Every public institution I've advised on AI in the last two years has already run a pilot. Most of them can point to a working proof of concept, a vendor demo that impressed the steering committee, a use case that clearly saved time in testing. Very few of them have anything running in production a year later. The technology worked. The institution didn't change.
That's the part the vendor conversation skips. A private company can adopt a new tool because one executive decides to and absorbs the consequences internally. A public institution, or one operating with development finance oversight, can't do that, because the decision the AI is assisting with was never just a decision. It was an accountable act, tied to a mandate, a budget line, an audit trail and, eventually, a parliamentary or board question about why it was made. Nobody redesigned that accountability chain to accommodate a machine-assisted recommendation. So the pilot stays a pilot, because putting it into production means answering a governance question nobody has been asked to own.
This is why procurement is rarely the real blocker, even though it gets blamed the most. Procurement can be fixed with a better RFP process. The harder problem is that most institutions still don't have a documented answer to who is responsible when an AI-assisted recommendation turns out to be wrong. Until that answer exists in writing, risk-averse officials will keep treating the tool as advisory in name only, checking its output line by line, which erases most of the efficiency gain it was bought for.
Data governance compounds the problem. Institutions holding citizen data, grant records or commercially sensitive deal information under development finance mandates are, correctly, cautious about what leaves their environment and what a model is trained or fine-tuned on. That caution is appropriate. What's usually missing isn't the caution, it's a clear internal policy translating that caution into what can actually be automated, so every team isn't relitigating the same risk question from scratch on every use case.
The institutions that do get past pilot stage share a pattern. They treat AI adoption as a governance and capability project first, with a named accountable owner and a documented decision-rights framework, and a technology rollout second. They budget for the change management, not just the licence. And they start with a narrow, well-bounded use case where the accountability question is easy to answer, then expand once the model for answering it is proven, rather than trying to solve accountability and scale at the same time.
None of this is a reason to slow-walk AI adoption in the public and development finance space. The institutions that get the governance right first will move faster later, because they won't be relitigating the same trust question every time they want to expand a use case. The ones that skip that step will keep collecting pilots.