From Pilots to Portfolios: How Development Finance Institutions Should Fund AI
Development finance has funded hundreds of AI pilots and few lasting capabilities. The fix is to fund AI the way good investors fund anything uncertain: as a managed portfolio with stage gates, shared foundations and clear exits.
Ask any development finance institution how many AI pilots it has funded, and the number is usually large. Ask how many are still running at scale three years later, and the number is usually small. The pattern is familiar enough to have a name: pilot purgatory.
The cause is not a shortage of ideas or of goodwill. It is the way the money is structured. AI work is typically funded as a component of a single project, with a fixed budget, a fixed timeline and an end date. When the project closes, the pilot closes with it, whether it worked or not.
Why the project model fails AI
AI capability is not delivered once. It improves with data, use and iteration, and it depends on foundations that no single project owns: clean and shared data, computing capacity, skilled people and governance. A project budget pays for a model; it rarely pays for those foundations. The result is a scatter of disconnected pilots, each rebuilding the same basics, none reaching the scale where it pays back.
The project model also makes it hard to stop. A pilot that is not working is often extended to protect the project’s results, rather than closed so its money can go somewhere better.
Fund it as a portfolio
Venture investors and pharmaceutical companies manage uncertainty the same way: many small bets, explicit criteria for moving to the next stage, and the discipline to stop. Development finance can apply the same logic to AI.
Stage gates with real exits. Fund discovery cheaply and broadly. Move a smaller number of ideas to field testing against pre-agreed criteria: measurable improvement over the current process, acceptable risk, and a credible owner who will run it. Scale only what passes. Close the rest on schedule and publish what was learned.
Shared foundations. Fund the things every use case needs as public goods: data standards and sharing agreements, computing access, model evaluation and governance tools. Digital public infrastructure such as India’s UPI payments system shows how shared rails can let many providers build on one foundation. AI needs the same thinking.
Outcome metrics, not activity metrics. Count improvements in loan processing time, disease detection rates or crop yields, not the number of pilots launched or models trained.
Local ownership from the start. Every pilot should have an institution in the country that will own and run it if it succeeds, with staff trained during the pilot rather than after it.
What changes for boards and leadership
For the boards of development finance institutions, this means approving an AI portfolio with an overall budget, a risk appetite and an annual review, rather than approving AI line by line inside unrelated projects. It means reporting how many ideas were stopped, not only how many were started. And it means measuring success by the capabilities left behind in client countries.
The institutions that make this shift will spend less on duplicated effort and produce the evidence that justifies the much larger investments AI will need in the decade ahead.