Building Government-Grade AI in Uganda: 5 Lessons from the Field

Under PPDA rules, the scope you wrote at bid stage becomes the scope you are held to eighteen months later, when the operating reality has changed. Teams that write specifications in terms of features get trapped. Teams that write them in terms of capabilities and measurable outcomes keep room to build the right thing.

Practically, that means describing 'a verified record of every published item, retrievable within two seconds, attributable to a named officer' rather than naming a specific database product. It also means budgeting for the evaluation cycle honestly: the gap between award and kickoff is rarely under a quarter, and any plan that assumes otherwise starts late and stays late.

Data residency is almost never raised in the first workshop. It arrives in month four, usually from someone senior who was not in the room earlier, and it can invalidate an entire hosting design. Under Uganda's Data Protection and Privacy Act, and the equivalent instruments in Kenya and Rwanda, the safe assumption is that citizen data stays in-country unless a specific case has been made and approved.

We now design for it from the start: national or regional hosting for anything containing personal data, model inference isolated so that no citizen record leaves the boundary, and a written data map that a compliance officer can read without a technical translator. It costs more up front. It costs far less than rearchitecting in month five.

A system that only works in English excludes a large share of the people it was built to serve, and, more damaging inside an institution, it excludes the field officers who actually enter the data. Luganda, Runyankole, Swahili, and Acholi are not localisation afterthoughts; they determine whether the record gets captured at all.

Model quality in these languages is uneven and improving fast, which means the design principle is graceful degradation rather than dependence. Voice input with human review. Templates in local language with English canonical records. Explicit confidence thresholds where a low-confidence transcription routes to a person instead of into the database.

Two-day training sessions produce certificates, not usage. What produces usage is a named person inside the institution whose job includes the system, who has authority to change a workflow, and who is still there in six months. Where we have had that person, adoption has held. Where we have not, usage decayed within a quarter regardless of how good the training was.

The corollary is uncomfortable for vendors: if the institution cannot identify that person during procurement, the honest response is to raise it as a delivery risk before signing, not to discover it at handover.

Public sector systems are judged on what they do in year three, not at launch. That means the deliverable is not the application; it is the application plus a runbook, plus a monitoring setup somebody actually watches, plus a funded support line, plus at least two people inside the institution who can deploy a change without calling us.

We now refuse engagements that have no maintenance budget. It is not a commercial position. A system that degrades quietly does more reputational damage to the institution, and to the case for AI in African government generally, than never having built it.