How Revenue Authorities Are Using AI to Close the Tax Gap

Ask most people what AI in tax administration looks like and they picture a robot doing an audit. The reality in every mature tax authority is quieter and more useful: AI mostly decides who to look at, not what to decide about them once you do.

The problem AI is actually solving

Every revenue authority faces the same structural limit: audit capacity is finite, but the number of returns filed each year is not. Historically, tax authorities have compensated with blanket rules — audit every return above a certain size, every business in a certain sector, every claim of a certain type. This catches some non-compliance, but it also burns audit hours on taxpayers who were always going to be compliant, while genuinely high-risk filings slip through because they don't fit the blunt rule. AI's real contribution is turning that blunt targeting into a ranked risk score, built from patterns in the data the authority already holds.

What this looks like in practice, internationally

• Estonia has built one of the most automated tax systems in the world. For most individual taxpayers, the return arrives already filled in, drawing on employer, bank, and other institutional data the tax authority already receives; the taxpayer's job is mostly to confirm it, not compile it.
• Chile's internal revenue service was an early adopter of machine learning models to detect VAT invoice fraud, particularly fake or "ghost" invoices used to inflate deductible costs — a pattern-matching problem that AI systems are well suited to, since fraudulent invoice networks tend to leave statistical fingerprints that are hard to see manually across millions of records.
• The UK's HMRC uses a system known internally as "Connect," which cross-references data from banks, property registries, and other government sources to flag inconsistencies between a taxpayer's declared income and their observable financial footprint.
• India's GST Network applies automated matching between the invoices a supplier declares and the credits a buyer claims, flagging mismatches for review rather than waiting for a manual audit cycle to catch them.

Where this maps onto Nigerian revenue administration

Nigerian revenue authorities are not starting from zero. FIRS's TaxPro Max platform already centralises filing, payment, and taxpayer records digitally, which is the actual precondition for any of the above — none of these AI systems work on paper files or fragmented spreadsheets. The realistic next steps sit in three areas:

The caution that matters more than the opportunity

Every one of these systems is only as good as the historical data it learns from, and that is where the real risk sits, not in the technology itself. If a risk-scoring model is trained on years of audit history that reflected uneven enforcement — certain sectors or regions audited more heavily than others for reasons unrelated to actual risk — the model will faithfully learn and repeat that pattern, and dress it up as an objective, mathematical result. This is not a hypothetical concern; it is the single most common failure mode reported by tax authorities that have deployed these systems internationally, and it is precisely why AI adoption in a revenue authority is a governance decision as much as a technical one.

The organisations that get the most value from this are the ones that ask "what did the model learn to associate with risk?" before they ask "does the model work?"

Nigeria's revenue authorities have an advantage most peer countries didn't have at this stage: the Nigeria Data Protection Act 2023 already sets clear rules for what fair, lawful data processing looks like, before any AI-specific rules are even in place. Building risk-scoring systems with that framework in mind from day one is considerably cheaper than retrofitting fairness and explainability into a system that's already been challenged in the field.