Somehow, it has been three years since GPT-4 was released(!). GPT-3 was viral but GPT-4 changed the texture of what seemed possible.

For me, this was the point where LLMs (Large Language Models) stopped feeling like an interesting frontier and started feeling like something that would change the economics of service delivery. And by that, I mean in the day-to-day sense of what could be built, by whom, and how quickly you could get something real in front of customers.

Gersi is a Business Analyst on my team who had never written a line of code before. We had an internal tool that was able to turn unstructured information into structured case notes. I watched him build an extension of that idea: the ability to upload an image and have the content pulled through into the notes as well. In half a day, he had something working well enough to drive discussion with the wider business. It was obviously rough, but it worked.

That was when it became clear that the bottleneck had moved.

For local government, building was assumed to be the part of a project fraught with cost and risk. That assumption shaped the way we approach development: who needs to be involved, how much documentation to produce up front, how many people needed to agree, how much certainty was expected before anything moved. This isn’t irrational. It reflects a world where software is slow to create, expensive to change, and risky to get wrong.

However, the underlying economics have changed without giving our ‘Ways of Working’ the time to catch up.

I experience this pattern regularly. A meeting with twelve people to discuss whether something is worth exploring. A placeholder calendar invite for yet another scoping session. A spreadsheet filled with owners, risks, dependencies and status before there is anything concrete to assess. Sometimes the cost of the conversation outweighs the cost of finding out.

We still behave as if building is the scarce part. So we preserve the old rituals: broad alignment, polished requirements, sequential approvals, the performance of certainty.

My team has shown, repeatedly, how quickly we can now move from a problem to something working. A recent example came through some FOI (Freedom of Information) work with Danielle, who leads the FOI team. In the past, a request like this would have followed the usual path: an initial brief, a project initiation document, requirements, an implementation plan, then a roadmap. Each stage had its logic. Each also moved us further away from the problem, and further away from anything real.

This time, we did something simpler. We had the conversation, then built an initial concept straight away. That working version became the artefact that carried the project. It changed the quality of the discussion almost immediately. Stakeholders were no longer responding to a projected future state described in documents. They were responding to something concrete. That made the thinking better, exposed the real questions earlier, and moved the work forward faster.

The next question is whether councils are willing to redesign decision-making in a world where the cost of testing an idea has collapsed.

If not, AI will not transform service delivery. It will just expose that our operating model still assumes that people who can build are slow, expensive and rare.