CBK's Draft AI Rules for Banks: Prior Approval, High-Risk Agentic AI and a 24-Hour Clock

On 10 September 2026 the Central Bank of Kenya published a revised package of draft prudential guidelines, risk management guidelines, guidance notes and a framework for domestic systemically important banks, and invited the public to comment. Inside that package sits the first dedicated artificial intelligence rulebook for Kenyan banks: the Draft Guidance Note on Artificial Intelligence in the Banking Sector.

It is a draft. It is also the clearest signal yet of how Kenya intends to supervise AI in the places where it moves money.

What CBK actually released

The AI guidance is not a standalone statement of principles. It is supervisory material that sits inside the framework CBK uses to examine banks, alongside a revised Risk Management Guidelines document and a new framework for identifying and supervising domestic systemically important banks. The consultation package also includes a draft Guidance Note on Cybersecurity and a draft Guidance Note on Third-Party Technology Service Providers.

Comments are open until 7 November 2026 and go to the Bank Supervision Department.

The timing was deliberate. CBK marked its 60th anniversary on 14 September, hosted the 2026 Annual Meetings of the Association of African Central Banks in Nairobi from 13 to 18 September, and ran an AI conference on 16 September under the theme "Trusted Artificial Intelligence in Central Banking Practice". The regulator spent the week making one argument to African central banks: AI is a supervisory matter. The draft guidance note is the domestic version of that argument.

The requirements that matter

The draft sets out a pre-deployment model that is stricter than anything currently in force in East Africa.

• Written approval before deployment. Institutions must seek CBK approval before deploying an AI system, and the application must be in writing, accompanied by the Data Governance and Artificial Intelligence Checklist in Annex III of the guidance note.


• Agentic AI classified high-risk. Any agentic AI system capable of independently initiating actions, interacting with external systems, executing transactions or materially affecting operations is classified as high risk, and subject to enhanced governance and independent review.


• Board responsibility. The board of directors holds ultimate responsibility for the ethical development and use of AI in the institution.


• An independent Model Risk Committee. Institutions must establish a Model Risk Committee that is independent of the functions responsible for developing, implementing and operating models.


• Bias impact assessments. Institutions must conduct bias impact assessments before deploying an AI system, with a Fairness-by-Design approach.


• Human oversight that works. Human oversight must include the ability to review a system's actions, intervene in or override its operation, and suspend or terminate it.


• Transparency toward customers. Customers and key stakeholders must be informed when and how AI is applied, with open disclosure on AI-reliant services.


• Incident reporting. CBK must be notified within 24 hours of any AI incident that materially affects service availability, confidentiality or integrity, or that could have a significant adverse impact on customers, reputation or financial condition, with quarterly incident records on top.

Read together, these are not aspirations. They are evidence requirements with deadlines attached.

Why the 24-hour clock is the real story

Of everything in the draft, the incident notification requirement will change institutional behaviour fastest.

A 24-hour window forces a bank to detect, classify and escalate an AI failure inside a single working day. That is an operational capability, not a compliance checkbox. It requires production monitoring that actually watches model behaviour, an escalation path that does not stall in committee, and a workable definition of "material" that survives contact with a real incident.

The guidance note lists what counts: model failures, unintended outcomes, bias or discriminatory effects, security breaches, data leakage, system outages and significant performance deterioration. It also requires audit logs, model inventories, retained training datasets and documented testing and validation results, including model drift detection.

The framing matters. An AI incident is treated as a security incident with a reporting duty, in the same family as a data breach or a system outage. That framing is right. A model that silently drifts, denies service unfairly, or fires an action it should not have fired is a live risk to customers and to financial stability, whether or not anyone acted maliciously.

The agentic AI classification is a warning

Classifying agentic AI as high risk is the most forward-looking line in the document, and it is the one with the longest tail.

The draft is specific about what agentic systems must do: operate within predefined authority limits, never modify their own objectives or operating parameters without authorisation, resist unauthorised access or manipulation, keep complete audit logs of material actions, and stay subject to continuous monitoring.

It also addresses the vendor problem directly. Where a bank procures an agentic system or AI agent from a third-party technology service provider, the bank must ensure the provider lets it exercise effective oversight, monitoring and control. Outsourcing the capability does not outsource the accountability.

Most Kenyan institutions are not running autonomous agents that execute transactions yet. CBK is setting the terms before they do.

The context: banks already moved

The draft is not hypothetical. The guidance note's own market overview, drawn from CBK's survey of AI in the banking sector, found that 50 per cent of surveyed institutions had adopted AI solutions, including 66 per cent of commercial banks, 57 per cent of microfinance banks and 43 per cent of digital credit providers. The top applications were credit risk assessment at 65 per cent, cybersecurity at 54 per cent and customer service at 43 per cent.

The governance lag is the striking part. Only 30 per cent of institutions had a formal AI strategy, even though 62 per cent had data strategies and dedicated data or AI teams. And 93 per cent of respondents told CBK they wanted comprehensive guidance on AI covering governance and compliance.

CBK has now answered that request. The written approval gate exists because a supervisor cannot oversee systems it has never seen. The bias and transparency requirements exist because a customer affected by an automated decision deserves to know it happened.

What it means if you build or buy AI for finance

For banks, the work is governance plumbing: an AI inventory, a Model Risk Committee that meets and minutes independently, bias testing before deployment, production monitoring wired to a 24-hour escalation clock, and disclosure language customers can actually understand.

For fintechs and vendors selling AI into financial institutions, the guidance resets the sales cycle. A model that cannot produce an explanation, an audit trail and a documented bias assessment becomes a liability for the buying bank. Explainability stops being a differentiator and becomes a procurement requirement.

For anyone building agentic products, the high-risk classification should shape the architecture now, while the comment window is open and the requirements are still malleable.

The deadline that matters this quarter

The single most useful date is 7 November 2026, when CBK closes comments on the draft package.

This is the point in the cycle where the industry still has influence. Once the guidance note is final, it becomes the baseline against which every bank in Kenya is supervised, and a precedent other East African regulators are likely to study. The comment template is published on the CBK website, and submissions go to the Bank Supervision Department.

The bottom line

Kenya's central bank has moved AI governance out of the policy debate and into supervisory practice: written approval before deployment, agentic AI treated as high risk, board-level accountability, independent model risk oversight, and a 24-hour incident clock.

The AI question in Kenyan banking is no longer whether models will be used. It is whether the institution using them can explain a decision, detect a failure, and report it before the day is out.