The AI Regulation Market Map by Oonagh van den Berg | RAW Compliance | Issue 1: Africa
This series maps what each country’s rules mean for banks deploying AI - the ground where prudential governance, model risk, data, financial crime and real-world implementation meet.
Africa does not have one AI rulebook. It has something more difficult for banks: a fast-moving patchwork of banking, data protection, cyber, outsourcing, consumer protection and emerging AI requirements.
Until now, the practical approach was usually to map an AI use case back to existing obligations. Kenya has just proposed something much more direct.
In its September 2026 draft Guidance Note on Artificial Intelligence, the Central Bank of Kenya says financial institutions should not merely govern AI internally. They must seek written CBK approval before deployment. They must conduct a bias impact assessment before deployment. And they must create an independent Model Risk Committee with validation and approval functions that cannot be performed by the AI or data governance committees.
That is a game changer.
The regulatory question is moving from “Can you defend this AI if we inspect it?” to “Have you earned permission to put it into production at all?”
Kenya has moved the gate
The draft is still under consultation, so it is not yet a binding final rule. But its direction could not be clearer. AI in banking is being treated as a governed prudential capability, not an IT tool that can be launched first and documented later.
The draft places ultimate responsibility on the board and separates three functions that firms often blur: data governance, AI governance and model risk. The Model Risk Committee must be independent from the teams developing, implementing and operating models. It can coordinate with the other committees, but they cannot take over its independent validation or approval role.
Then comes the deployment gate. The draft states: “Institutions shall seek approval of the Central Bank prior to deployment of AI systems.” The application must be in writing and accompanied by CBK’s Data Governance and Artificial Intelligence Checklist.
This is broader than a notification. It is prior regulatory approval.
The bias requirement is equally direct. Financial institutions must conduct bias impact assessments before deploying an AI system, use fairness-by-design, test for protected characteristics and proxies, and continue fairness and bias audits after deployment. The draft also requires human review, intervention and override across relevant lifecycle stages.
In other words, governance cannot be retrofitted after the pilot works. The evidence must exist before production.
Why this matters beyond Kenya
Kenya’s proposal creates a reference point other African supervisors can now assess against. They do not need to copy the wording for bank boards and vendors to feel the effect.
A regional bank cannot sensibly run one governed AI model for Kenya and a less controlled version everywhere else. A vendor cannot build auditability, bias testing and independent validation for one customer at the end of the product cycle. The strictest credible market begins to shape the operating standard.
This creates five immediate changes for banks:
* AI inventories need deployment status, materiality and regulator-approval status, not just a list of experiments.
* Model validation must be independent from builders and business sponsors.
* Bias assessment becomes a pre-production control with evidence, owners and remediation thresholds.
* Third-party AI cannot outsource accountability; the bank must understand models, data, hosting, monitoring and exit.
* Regulatory engagement becomes part of the delivery plan, not a discussion after launch.
The African market map
Kenya is the clearest dedicated bank-AI regime in the markets reviewed. Elsewhere, the perimeter is being built by combining existing rules. That does not mean those markets are unregulated. It means the board must assemble the obligations across several regimes.
Kenya - dedicated bank-AI framework in consultation
The proposed CBK framework is the strongest pre-deployment model: board accountability, independent model risk governance, written approval before deployment, bias impact assessment, lifecycle monitoring, human oversight and continuing accountability for vendors.
Morocco - prudential and cloud rules do the heavy lifting
Bank Al-Maghrib has no dedicated bank-AI instrument identified in the official materials reviewed, but its governance, internal-control, operational-risk and cloud-outsourcing expectations apply. Directive 4/W/2022 requires risk-based cloud governance, oversight, audit and access rights, security, continuity and exit controls. Law 09-08 and CNDP oversight add the personal-data perimeter. For material or customer-facing AI, early supervisory engagement is the sensible posture.
Nigeria - data protection turns AI risk into board risk
The CBN’s technology, cyber, outsourcing and consumer rules remain relevant. The Nigeria Data Protection Act 2023 adds lawful and fair processing, impact assessment for high-risk processing and rights relevant to automated decisions. Explainability, contestability and cross-border data control are not optional design extras.
South Africa - automated decisions meet prudential scrutiny
SARB and the FSCA have mapped AI use and supervisory risk while existing prudential, operational-risk and third-party requirements remain in force. POPIA section 71 restricts solely automated decisions with legal or similarly significant effects unless an exception applies, with safeguards including representations and meaningful information about the logic.
Egypt - national AI governance sits over the banking baseline
Egypt’s national AI governance framework provides risk-based principles, while Banking Law 194/2020, CBE technology supervision and personal-data law form the binding perimeter. The framework is not a bank licence to deploy. Material use cases still need privacy, security, provider and human-accountability controls.
Ghana - cyber governance and automated-decision rights
The Bank of Ghana’s 2026 cyber directive makes board accountability, resilience, incident management and third-party control central. The Data Protection Act 2012 contains rights around automated decision-taking. Banks should connect AI governance directly to security, human escalation and customer redress.
Rwanda - policy ambition, privacy and banking controls
Rwanda’s National AI Policy sets direction rather than a bank-specific deployment rule. Its personal-data law and National Bank requirements on cyber and consumer protection shape operational use. The practical board questions remain human review, transfers, processor accountability, resilience and evidence.
Mauritius - cloud and outsourcing rules shape the architecture
Bank of Mauritius guidance on cloud, outsourcing, cyber and technology risk applies directly to AI delivery. The Data Protection Act 2017 gives rights concerning solely automated decisions with significant effects. Architecture, data location, audit access, sub-processors and exit planning therefore become regulatory design decisions.
One continent, one unavoidable direction
The instruments differ, but the direction converges around the same control spine: named accountability; documented use cases; proportionate risk classification; lawful and high-quality data; independent validation; human authority; bias and performance testing; secure and resilient deployment; third-party accountability; audit evidence; incident response; and an exit path.
Kenya’s contribution is to make the sequence explicit. Assess first. Validate independently. Test bias. Ask the supervisor. Deploy only after approval.
That sequence will feel slow to teams that still measure AI progress by the number of proofs of concept. It may prove much faster than explaining an ungoverned model after a customer is harmed, a suspicious activity case is mishandled or a regulator asks who authorised production.
What bank boards should do now
* Approve an AI policy, risk appetite, inventory and classification method.
* Name executive ownership and a board committee that receives material incidents, model performance, fairness results and independent assurance.
* Separate model-risk validation and approval from development and operational ownership.
* Build privacy, bias, security, outsourcing and customer-impact assessments into the deployment gate.
* Map regulator approval and notification requirements market by market.
* Require an auditable chain: inputs, model and version, outputs, human decisions, overrides, approvals and final filings.
* Choose one bounded financial-crime workflow where humans retain decision authority, then prove the controls before scaling.
For financial crime teams, the message is particularly important. AI can accelerate triage, entity resolution, investigation and drafting. It must not become autonomous regulatory judgement. A bank remains accountable for the alert it closes, the customer it excludes, the case it escalates and the report it files.
The winning institutions will not be those that deployed first. They will be those that can show, before deployment, why the model should be trusted, who can stop it and who owns the answer.
Kenya has drawn the new borderline. The rest of the market should assume it will not stay inside Kenya for long.
Sources and status note
Kenya’s Guidance Note is a draft issued for consultation on 10 September 2026, not yet a final binding rule. Kenya claims in this article are taken directly from CBK’s official draft. Other market positions reflect the official banking, privacy, cyber, outsourcing and AI-policy sources cited in the accompanying editorial notes. This article is a market overview, not legal advice.
CBK consultation and official guidance bundle
The AI Regulation Market Map | Issue 1: Africa | Oonagh van den Berg, RAW Compliance
The AI Regulation Market Map by Oonagh van den Berg | RAW Compliance | Issue 1: Africa
This series maps what each country’s rules mean for banks deploying AI - the ground where prudential governance, model risk, data, financial crime and real-world implementation meet.
Africa does not have one AI rulebook. It has something more difficult for banks: a fast-moving patchwork of banking, data protection, cyber, outsourcing, consumer protection and emerging AI requirements.
Until now, the practical approach was usually to map an AI use case back to existing obligations. Kenya has just proposed something much more direct.
In its September 2026 draft Guidance Note on Artificial Intelligence, the Central Bank of Kenya says financial institutions should not merely govern AI internally. They must seek written CBK approval before deployment. They must conduct a bias impact assessment before deployment. And they must create an independent Model Risk Committee with validation and approval functions that cannot be performed by the AI or data governance committees.
That is a game changer.
The regulatory question is moving from “Can you defend this AI if we inspect it?” to “Have you earned permission to put it into production at all?”
Kenya has moved the gate
The draft is still under consultation, so it is not yet a binding final rule. But its direction could not be clearer. AI in banking is being treated as a governed prudential capability, not an IT tool that can be launched first and documented later.
The draft places ultimate responsibility on the board and separates three functions that firms often blur: data governance, AI governance and model risk. The Model Risk Committee must be independent from the teams developing, implementing and operating models. It can coordinate with the other committees, but they cannot take over its independent validation or approval role.
Then comes the deployment gate. The draft states: “Institutions shall seek approval of the Central Bank prior to deployment of AI systems.” The application must be in writing and accompanied by CBK’s Data Governance and Artificial Intelligence Checklist.
This is broader than a notification. It is prior regulatory approval.
The bias requirement is equally direct. Financial institutions must conduct bias impact assessments before deploying an AI system, use fairness-by-design, test for protected characteristics and proxies, and continue fairness and bias audits after deployment. The draft also requires human review, intervention and override across relevant lifecycle stages.
In other words, governance cannot be retrofitted after the pilot works. The evidence must exist before production.
Why this matters beyond Kenya
Kenya’s proposal creates a reference point other African supervisors can now assess against. They do not need to copy the wording for bank boards and vendors to feel the effect.
A regional bank cannot sensibly run one governed AI model for Kenya and a less controlled version everywhere else. A vendor cannot build auditability, bias testing and independent validation for one customer at the end of the product cycle. The strictest credible market begins to shape the operating standard.
This creates five immediate changes for banks:
* AI inventories need deployment status, materiality and regulator-approval status, not just a list of experiments.
* Model validation must be independent from builders and business sponsors.
* Bias assessment becomes a pre-production control with evidence, owners and remediation thresholds.
* Third-party AI cannot outsource accountability; the bank must understand models, data, hosting, monitoring and exit.
* Regulatory engagement becomes part of the delivery plan, not a discussion after launch.
The African market map
Kenya is the clearest dedicated bank-AI regime in the markets reviewed. Elsewhere, the perimeter is being built by combining existing rules. That does not mean those markets are unregulated. It means the board must assemble the obligations across several regimes.
Kenya - dedicated bank-AI framework in consultation
The proposed CBK framework is the strongest pre-deployment model: board accountability, independent model risk governance, written approval before deployment, bias impact assessment, lifecycle monitoring, human oversight and continuing accountability for vendors.
Morocco - prudential and cloud rules do the heavy lifting
Bank Al-Maghrib has no dedicated bank-AI instrument identified in the official materials reviewed, but its governance, internal-control, operational-risk and cloud-outsourcing expectations apply. Directive 4/W/2022 requires risk-based cloud governance, oversight, audit and access rights, security, continuity and exit controls. Law 09-08 and CNDP oversight add the personal-data perimeter. For material or customer-facing AI, early supervisory engagement is the sensible posture.
Nigeria - data protection turns AI risk into board risk
The CBN’s technology, cyber, outsourcing and consumer rules remain relevant. The Nigeria Data Protection Act 2023 adds lawful and fair processing, impact assessment for high-risk processing and rights relevant to automated decisions. Explainability, contestability and cross-border data control are not optional design extras.
South Africa - automated decisions meet prudential scrutiny
SARB and the FSCA have mapped AI use and supervisory risk while existing prudential, operational-risk and third-party requirements remain in force. POPIA section 71 restricts solely automated decisions with legal or similarly significant effects unless an exception applies, with safeguards including representations and meaningful information about the logic.
Egypt - national AI governance sits over the banking baseline
Egypt’s national AI governance framework provides risk-based principles, while Banking Law 194/2020, CBE technology supervision and personal-data law form the binding perimeter. The framework is not a bank licence to deploy. Material use cases still need privacy, security, provider and human-accountability controls.
Ghana - cyber governance and automated-decision rights
The Bank of Ghana’s 2026 cyber directive makes board accountability, resilience, incident management and third-party control central. The Data Protection Act 2012 contains rights around automated decision-taking. Banks should connect AI governance directly to security, human escalation and customer redress.
Rwanda - policy ambition, privacy and banking controls
Rwanda’s National AI Policy sets direction rather than a bank-specific deployment rule. Its personal-data law and National Bank requirements on cyber and consumer protection shape operational use. The practical board questions remain human review, transfers, processor accountability, resilience and evidence.
Mauritius - cloud and outsourcing rules shape the architecture
Bank of Mauritius guidance on cloud, outsourcing, cyber and technology risk applies directly to AI delivery. The Data Protection Act 2017 gives rights concerning solely automated decisions with significant effects. Architecture, data location, audit access, sub-processors and exit planning therefore become regulatory design decisions.
One continent, one unavoidable direction
The instruments differ, but the direction converges around the same control spine: named accountability; documented use cases; proportionate risk classification; lawful and high-quality data; independent validation; human authority; bias and performance testing; secure and resilient deployment; third-party accountability; audit evidence; incident response; and an exit path.
Kenya’s contribution is to make the sequence explicit. Assess first. Validate independently. Test bias. Ask the supervisor. Deploy only after approval.
That sequence will feel slow to teams that still measure AI progress by the number of proofs of concept. It may prove much faster than explaining an ungoverned model after a customer is harmed, a suspicious activity case is mishandled or a regulator asks who authorised production.
What bank boards should do now
* Approve an AI policy, risk appetite, inventory and classification method.
* Name executive ownership and a board committee that receives material incidents, model performance, fairness results and independent assurance.
* Separate model-risk validation and approval from development and operational ownership.
* Build privacy, bias, security, outsourcing and customer-impact assessments into the deployment gate.
* Map regulator approval and notification requirements market by market.
* Require an auditable chain: inputs, model and version, outputs, human decisions, overrides, approvals and final filings.
* Choose one bounded financial-crime workflow where humans retain decision authority, then prove the controls before scaling.
For financial crime teams, the message is particularly important. AI can accelerate triage, entity resolution, investigation and drafting. It must not become autonomous regulatory judgement. A bank remains accountable for the alert it closes, the customer it excludes, the case it escalates and the report it files.
The winning institutions will not be those that deployed first. They will be those that can show, before deployment, why the model should be trusted, who can stop it and who owns the answer.
Kenya has drawn the new borderline. The rest of the market should assume it will not stay inside Kenya for long.
Sources and status note
Kenya’s Guidance Note is a draft issued for consultation on 10 September 2026, not yet a final binding rule. Kenya claims in this article are taken directly from CBK’s official draft. Other market positions reflect the official banking, privacy, cyber, outsourcing and AI-policy sources cited in the accompanying editorial notes. This article is a market overview, not legal advice.
CBK consultation and official guidance bundle
The AI Regulation Market Map | Issue 1: Africa | Oonagh van den Berg, RAW Compliance
The AI Regulation Market Map by Oonagh van den Berg | RAW Compliance | Issue 1: Africa
This series maps what each country’s rules mean for banks deploying AI - the ground where prudential governance, model risk, data, financial crime and real-world implementation meet.
Africa does not have one AI rulebook. It has something more difficult for banks: a fast-moving patchwork of banking, data protection, cyber, outsourcing, consumer protection and emerging AI requirements.
Until now, the practical approach was usually to map an AI use case back to existing obligations. Kenya has just proposed something much more direct.
In its September 2026 draft Guidance Note on Artificial Intelligence, the Central Bank of Kenya says financial institutions should not merely govern AI internally. They must seek written CBK approval before deployment. They must conduct a bias impact assessment before deployment. And they must create an independent Model Risk Committee with validation and approval functions that cannot be performed by the AI or data governance committees.
That is a game changer.
The regulatory question is moving from “Can you defend this AI if we inspect it?” to “Have you earned permission to put it into production at all?”
Kenya has moved the gate
The draft is still under consultation, so it is not yet a binding final rule. But its direction could not be clearer. AI in banking is being treated as a governed prudential capability, not an IT tool that can be launched first and documented later.
The draft places ultimate responsibility on the board and separates three functions that firms often blur: data governance, AI governance and model risk. The Model Risk Committee must be independent from the teams developing, implementing and operating models. It can coordinate with the other committees, but they cannot take over its independent validation or approval role.
Then comes the deployment gate. The draft states: “Institutions shall seek approval of the Central Bank prior to deployment of AI systems.” The application must be in writing and accompanied by CBK’s Data Governance and Artificial Intelligence Checklist.
This is broader than a notification. It is prior regulatory approval.
The bias requirement is equally direct. Financial institutions must conduct bias impact assessments before deploying an AI system, use fairness-by-design, test for protected characteristics and proxies, and continue fairness and bias audits after deployment. The draft also requires human review, intervention and override across relevant lifecycle stages.
In other words, governance cannot be retrofitted after the pilot works. The evidence must exist before production.
Why this matters beyond Kenya
Kenya’s proposal creates a reference point other African supervisors can now assess against. They do not need to copy the wording for bank boards and vendors to feel the effect.
A regional bank cannot sensibly run one governed AI model for Kenya and a less controlled version everywhere else. A vendor cannot build auditability, bias testing and independent validation for one customer at the end of the product cycle. The strictest credible market begins to shape the operating standard.
This creates five immediate changes for banks:
* AI inventories need deployment status, materiality and regulator-approval status, not just a list of experiments.
* Model validation must be independent from builders and business sponsors.
* Bias assessment becomes a pre-production control with evidence, owners and remediation thresholds.
* Third-party AI cannot outsource accountability; the bank must understand models, data, hosting, monitoring and exit.
* Regulatory engagement becomes part of the delivery plan, not a discussion after launch.
The African market map
Kenya is the clearest dedicated bank-AI regime in the markets reviewed. Elsewhere, the perimeter is being built by combining existing rules. That does not mean those markets are unregulated. It means the board must assemble the obligations across several regimes.
Kenya - dedicated bank-AI framework in consultation
The proposed CBK framework is the strongest pre-deployment model: board accountability, independent model risk governance, written approval before deployment, bias impact assessment, lifecycle monitoring, human oversight and continuing accountability for vendors.
Morocco - prudential and cloud rules do the heavy lifting
Bank Al-Maghrib has no dedicated bank-AI instrument identified in the official materials reviewed, but its governance, internal-control, operational-risk and cloud-outsourcing expectations apply. Directive 4/W/2022 requires risk-based cloud governance, oversight, audit and access rights, security, continuity and exit controls. Law 09-08 and CNDP oversight add the personal-data perimeter. For material or customer-facing AI, early supervisory engagement is the sensible posture.
Nigeria - data protection turns AI risk into board risk
The CBN’s technology, cyber, outsourcing and consumer rules remain relevant. The Nigeria Data Protection Act 2023 adds lawful and fair processing, impact assessment for high-risk processing and rights relevant to automated decisions. Explainability, contestability and cross-border data control are not optional design extras.
South Africa - automated decisions meet prudential scrutiny
SARB and the FSCA have mapped AI use and supervisory risk while existing prudential, operational-risk and third-party requirements remain in force. POPIA section 71 restricts solely automated decisions with legal or similarly significant effects unless an exception applies, with safeguards including representations and meaningful information about the logic.
Egypt - national AI governance sits over the banking baseline
Egypt’s national AI governance framework provides risk-based principles, while Banking Law 194/2020, CBE technology supervision and personal-data law form the binding perimeter. The framework is not a bank licence to deploy. Material use cases still need privacy, security, provider and human-accountability controls.
Ghana - cyber governance and automated-decision rights
The Bank of Ghana’s 2026 cyber directive makes board accountability, resilience, incident management and third-party control central. The Data Protection Act 2012 contains rights around automated decision-taking. Banks should connect AI governance directly to security, human escalation and customer redress.
Rwanda - policy ambition, privacy and banking controls
Rwanda’s National AI Policy sets direction rather than a bank-specific deployment rule. Its personal-data law and National Bank requirements on cyber and consumer protection shape operational use. The practical board questions remain human review, transfers, processor accountability, resilience and evidence.
Mauritius - cloud and outsourcing rules shape the architecture
Bank of Mauritius guidance on cloud, outsourcing, cyber and technology risk applies directly to AI delivery. The Data Protection Act 2017 gives rights concerning solely automated decisions with significant effects. Architecture, data location, audit access, sub-processors and exit planning therefore become regulatory design decisions.
One continent, one unavoidable direction
The instruments differ, but the direction converges around the same control spine: named accountability; documented use cases; proportionate risk classification; lawful and high-quality data; independent validation; human authority; bias and performance testing; secure and resilient deployment; third-party accountability; audit evidence; incident response; and an exit path.
Kenya’s contribution is to make the sequence explicit. Assess first. Validate independently. Test bias. Ask the supervisor. Deploy only after approval.
That sequence will feel slow to teams that still measure AI progress by the number of proofs of concept. It may prove much faster than explaining an ungoverned model after a customer is harmed, a suspicious activity case is mishandled or a regulator asks who authorised production.
What bank boards should do now
* Approve an AI policy, risk appetite, inventory and classification method.
* Name executive ownership and a board committee that receives material incidents, model performance, fairness results and independent assurance.
* Separate model-risk validation and approval from development and operational ownership.
* Build privacy, bias, security, outsourcing and customer-impact assessments into the deployment gate.
* Map regulator approval and notification requirements market by market.
* Require an auditable chain: inputs, model and version, outputs, human decisions, overrides, approvals and final filings.
* Choose one bounded financial-crime workflow where humans retain decision authority, then prove the controls before scaling.
For financial crime teams, the message is particularly important. AI can accelerate triage, entity resolution, investigation and drafting. It must not become autonomous regulatory judgement. A bank remains accountable for the alert it closes, the customer it excludes, the case it escalates and the report it files.
The winning institutions will not be those that deployed first. They will be those that can show, before deployment, why the model should be trusted, who can stop it and who owns the answer.
Kenya has drawn the new borderline. The rest of the market should assume it will not stay inside Kenya for long.
Sources and status note
Kenya’s Guidance Note is a draft issued for consultation on 10 September 2026, not yet a final binding rule. Kenya claims in this article are taken directly from CBK’s official draft. Other market positions reflect the official banking, privacy, cyber, outsourcing and AI-policy sources cited in the accompanying editorial notes. This article is a market overview, not legal advice.
