Bringing AI Into Public Service Delivery: Where to Start

Ask a ministry official what they want from AI and the answer is often a chatbot on the website. Ask the countries that have actually improved citizen service delivery how they did it, and the chatbot is the last thing they built, not the first.

The uncomfortable order of operations

Estonia is the country most often cited as a model of digital government, and its own account of how it got there is instructive: the visible, citizen-facing services came after roughly a decade of unglamorous work connecting government databases so they could talk to each other reliably, through a backbone system known as X-Road. Rwanda followed a similar sequence with its Irembo platform, which consolidated dozens of separate government services into one portal — again, integration work, not AI. In both cases, the "smart" layer that citizens actually notice was the last five percent of the project, sitting on top of the ninety-five percent that was data cleanup, standardisation, and system integration.

Why this order matters more for AI than for ordinary digitisation

A basic e-government portal can tolerate messy underlying data — a human civil servant on the other end can use judgement to interpret an inconsistent record. An AI system cannot. If a system is meant to triage pension verification requests, pre-screen permit applications, or route citizen enquiries, it will confidently produce wrong answers from inconsistent, duplicated, or incomplete records, and it will do so at scale and speed, which is precisely what makes the failure mode dangerous rather than merely inconvenient.

The single most common failure in public-sector AI pilots isn't a bad model. It's a good model given a bad, fragmented dataset to learn from.

What Nigeria already has to build on

This is not a story of starting from nothing. Nigeria has real foundational infrastructure already in place that most peer countries at this stage did not have:

• The National Identification Number (NIN) provides a unifying identifier that many AI-assisted service triage systems elsewhere had to build from scratch.
• JAMB's computer-based testing system demonstrated, at national scale, that a Nigerian public institution can run a high-volume, high-stakes digital process reliably.
• The 3MTT digital skilling programme is actively building the technical workforce that public-sector AI projects will eventually need to staff and maintain, at a scale — over a million applicants by early 2026 — that few African countries can match.

The realistic next step for most MDAs is not a flagship AI system. It's an honest audit of how clean, complete, and joined-up the data behind a specific service already is, followed by a narrow pilot on the process where that data is strongest — not the process that would look most impressive in a press release.

A practical starting sequence

1. Pick one service, not a strategy. A single, bounded process — pension verification, permit renewal, a specific benefit application — is governable and measurable. "AI transformation" as an agency-wide initiative usually is neither.
2. Audit the data before the model. Establish how complete, current, and consistent the underlying records are. This step is unglamorous and it is also where most of the real budget should go.
3. Start with triage, not decisions. Have the system rank or flag cases for a human officer rather than decide outcomes outright. This preserves accountability and builds institutional trust in the tool before higher-stakes use is considered.
4. Measure the human experience, not just throughput. A system that processes more cases per day but generates more citizen complaints has not actually improved service delivery.

None of this is a reason to wait. It's a reason to start in the right place — with the unglamorous data work that every country now held up as an AI success story did first, whether or not it made the press release.