The Central Bank of Nigeria has not issued an AI agent training rule, yet its expectations reach every person who oversees one. Nigerian banks must run board-approved risk management, sound governance, cybersecurity programmes that include staff, and AML controls that work. Once an agent drafts or executes banking work, the reviewers become those controls and need provable skill.
What does the CBN expect, in substance?
The CBN supervises banks through a familiar set of expectations. Boards approve the risk appetite and answer for it. Management must be fit for its roles. Banks must identify, measure and control their risks, including operational and technology risk, and must run cybersecurity programmes that cover people as well as systems, with awareness and training for staff. AML and counter-terrorist financing controls must follow a risk-based approach, consistent with the FATF Recommendations: know the customer, monitor transactions, and report suspicious activity.
None of these was written with agents in mind, and none needs to be. Each assumes that where a bank relies on a person to check something, that person can check it. An agent changes how many checks there are and how fast they arrive, not who is accountable for them.
Data adds a second authority. The Nigeria Data Protection Act 2023, enforced by the Nigeria Data Protection Commission, governs how banks process customer data, including through automated means. Banks with capital market or insurance arms will also face SEC Nigeria or NAICOM for those businesses.
Who in a Nigerian bank does this reach?
Every licensed bank, from a large commercial bank to a microfinance bank, though expectations scale with size and complexity. Inside the bank, five groups need training that matches what they do:
- Operations staff and team leads who approve or decline actions an agent's platform holds for them, such as a transfer above a limit or a reversal request.
- Compliance and AML analysts who review agent-prepared alerts and draft suspicious transaction reports.
- Customer service leads who approve agent-drafted replies on complaints and disputes.
- Risk and internal audit who test the controls and reconstruct decisions.
- Directors, who approve the risk appetite that sets each agent's limits.
One local point deserves a line in every training plan. In Nigerian banking, "agent" already means a person running a banking point for the bank. Training materials, policies and board papers should name AI agents clearly so no one confuses the two, especially in incident reports.
What changes when the agent acts on a Nigerian bank's systems?
Instant transfers settle in seconds, so an approval is close to final. That is why depth matters so much here. An agent at depth 2 drafts a transfer for a person to send. At depth 3 it executes routine work within bounds, such as standard account updates. At depth 4 it may move money or take actions that are hard to undo, and on a governed platform those actions are held for a named person whenever they cross the bank's limits.
The reviewer of a held transfer must be able to answer three questions in under a minute. Where did the instruction come from: a sealed record in the core banking ledger, a signed instruction from a verified party, a declared request from the customer, or an external email? An external claim cannot authorise a payment. Is the action inside the agent's certified depth for this capability? Does anything in the request look like the patterns the bank's fraud and AML training describes?
Refusals and withheld actions need their own training. A refused action falls outside the agent's bounds and should be routed, not retried with different wording. A withheld action, such as reversing a settled payment on the platform's own rail, is blocked for every agent at every depth by design. Staff who understand that difference stop treating governance as an obstacle.
Finally, the agent's certificate lapses after 90 days, and a run that clears no depth drops it to depth 1 the same day. Branch and operations teams must be ready to take back the work immediately, which means people trained to do it by hand. The case for doing this before go-live is made in why payments operations staff should be certified before AI agents go live.
How do you meet the CBN's expectations?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Board-approved risk management | Each agent's depth and limits approved within the risk appetite | Board or committee minutes; the agent's configured limits |
| Competent people in control roles | Only staff with a current credential may approve held transfers and reversals | Approval records matched to credential expiry dates |
| Cybersecurity awareness | Staff trained to recognise external instructions and social engineering aimed at agents | Training records; phishing and injection drill results |
| Risk-based AML | Analysts trained to review agent-prepared alerts and draft reports with their own judgement | Alert dispositions with reasons; report quality reviews |
| Data protection under the NDPA | Reviewers trained to check data use in agent drafts | Decline reasons citing data; DPIA updates |
For the wider regulatory picture of agents in Nigerian banks, see CBN and AI agents in Nigerian banks. Our guide to AI literacy and certification for people who work with agents sets out the levels each role needs.
What will a CBN examiner ask?
Expect examiners to sample. Pick a held transfer that was approved: who approved it, what was their training status that day, and what did the record show them? Pick an AML alert the agent closed as low risk: who reviewed it, and on what basis? Ask how the bank would operate for a week if the agent dropped to depth 1. Ask how directors are kept informed of agent incidents.
The practical method for answering is covered in how banks can prove staff are trained to oversee AI agents. A comparison with Indian practice is in RBI expectations for training staff who oversee agents.
What do the CBN's expectations not cover?
They do not prescribe AI-specific training content, hours or a named credential. They do not decide which depth an agent may hold; the bank decides that within its risk appetite and on evidence from the agent's own certification. Data subject rights sit with the NDPC, not the CBN. Training also cannot fix a poorly bounded agent: if limits are set too high, a trained reviewer is catching what the design should have stopped. Some banks use an external credential, such as Pryme Intelligence Academy's Certified Pryme Operator, as one piece of evidence, but the duty to judge competence stays with the bank.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Do microfinance banks need the same training as commercial banks?
The principles are the same, scaled to the bank's size and the agent's depth. A microfinance bank using an agent only to draft replies needs less than one letting an agent prepare transfers, but both need reviewers who can check what they approve.
Can an AML analyst rely on the agent's risk rating?
The analyst can use it, but must form their own view before closing or escalating an alert. A suspicious transaction report is the bank's judgement, and the analyst must be able to explain it without pointing at the agent.
How should training handle social engineering aimed at agents?
Teach staff that instructions arriving from outside, such as an email asking to change a beneficiary, are external facts and cannot authorise a payment however convincing they look. Drills that plant such instructions are a good test of whether training worked.