Kubair Shirazee

Transformation delivery · Enterprise agility · AI-enabled decision systems



5 min read

Build vs buy AI: the product I never built

The knowledge map worked. It had become a second brain for the programme.

By earlier this year, Sahar and I could ask a question about our own programme and get an answer back in near real time. Not in two weeks of evenings. Not after an argument over whose sheet was right. While we were still thinking about the question.

In the last post, I left that part out on purpose. The definitions were the hard part, and I wanted them to have the room. But they did get written down, the context did go in, and the pile that had eaten our evenings since 2011 started answering back.

Which led straight to the obvious thought. Most organisations I’ve worked with have this pile, and pay someone to read it. If it worked for a self-funded programme across three countries, it should work for a 300-person company, facing its own build vs buy AI decision.

This should be a product.

Why didn’t I build it?

In a sense, we already had. The model Sahar and I put together is still how Peace Through Prosperity (PTP) reads its field reports today. But it was built for us: our teams, our data, our definitions.

What I had was a working answer to our own problem. A product is a working answer to other people’s problems, every day, without me in the room. The gap between those two things turned out to be three things I didn’t have.

A product team. People who build, test, secure and maintain it, so it still works on the day I’m not looking at it. And willing partners whose data looks nothing like ours, so the model could learn more than one shape of field.

Distribution. A working tool that nobody hears about is a hobby. Getting it in front of the leaders who carry this pile is a full-time job on its own.

And the discipline that only comes from other people depending on the method. When it was just Sahar and me, we could forgive the system its quirks, because we knew them. A customer can’t, and shouldn’t have to.

None of those is a technology problem. Each one needs other people. And building it alone would have taken the time PTP and our other community commitments already ask of us.

Sahar and I had argued about this for years: was a product ever our core business? We’ve worked hard, and given things up, to keep PTP an independent social lab. Putting it within the pull of investors was never part of the plan. Experimenting, validating, open-sourcing and scaling social development programmes is.

Joining someone else’s product meant PTP never had to become one.

What changed?

In summer 2026, a recruiter messaged me on LinkedIn about a company that had already built what I was circling. A few weeks later, I spent a day with a made-up company’s chat threads and spreadsheets that disagreed with each other, writing down which source wins and why. It felt like a PTP quarter, compressed into a day.

It was the thing itself, built by a small team who had already settled the questions I kept repeatedly circling.

So I had a choice. Try to turn ours into a product alongside PTP, or join a team that already had one, and spend my time where I’m most useful: taking it into organisations and making it work there.

I chose the second.

To be plain about it, because you should read the rest of this post knowing it: I’ve joined the team that built it, and I’ll say more about them soon. Everything I’ve written in this series happened before I did, and it’s one of the reasons I did.

How do you decide on build vs buy AI?

Build your own organisational brain if reading the field is your core business and you have people to keep its definitions current. Otherwise, join someone who has. You’ll face the same choice I did, just with a bigger pile.

You’ll learn more than you expect if you do.

Most won’t be in that position. The honest test is the one I couldn’t pass on my own. Who maintains this on the day the person who built it isn’t looking? If the answer is one person, you haven’t built a system. You’ve built another integration layer, with a single point of failure.

Whichever way you go, three things held for us, and I’d expect them to hold for you.

Write your definitions down before you ask the machine anything. Definition debt doesn’t go away because the tool got better.

Keep the person who knows in the loop. Someone has to read the answer back before anyone acts on it.

And give your managers their judgement back. The point was to stop wasting the translators on translation.

The technology was the easy part. The people who make it work were never optional.

So here’s my question

Whichever side of build vs buy AI you land on, who would keep the pile honest?

Tell me where you’d start, in the comments, or grab 20 minutes on my calendar.

This is the last of three posts on what happens when the field reports faster than anyone can read it. Part one: Your managers aren’t the problem. They’re the integration layer. Part two: Same question, different answer: where inconsistent AI answers come from.

Thank you for reading. If you’re new here, I’m Kubair Shirazee, I help people, teams, organisations and communities navigate past the status quo and explore better ways of collaborating, deciding and delivering impact, increasingly with AI.


More writing