Why AI projects fail, and how a 1969 method reveals the conflict you didn't know you had

Watch on YouTube →

A kick-off meeting. Someone says "we're going to roll out an AI coding assistant across engineering." Six people nod. Everyone understood the sentence. Almost nobody meant the same thing by it.

That sentence isn't one project. In this example it's four, and the fourth one belongs to the developers who have to use the thing. Nobody writes that one down, and it's the one that comes true.

The method for finding them is Soft Systems Methodology, developed by Peter Checkland at Lancaster University over ten years from 1969. I was taught it as an undergraduate in 1991, then spent thirty-five years in an industry that reaches for the hard methods every single time.

I ran it on my own project and got two answers I didn't want. Both are in the video.

Two honest caveats, said in the video and worth repeating here. It won't save you from power: it surfaces disagreement, it doesn't settle it. And nobody has run the study. There's no published trial showing this improves AI project outcomes. I'm arguing from the fit.

CATWOE in full, as promised:

C — Customers. Who benefits or suffers from this. A — Actors. Who does the work. T — Transformation. What goes in, what comes out changed. W — Worldview. The belief that has to be true for any of it to be worth doing. O — Owner. Whoever could stop it. E — Environment. What you have to take as given.

The two that do most of the work are W and O.

Written version: https://jamiewatters.work/journey/soft-systems-methodology-explained

The project I ran it on is Executor File. It's open source, and the three definitions in the video are the ones I actually wrote for it.

Executor File: https://executorfile.com The code: https://github.com/TheWayWithin/executor-file Everything else I build: https://github.com/TheWayWithin

If you try it on something real, tell me what came back. I've run it once, on one project. I'd like to know whether that fourth one is always the one nobody wrote down.

Say hello, I answer:

If any of this saved you a meeting, you can buy me a coffee: https://buymeacoffee.com/jamiewatters

Chapters

0:00 Six people nod 0:55 Write the project as one sentence 1:38 The four projects hiding in one brief 2:57 You can build any one of them, not all four 3:51 Peter Checkland, Lancaster, 1969 4:38 Don't use the version Google gives you 4:58 Four activities, and it's a loop 5:33 The two tests 6:35 I was taught this in 1991 7:42 The move: a system is something you make up 8:56 CATWOE, and the two letters that matter 10:14 I ran it on my own project 12:26 It won't fail, it'll go quiet 12:43 Two honest limits 13:18 Why bother now 14:10 A chatbot answers, an agent acts 14:27 What to do with your next AI project