Every few years, a government somewhere decides its startup ecosystem needs fixing. It commissions research, identifies the gap, and funds a wave of accelerators and pitch competitions. Activity goes up. Founding rates tick upward. Politicians cut ribbons. Then the cohort graduates, and most of the companies quietly die, because the binding constraint was not only never early-stage capital. It was the absence of customers, or procurement rules that locked out young firms, or a talent market that kept draining to larger cities. The signal was real. The diagnosis was wrong.
This is the policy trap that more data alone cannot escape.
Governments are using AI more than ever, but mostly to do paperwork faster: summarising consultations, grouping evidence, flagging patterns. Useful, yes. Transformative, not yet. A recent white paper by Systemic Innovation, sistemaFutura and i4Policy puts a name to the problem: the interpretation gap. What we mean by this is the distance between knowing what the data shows and understanding what will actually happen when a policy hits the real world. More information (or signals) doesn’t close that gap. It can even make it worse, burying teams in signals nobody has time to reason through.
Marrying AI with ADDIS Decision Thinking
Closing the interpretation gap is a process design problem.
That’s where AI can meet methodologies like i4Policy’s ADDIS Decision Thinking; a structured, human-centered framework for moving through the full policy cycle. Embedding AI inside that kind of workflow shifts the question from “what does the data show?” to “what will this decision actually do?”
In practice, that means five things:
- Build causal maps, not just trend lines. Good policy analysis states explicitly why a policy should work. AI can help teams construct evidence-backed maps of how a system actually functions, making the assumptions visible before anyone commits to a direction.
- Simulate before you decide. Choices interact and compound over months or years in ways intuition struggles to track. Running structured scenarios lets policymakers see those feedback loops before implementation, when changing course is still cheap or where designing adaptive policies makes sense.
- Look across the whole portfolio. Different grants, rules, and programmes influence, or even quietly undermine, each other all the time. Modeling the full picture can surface those conflicts early.
- Take behaviour seriously. A model that assumes people will change what they do without the right motivation or support structure will fail in the field. Every scenario needs a reality check against how institutions and humans actually behave under pressure and who reacts to what incentive.
- Keep humans in charge. AI should sharpen judgment complementing human reasoning. Assumptions need to be visible and open to challenge; by citizens, practitioners, and decision-makers alike.
The bottom line
There’s a version of AI-for-policy that just makes bureaucracies faster at being wrong. And there’s a version that helps decision-makers reason more carefully about complex systems, test assumptions before committing resources, and learn from what actually happens.
The difference isn’t the algorithm. It’s the process wrapped around it.
Read the paper here: https://www.yumpu.com/en/document/view/71107687/ai-for-policymaking-beyond-signal-detection
Written by Tim Gelissen
