Building Legal AI for Africa: Lessons from Lagos State

In 2024, TNG was approached by Lagos State Government with an unusual brief: translate every statute on the books into Yoruba, using AI, at scale. The Lagos Law digitisation project became our most technically complex and personally formative engagement to date. This post is about what we learned — the hard way.

The Problem Is Not the Model

Every AI project I have led has taught me the same lesson: the model is the easy part. When we began the Lagos project, our engineering team was already deep into transformer architectures, transfer learning from mBERT, and custom tokenisation pipelines for Yoruba's tonal orthography. We were confident in our technical approach.

What we were not prepared for was the legal interpretation question. A statute is not a novel. Words carry precise, jurisdiction-specific meanings built up over decades of case law and legislative intent. When our model translated "beneficial owner" into Yoruba, it produced a grammatically correct phrase that meant, roughly, "the person who benefits." Any Yoruba speaker would understand it. No Lagos High Court judge would accept it.

Legal translation is not linguistic equivalence. It is jurisdictional equivalence — and that requires domain experts, not just language models.

Building the Human-in-the-Loop Pipeline

We rebuilt our pipeline around a collaboration model: AI produces a first-pass translation, a professional legal translator reviews and annotates errors, and those corrections feed back into fine-tuning. After three iterations, our model's outputs required 60% fewer human edits than the baseline. After six, we were close to production quality for standard statutory language.

The more important outcome was institutional: we helped Lagos State build internal capacity to review AI-generated legal content. The technology transfers as much as the translation does.

The Infrastructure Question

Serving an AI system for a government client in Lagos meant confronting realities that AWS documentation does not prepare you for. Power interruptions during peak hours, latency spikes caused by international routing, browser compatibility requirements for older devices used in government offices. We built an aggressively cached, progressive-enhancement architecture with graceful degradation — the full AI interface for modern browsers, a simplified statutory lookup for everything else.

This experience shaped our entire approach to African government technology. Resilience is not a nice-to-have. It is the product.

What We Would Do Differently

Involve legal practitioners from day one, not after the first model is trained. The annotations they provide in early rounds are exponentially more valuable than corrections made to a near-final system. We also underestimated the sensitivity of presenting AI-generated legal text to government stakeholders — framing matters enormously. We now lead with "AI-assisted" rather than "AI-generated," because accuracy aside, the perception of human oversight increases institutional trust and adoption.

The Broader Implication

Lagos State has 864 active statutes. Nigeria has 36 states. Africa has 54 countries. The opportunity to digitise, translate, and make accessible the legal infrastructure of a continent is not a niche product idea. It is a foundational infrastructure project — as consequential as building roads or electrifying cities.

That is why we keep building.