Pilots can hide the service around the technology

A funded pilot often contains resources that are easy to miss when calculating the cost of the AI itself. A vendor may absorb product-development effort. Researchers may test outputs. Development-program staff may prepare local data, train users and coordinate partners. A grant may pay for connectivity, messaging or expert review. When the pilot becomes a service, those activities either have to continue, change or disappear.

The model may still need access to current agronomic data, weather information, and local-language content. Users may need support. Responses may need monitoring. Expert escalation may remain necessary when confidence is low or the consequences of poor advice are material. The cost of producing an AI response is therefore only one component of the cost of delivering a reliable service.

The buyer question should arrive earlier

The World Bank’s World Development Report 2026 makes this transition explicit. It argues that governments need frameworks for testing, procuring and evaluating AI solutions to move successful applications from pilots to large-scale deployment, together with investment in staff capability. [3] That reframes scale. If evidence shows that an AI advisory service improves access to useful agricultural knowledge, which institution becomes responsible for purchasing it after the development program ends?

The answer could be a public agency budget, an existing agricultural-extension contract, a shared digital service, a producer organization, a public-private arrangement, or another model suited to the service and country. There is no universal purchasing model. There must, however, be an institutional route by which a successful service continues.

Asking which institution will buy the service is the right instinct and the wrong unit. A service that continues needs four things to have owners, and they are frequently four different organizations.

1.  The service owner is accountable for the outcome the citizen or the farmer receives. Not for the system. For the outcome.
2. The economic buyer funds the recurring cost. This may be a ministry budget line, an existing extension contract, a producer organization’s subscription, or a development partner during transition.
3. The operator runs it day to day: availability, incidents, content updates, user support.
4. The asset steward maintains the capability the service depends on. The local-language content, the agronomic data, the evaluation set, the model relationship. This is the role most often unassigned, because it is invisible while it is being done and catastrophic when it stops.

A sustainable service does not require one institution to perform all four. It requires each of the four to have a durable owner, with a budget and a named accountability. A pilot that ends with a service owner and no asset steward has not handed over a service. It has handed over something that will work for a year.

A pilot without that route has not failed. It has an unfinished service model.

What leaves, and what has to stay

The pilot-to-service handoff becomes clearer when temporary support and continuing obligations are viewed together. Grant funding, vendor absorption, researchers, and project staff can fall away at the boundary between validation and procurement. Six obligations do not: technology, data, and content, delivery, human support, assurance, and institutional capability. Each needs a permanent owner and budget before the service is expanded.

Build the buyer before expanding the technology

The next investment may need to build the purchasing institution before it expands the AI. A public organization may value a pilot and still lack a budget line for the recurring service. It may also lack a procurement specification capable of buying it, staff who can manage the supplier, or an accountable owner for its outcomes. Expanding user numbers before resolving those issues can make the eventual handoff harder rather than easier.

World Bank work on AI in agriculture reaches a similar conclusion from the sector side. Its Harnessing Artificial Intelligence for Agricultural Transformation report emphasizes the complementary investments required in infrastructure, governance, skills, local data, and institutional collaboration if agricultural AI is to scale responsibly. [4] The African Union’s Continental AI Strategy similarly places infrastructure, skills, datasets, innovation, investment, and cooperation inside an Africa-centered development approach to AI. [5] The service model therefore has to be larger than the model itself.

What if nobody can afford it

There is a case the framework has to be able to reach, and it is not a failure of the framework.

The institution best placed to own a service may not be able to fund it. An agricultural extension agency may have the mandate, the relationships and the staff. It may still have no budget line capable of absorbing a recurring technology cost, and no realistic prospect of one within the planning horizon. The pilot worked. The evidence is good. The economics do not close.

Three honest responses exist, and pretending otherwise is how pilots get extended indefinitely. The service can be restructured so its recurring cost fits what the institution can actually carry, which usually means a narrower service. The economic buyer can be separated from the service owner, with a funder, a levy, or a cross-subsidy carrying the cost while the agency owns the outcome. Or the service does not proceed at scale, and the program records that the intervention was effective and not affordable.

That last finding is worth publishing rather than burying. What should not happen is expansion on the assumption that the funding question resolves itself later. It resolves less easily with more users, not more.

Development finance should fund the handoff

This does not mean that every public-purpose AI service needs to become commercially self-sustaining. Some services create public value and should continue to be publicly financed. Development partners and DFIs can also play an important role in absorbing early uncertainty, building shared infrastructure, funding evaluation, or helping institutions acquire capabilities they do not yet possess. The issue is whether temporary financing is deliberately building towards a durable arrangement.

A useful financing plan should show which costs development funding is absorbing during demonstration, and which capabilities the permanent institution needs to acquire. It should also show when budget responsibility moves, and what evidence must be available before the service is expanded again. That is a different definition of scale. It replaces one question with another. Not how many more users can the pilot reach, but what has to become permanent for those users to keep receiving the service.

Africa does not need fewer experiments with AI for public purpose. It needs more experiments designed from the beginning to reveal what operating them successfully will require. The questions at the start of the next pilot should therefore be paired.

Can this work? And if it does, who will still own, improve and pay for the service when the pilot is no longer a project?

Scope. Agricultural advisory services carry liability, data-protection, and duty-of-care obligations that differ by country and by sector. Confirm the applicable regime and the local decision owner before a pilot becomes a public service.

‍