Anti-corruption technology in Cameroon: after Yaoundé

What was proposed in Yaoundé

From 4 to 8 May 2026, Yaoundé hosted the 16th Commonwealth Regional Conference of Heads of Anti-Corruption Agencies in Africa, convened by the Government of Cameroon in partnership with the Commonwealth Secretariat under a deliberately pointed theme: “Deploying Artificial Intelligence in the Fight Against Corruption in Commonwealth Africa.” For anyone following anti-corruption technology in Cameroon, the event marked a shift worth recording: the question in the room was no longer whether machine analysis belongs in the anti-corruption toolkit, but how to deploy it.

The sessions were concrete about the intended uses. According to the Commonwealth Secretariat's account of the conference, delegates examined how AI-driven tools can support investigations, sharpen data analysis, improve risk detection and strengthen public-sector oversight, with national case studies turning on the ethical safeguards responsible deployment requires. Dr Roger Koranteng of the Secretariat's Governance and Peace Directorate stated the conditional that matters: AI offers new tools, new insights and new opportunities — but only if it is harnessed with integrity.

The scale of the problem was quantified in the host country's own figures. As reported from the conference, Cameroon's National Anti-Corruption Commission (CONAC) received 10,520 corruption reports in 2024, with the resulting loss to the state estimated at just over 4 billion CFA francs; the Commonwealth noted that it has trained more than 7,000 officials across its 21 African member countries since 2011.

The conference closed with practical recommendations rather than a declaration of victory. Coverage of the closing sessions records more than 200 participants and three recurring asks: data sharing among member states, sustained training of officials, and harmonisation of legal frameworks. The Secretariat said the outcomes would inform Commonwealth Heads of Government discussions on governance reform and progress toward SDG 16.

The ambition is the right one, and this note takes it seriously. Precisely for that reason, the useful next question is unglamorous. Between a conference recommendation and a working detection system sits a layer nobody photographs: the procurement data itself, and the properties of the system that reads it. That layer decides whether AI against corruption becomes an instrument or a press release.

What anti-corruption technology requires of the data

Corruption-pattern detection is not a model problem first; it is a disclosure problem first. An algorithm cannot find bid rigging in a scanned PDF of an award notice. It needs structured records — who bought, who bid, who won, at what price, on what dates, amended how — published as data across the whole contracting cycle, from planning through tender, award, contract and implementation. This is what the Open Contracting Data Standard (OCDS) exists to make possible: a common, open data model for the full lifecycle, so that analysis is not locked to one buyer's filing habits or one vendor's schema.

Completeness matters more than sophistication. The signature patterns of procurement fraud — a tender window too short to be real, repeated single-bidder awards, amendments that quietly double a contract's value after signature — live in exactly the fields that partial disclosure omits. A model trained on incomplete records learns the shape of the gap, not the shape of the fraud. And the discipline has a corollary any honest system must state plainly: in a dataset with holes, the absence of a red flag is not evidence of a clean process. It is evidence of nothing.

Entity resolution: knowing who is actually bidding

The unit of procurement corruption is rarely a single tender; it is a network. The same supplier appearing under three spellings. Two “competing” bidders sharing a director, an address or an owner. A winning company connected to the official who signed the award. None of this is visible in tender data alone; it becomes visible only when records about companies and people are resolved to stable identities and linked across sources.

Entity resolution is therefore where procurement analytics quietly succeeds or fails, and it must be engineered as evidence, not as guessing. A match between two names is a claim with a measurable confidence, never a fact by default. Merge too eagerly and the system accuses the innocent — two genuinely distinct companies fused into one suspicious profile. Merge too timidly and a coordinated ring appears as unrelated strangers. A serious system quantifies the uncertainty of every link, records the basis of every merge, and routes the ambiguous cases to a human reviewer rather than resolving them by default.

The analytical system must itself stand audit

A system that flags a company for investigation is exercising a fragment of public power. That carries an implication the Yaoundé theme, to its credit, acknowledged in its attention to safeguards: an unauditable anti-corruption tool is a new opaque authority — the very thing anti-corruption work exists to eliminate.

Auditability here is specific, not rhetorical. For any flag the system raises, it must be possible to establish afterwards: which version of which data it read; which rule or model fired, and why; what evidence supports the finding; who saw it, and when; and what was decided. Findings must be re-verifiable outside the tool that produced them — carried with their evidence, so that an inspector, a judge or a defence lawyer can re-run the check without trusting the platform's word. A finding that cannot survive leaving the system that produced it is not yet a fact; it is a lead, and it must be labelled as one.

This is also what protects the analysis from its own operators. Tamper-evident logs and signed findings do not only prove what the system did; they make it detectable when someone tries to make the system un-see something.

Decision authority stays human

The last requirement is the one most easily eroded in deployment: no model output should, by itself, open an investigation, refer a company or trigger a sanction. The machine compiles evidence and states its confidence; an accountable institution decides. The distinction is not a courtesy — it is an integrity control. A capability to flag is not an authorization to act, and a pipeline that lets the first quietly become the second has recreated, in software, the unaccountable discretion it was built to remove.

Human authority must itself be structured to survive pressure: consequential decisions taken by more than one person, with written justification, on the record. That protects citizens from the model's errors — and it protects officials too. A decision made by named people on recorded evidence survives review; “the system said so” does not.

What this means for Central Africa

It matters that this conversation happened in Yaoundé. Coverage of the conference pointed to the breadth of Cameroon's existing control institutions — CONAC, the National Financial Investigations Agency (ANIF), the Chamber of Accounts, the Supreme State Audit (CONSUPE). The machinery of oversight exists; the question the conference posed is what that machinery can see. For Central Africa, the answer will be decided less by which model a country licenses than by which data its procurement process emits.

The practical opportunity is sequencing. Structured disclosure, resolved entities, auditable analysis, human authority: the first item is the cheapest and the least glamorous, and everything else depends on it. A jurisdiction that invests first in complete, structured, open-standard procurement data will find that even simple analytics begin to work — and that the sophisticated tools finally have something real to read. We map the institutions and frameworks that shape this terrain at procurement integrity in Cameroon.

Where SIGIL stands

SIGIL is a Cameroonian company, in Buea, building in exactly this space — and it is early. APEX, our anti-corruption intelligence system for public procurement, is designed to the requirements this note describes: it reads public data, detects defined patterns of procurement fraud, scores each finding with a calibrated certainty engine, and keeps every consequential decision with a human institution — no single person, and no model, can escalate a case. APEX is pre-deployment: it has no customers, no pilots and no deployments, and we say so plainly. Our standing rule is the one against which Yaoundé's theme will ultimately be judged: the institution holds the keys; the data stays under the institution's control; every consequential action is authorized, bounded, and proven.