Three years in: what we’ve learned about making AI stick

Most AI pilots in local government do not fail because the technology does not work. They fail because nobody changed the way the work happens around it.

That is the single most useful thing we have learned in three years of doing this. We started Outcomes Matter Consulting in 2023, and since then we have sat alongside caseworkers, social workers, SEND officers and service leads while they try to fold a new tool into a working day that was already full. The pattern is consistent. The licence gets bought. The training gets booked. And then, six weeks later, half the team has quietly gone back to the old way, because the old way was the one that fit around everything else they have to do.

This is not a workforce problem. It is a design problem. A tool dropped into an unchanged process is just one more thing to learn on a day that has no room in it. People are not resistant to AI. They are resistant to being handed something that makes their day harder before it makes it easier, with no support for the gap in between.

So the work that matters is rarely the technical work. It is the behaviour and culture work that sits around it. What does a good day look like once this tool exists? Who reviews the output, and how do they build trust in it? Where does the time saved actually go, and who decides? What happens the first time it gets something wrong, because it will, and does the team have somewhere to take that? None of these questions are answered by the software. All of them decide whether the software is still in use by Christmas.

The honest version is that adoption is slow, and in public services it should be. These are teams making statutory decisions about children and vulnerable adults. Caution is not a blocker to design around. It is a feature of doing the job responsibly. The mistake is treating careful adoption as a reason to do nothing, rather than a reason to do it properly: governance built in parallel with the work, a human-in-the-loop boundary the team can see and trust, and impact measured honestly rather than claimed.

Here is what changes when you get the conditions right rather than just the kit. Social workers tell us, consistently, that around 80 percent of their time goes on administration and only 20 percent on the direct work that actually changes outcomes. That is the worst admin ratio of any major UK profession, and it is not the fault of the people inside it. It is what a system configured this way produces. The prize from AI is not novelty. It is shifting that ratio, even slightly, so that the most expensive and most human part of the job gets more of the time.

But the prize only lands if the time saved is protected, named and pointed somewhere. We have watched teams claw back hours and then watched those hours silently absorbed back into the backlog. The technology created the space. The absence of any plan for the space wasted it. That is a behaviour change failure, not a technology one, and no upgrade fixes it.

So the advice we would offer, after three years, is short. Start with a real pressure the team already feels, not a tool somebody wants to try. Bring the frontline in early, because they will tell you where it breaks. Build the governance alongside, not after. And treat the rollout as the beginning of the work, not the end of it, because the login is the easy part.

None of this is the exciting bit. It does not demo well. But it is the difference between a pilot that quietly disappears and a change that is still there a year later, doing real work, freeing real time, for the conversations that actually matter to families.

Get the technology right and you have a tool. Get the conditions right and you have a change. Only one of them is still standing when you come back to look.

Because Outcomes Matter.