The institutional difference
Consumer artificial intelligence is judged by surface fluency. Public sector artificial intelligence is judged by accountability. A consumer chatbot that occasionally invents facts is a minor irritation. An AI assistant on a continental Public Financial Management platform that invents a citation, fabricates a reform statistic, or misattributes a policy position has caused institutional damage that is hard to undo.
The temptation on every AI deployment is to treat hallucination as a model selection problem. Better model, fewer hallucinations. The institutional reality is that no model selection eliminates the failure mode. The discipline that makes AI viable on a public sector platform is structural, not statistical, and it has to be designed in from the start.
Grounding as the structural commitment
The first structural commitment is grounding. Every AI answer is generated against a defined corpus, the platform's own curated content, with citations back to the source material. The user sees the answer and the source in the same view. If the source is not in the corpus, the answer is not generated. The model does not get to fall back on its parametric memory, because parametric memory is exactly where hallucination lives.
Grounding is not free. The corpus has to be curated, indexed, and kept current. The retrieval layer has to surface the right source material for the question. The generation layer has to produce answers that respect the source rather than embellishing it. Each layer has institutional weight, and the work of building them is the work of making the AI viable.
Editorial governance as the second commitment
The second structural commitment is editorial governance. AI outputs that affect what users see on the platform are routed through the same editorial workflow as human-authored content. AI-generated case study drafts are queued for administrator review before publication. AI-summarised content from autonomous ingestion is queued for review before it becomes part of the corpus. AI-assisted moderation flags are reviewed by human moderators before action. The AI does not get to bypass the governance that the rest of the platform operates under, because the governance is what makes the platform institutionally credible.
Methodological guardrails as the third commitment
The third commitment is methodological guardrails for the operations the AI does perform autonomously. When a user asks the AI to compare two countries on a reform domain, the comparison only proceeds if the underlying diagnostics are methodologically compatible. When the user attempts a comparison across incompatible assessment periods, or between a sub-national and a national evaluation, the AI surfaces a contextual advisory message rather than producing an analytically unsound output. The user's judgement is preserved in the final action, but the AI does not silently produce comparisons that practitioners cannot defend.
What we are building for ACBF
PANEOTECH is building exactly this AI layer for the Public Sector Collaboration Hub commissioned by the African Capacity Building Foundation, a specialised agency of the African Union, with funding from the Bill and Melinda Gates Foundation, in joint venture with JAMII LAB. The AI layer is delivered on a dedicated instance of Rafiki AI, branded for ACBF, with grounded conversational question answering against the Hub corpus, document intelligence including summarisation and structured analysis, semantic search and recommendations, multi-channel reach to WhatsApp and SMS, and an autonomous content ingestion engine that queues new material for administrator review prior to publication. Every output cites its sources. Every administrative action goes through the editorial workflow. Every comparison is bounded by the methodological guardrails.
The institutional lesson
AI on public sector platforms succeeds when it is treated as a content layer subject to the same governance as every other content layer. Treat it as a magic box and the platform inherits all the consumer-AI failure modes that public sector institutions cannot tolerate. Treat it as part of the editorial system, and it becomes a continental reference rather than a continental embarrassment.
