Build, buy, or something else: nine questions before your school commits to any AI tool

Sep 6

5 min read

Updated: Sep 13

On Wednesday 9 September at 16:50 I’ll be speaking at the AI in Schools conference run by SecEd and Headteacher Update. It’s the third year of the event, it runs online across two afternoons, and every delegate sits in every session, which shapes what a speaker can usefully say. My session is called Build versus buy: a teacher’s case study in scaling AI coaching. This is a preview of what I’ll cover, and the questions I’d want anyone to take away whether they attend or not.

The problem that started it

A teacher teaches somewhere around 800 lessons a year. Formal observation, under most union protocols, sees three of them. Schools that add learning walks and drop-ins see more, but in glimpses rather than whole lessons. Whatever the exact figures in your school, the shape is the same: well under one per cent of a teacher’s practice ever receives specific, whole-lesson feedback, and the other 790 or so lessons are judged only by the teacher’s own sense of how they went.

That isn’t a failure of effort. Nobody can watch 800 lessons per colleague. It’s arithmetic, and it was the problem I was trying to solve as an assistant headteacher well before there was any product attached to it.

Why the title is only half right

When Pete Henshaw and I first discussed the session, he wanted a case study of someone who chose to build an AI tool rather than buy one. That is roughly what happened, but not in the way the phrase suggests, and the difference is the point of the talk.

I looked at what was on the shelf. Nothing fitted. I needed whole-lesson audio rather than video, feedback grounded in a transcript rather than an impression, reports that were private to the teacher with no score attached, and a data protection position a union rep would be pleased to support. I also decided not to build it inside the school, because the hidden costs of doing that (liability, security, maintenance, models that change under you, the enthusiast who eventually moves on) land on the school and nobody else. What I did instead was co-found a company to build it properly, with an engineer and a commercial lead, and with my school as the pilot site rather than the developer. I'll write more about that in a future post.

So the honest framing is that build versus buy is rarely binary. There is a third route, which for most schools means being a pilot site that shapes a specialist provider’s roadmap rather than founding one. I’ll try to describe all three routes with their real costs, risks and trade-offs rather than the brochure versions.

What the session covers

Pete’s brief was practical: what would you tell other schools. So most of the 35 minutes goes on the things that turned out to matter more than the technology.

Staff and union engagement. What we put in place before anything was switched on, why recording is teacher-led and outside appraisal, and where our union reps started compared with where they are now. One rep has since written up their own view of the year in their own words, which is worth reading if the union question is the one holding your school back.

Data protection. The four things to have in writing before any classroom audio or text goes anywhere: roles, data flow, retention and transparency. And the one-page test I’d apply to any provider.

Safeguarding. What happens on the day an allegation is made, decided before it happens rather than after. Recordings turn out to protect staff as much as pupils, which is not how most people expect that conversation to go.

Integration with CPD, and the limits. The AI gathers the evidence; people do the coaching. I’ll be blunt about what transcript-based feedback cannot do, because the audience will find out anyway.

Costs, risks and trade-offs. Cost shapes rather than price lists, for all three routes, including the one I have a commercial interest in.

Nine questions before you commit to any route

This is the part I most want delegates to leave with, because it applies to any AI tool in any phase, not just coaching.

Is there actually a problem? Name it in one sentence, and name who feels it. If nobody does, you have a solution looking for a problem.

Is AI the right answer to it? What does it do that a person, a process or a spreadsheet can’t do more cheaply, and what does it do worse?

How will you know it’s working? Decide now what evidence you’d expect after a term, and what result would make you stop.

Who is it for, and who could it be used against? The user, the people in the data, and anyone who could be judged by its output.

What does it change about people’s work? What it replaces, what it adds, whose habits have to change, and whether they shaped it.

Where does the data go, and who profits from it? Your data is the asset. Ask who processes it, where, for how long, whether it trains anyone’s model, and how you get it back.

How hard is it to leave? Contract length, export format, switching cost, and how much of your practice now depends on one provider.

Could you run it yourself, and should you? Open-weight models now make keeping data in-house realistic. The price is everything below the waterline. Ask this one again every year. I wrote more about that here.

What happens when it goes wrong? A wrong output, a complaint, an allegation. Who notices, who owns it, and what the provider does about it.

Booking, and the rest of the programme

Registration is here, with rates on the conference site. The full programme runs from 14:30 on both days, and the session descriptions are worth a look. Three I’d point to: Anthony David on building a sustainable whole-school AI strategy, Matt Bromley on making assessment accessible with AI, both on the Wednesday before me, and Laura Knight on the procurement challenge on the Thursday at 16:10, which picks up the DfE AI and digital standards in far more depth than I will. Read together, her session and mine are the buy and build halves of the same decision.

If you’re coming, say hello in the chat. If you can’t, the nine questions are yours anyway, and if a version of this session would be useful for your trust or network, there are details on my speaking page.

Adam Sturdee is a senior leader and co-founder of Starlight, the UK’s teacher-first AI-powered transcript-based coaching platform for educators.

His work sits at the intersection of dialogic practice, instructional leadership and responsible AI strategy for schools and trusts.

He recently presented his research on AI-supported coaching at the BERA TEAN Conference 2026: https://www.bera.ac.uk/conference/bera-tean-conference-2026

Build, buy, or something else: nine questions before your school commits to any AI tool

Recent Posts

Comments