2025
- Bastani et al., Generative AI without guardrails can harm learning, PNAS, vol. 122, no. 26, 2025
- Gallup and Walton Family Foundation, report on teacher AI use and time savings, June 24, 2025
- OECD, Digital Education Outlook 2026, February 3, 2026
- UK Department for Education, Generative artificial intelligence in education, updated August 12, 2025
- UNESCO, Recommendation on the Ethics of Artificial Intelligence, adopted November 2021
- NIST AI 600-1, Generative AI Profile, July 2024
Frequently asked questions
Where should a school start with AI?
Our own suggestion is to start with teacher-facing tasks, such as drafting lesson materials, before giving students open access. The Gallup teacher survey reports time savings for those tasks, while student-facing tools carry the risks seen in the tutoring trials.
Should students be allowed to use AI chatbots?
That is a school decision. The evidence suggests an answer-giving chatbot can raise practice scores and lower later performance, as the PNAS study found. If students use AI, supervised tools built to give hints and steps, plus tests without the tool, are the safer approach.
What student data is safe to put into an AI tool?
Our own suggestion is to avoid personal data unless you have checked the supplier. The UK Department for Education recommends that personal data is not used in generative AI tools, and says personal data must be protected under data protection law.
How do we know a pilot worked?
Compare against a baseline you recorded before the pilot, and check student performance without the tool. Practice scores while using AI can mislead, as the Turkish trial showed.
What you get:
- Eight numbered steps from problem definition to a measured pilot
- A table ranking uses by risk to learning, labelled as our reading of the sources
- Supplier and data questions to ask before any student data is shared
- A ten-item launch checklist you can copy
- How to test learning without the tool, as the trials did