Bank of Ghana CISD 2026: What the Directive Requires of AI Systems
The Bank of Ghana's Cyber and Information Security Directive 2026 places AI and machine learning inside every regulated institution's cyber risk framework. Where the obligations sit and what they require.
The Bank of Ghana's Cyber and Information Security Directive 2026 (CISD 2026) brings artificial intelligence and machine learning systems inside the cyber risk framework of every institution the Bank regulates. AI obligations appear across the directive: in risk classification, staff awareness, security controls, incident notification, model-specific testing, supply chain oversight and a dedicated annexure on AI and ML governance. This article sets out where those obligations sit, what each requires of an AI system in practice, and how AdwumaTech designs and deploys systems to meet them.
The Directive at a Glance
| Item | Detail |
|---|---|
| Issued | March 2026, by the Bank of Ghana; revises the 2018 directive |
| Scope | Institutions licensed under Acts 930, 1032, 774, 987, 726, 732 and 1154, and any other institution licensed or regulated by the Bank of Ghana (Section 2) |
| Certification | Regulated institutions must achieve ISO/IEC 27001 certification and conduct due diligence on third-party service providers (Section 4) |
| Board | At least one board member with verifiable qualification or expertise in cyber and information security risk management (Section 5(3)) |
| Incident reporting | Monthly reports to the Bank by the 15th of each month, with immediate notification for events including customer data exposure and data integrity damage (Section 50) |
| Data location | Sensitive financial and personal data remains within Ghana, in alignment with the Data Protection Act 2012 and the Cybersecurity Act 2020 |
Sources: Bank of Ghana, CISD 2026; News Ghana.
Where the AI Obligations Sit
CISD 2026 distributes AI governance across the document. A July 2026 analysis by Dennis Gborglah found AI and ML obligations built into "at least six distinct locations" of the directive, spanning board literacy, vendor due diligence, cybersecurity specifications, technical testing and dedicated governance guidelines. The consequence for an institution is that AI compliance spans several functions at once: risk, security, technology, procurement and the board.
| Provision | Subject | Effect on an AI system |
|---|---|---|
| Section 16(2) | Risk categories | AI and ML enter the institution's cyber risk framework alongside every other risk category, with named owners |
| Section 61 | AI and ML awareness and education | Staff and leadership are trained on the risks specific to AI systems |
| Section 100 | Cyber and information security requirements for AI systems | AI systems are held to documented security controls covering access, logging and secure development |
| Sections 114 and 115 | AI principles, reporting and notification | AI systems are governed against confidentiality, integrity and availability, and significant incidents involving them are reported to the Bank |
| Section 122 | AI and ML model-specific testing | Testing extends to vulnerabilities particular to models and their training data |
| Section 96 and Annexure D | Supply chain cyber resilience | AI vendors fall under audit rights, incident notification, forensic cooperation and remediation or termination terms |
| Annexure E | AI and ML governance and control guidelines | Accountability, model auditing, bias detection and continuous monitoring |
Coverage of the directive's launch reports that the AI rules target systems used in fraud detection, credit scoring and customer service, to ensure they meet standards of fairness, transparency and security.
What the Directive Requires of an AI System in Practice
Read as engineering requirements, the AI provisions resolve into five properties an AI system in a regulated Ghanaian institution has to demonstrate.
-
Data residency. Sensitive financial and personal data remains within Ghana. For an AI system, that requirement governs where inference runs. A model that processes customer records, call audio or transaction data runs on infrastructure inside the jurisdiction: the institution's own data centre, an approved in-country private cloud, or local hardware. The model's prompts, retrieval indexes and logs sit inside the same boundary, because each contains the sensitive data it was built from.
-
Model-specific test evidence. Section 122 extends the testing regime to AI and ML models. Test coverage for a model includes adversarial and malformed inputs, attempts to extract training or retrieval data through outputs, degradation under shifted input distributions, and the integrity of the data the model was trained or tuned on. Each test run is retained as evidence with the model version it was run against.
-
Classified AI incidents. Section 115 requires notification of significant incidents involving AI systems, and Section 50 sets the reporting cadence. An institution needs a definition of what an AI incident is before one occurs: a model output that exposes customer data, an integrity failure in a decision the system informs, an outage of an AI-dependent service. Each definition maps to the monthly return or to immediate notification.
-
Vendor evidence. Where a third party supplies or operates the AI system, the supply chain provisions carry through to it. The institution holds audit rights over the vendor, notification obligations flow back from the vendor, and the vendor cooperates in forensic work. A vendor that hosts inference abroad, retains customer data for its own training, or declines audit access creates a compliance gap for the institution that engaged it.
-
A standing governance record. Annexure E expects accountability, model audit, bias detection and continuous monitoring. That translates into a named owner for each AI system, a record of its intended use and limits, performance and fairness measurements run on a schedule, and a change log linking every model or data change to the evaluation that approved it.
Implications for Boards and Senior Management
The directive places cyber expertise on the board and a Bank-approved chief information security officer in senior management, reporting directly to the chief executive. AI systems now fall inside that line of accountability. A board member with cyber expertise will be asked about the institution's AI inventory, the testing each system has passed and the incidents recorded against it. The awareness requirement in Section 61 extends the same literacy to the staff who build, buy and operate those systems.
How AdwumaTech Builds to the Directive
AdwumaTech builds AI systems for Ghanaian financial institutions to satisfy CISD 2026 by design: deployed inside the institution's perimeter, tested against model-specific threats, and delivered with the governance record Annexure E describes. AdwumaTech holds ISO/IEC 27001 certification, the information security standard the directive requires of the institutions it regulates, and aligns its AI management practice to ISO/IEC 42001 and GDPR.
The CISD 2026 Control Map
AdwumaTech maps each AI obligation in CISD 2026 to a control, an owner and a record before a system's architecture is fixed. The map below shows how each provision is met and what the institution holds as evidence.
| CISD 2026 provision | AdwumaTech practice | Record the institution holds |
|---|---|---|
| Data location | Deployment on private cloud or local hardware inside the institution's perimeter in Ghana; customer data, model weights, retrieval indexes and logs remain under the institution's control | Deployment architecture and data flow record |
| Section 16(2), risk categories | Each system entered in the institution's risk framework with a named owner, intended use and documented limits | AI system register entry |
| Section 100, security requirements for AI systems | Access through the institution's identity systems, logging into its security monitoring, secure development practice across the build | Security control mapping |
| Section 122, model-specific testing | Test suites covering adversarial and malformed inputs, data extraction attempts, distribution shift and training data integrity, rerun on every model version | Test results by model version |
| Sections 50 and 115, incident reporting | AI incident definitions mapped to the monthly return and to immediate notification, with monitoring alerts tied to each definition | Incident classification and response runbook |
| Section 4 and Section 96, third-party due diligence | ISO/IEC 27001 certification and supporting security documentation supplied for the institution's vendor assessment | Vendor due diligence file |
| Section 61, awareness and education | Training for operators, risk and compliance staff and leadership on the systems delivered | Training records |
| Annexure E, AI and ML governance | AdwumaTech's Evidence Chain: impact assessment, data lineage, disaggregated evaluation, decision logs, oversight records and change control | Evidence Chain record for each system |
Deployment Inside the Perimeter
AdwumaTech deploys every system it builds for a regulated institution on infrastructure the institution controls: its own data centre, an approved in-country private cloud, or local hardware. Inference, retrieval and logging run inside that boundary. The design satisfies the directive's data location requirement at the architecture level, so compliance holds regardless of how the system is later extended.
Nebula and the Systems the Directive Targets
CISD 2026 coverage identifies customer service, fraud detection and credit scoring as the AI systems its rules target. Nebula, AdwumaTech's platform for customer interaction in Twi, Ewe and Ga, serves customers in their language through voice and messaging channels and produces an Interaction Record: a structured English record of each interaction, written to the institution's backend systems. Nebula deploys entirely inside the institution's security perimeter, and each Interaction Record gives risk, compliance and fraud teams an auditable account of what the customer reported and how the system responded.
Evidence Produced With the System
AdwumaTech produces model-specific test results, evaluation reports and change logs during the build and hands them over with the system, so the evidence Section 122 and Annexure E call for exists from the first release. Each model or data change after handover runs against the same test suite, and the result is appended to the change log. The full evidence model, developed for public institutions, is set out in AI Assurance in the Public Sector.
Where Engagements Begin
AdwumaTech's AI Consulting Governance engagement produces the CISD 2026 control map for an institution's existing and planned AI systems. The AI Opportunity Assessment identifies the AI worth building under those controls, and the Deployment Readiness Score, a 12-question pre-deployment assessment, gives an institution its starting position.
Related Resources
- AI Assurance in the Public Sector: The Evidence a Government AI System Has to Produce
- The Verification Gap: Why Capable Models Produce Unreliable Systems
- Introducing Nebula
- Deployment Readiness Score
About AdwumaTech
AdwumaTech AI designs, builds and deploys AI systems within institutions and governments across Africa. AdwumaTech is ISO/IEC 27001 certified, aligned to ISO/IEC 42001 and GDPR, and its mGhana-ST dataset won Best African Dataset at Deep Learning Indaba 2026.
Frequently Asked Questions
What is the Bank of Ghana CISD 2026?
The Cyber and Information Security Directive 2026 is the Bank of Ghana's revised cybersecurity directive for regulated financial institutions, issued in March 2026 to replace the 2018 directive. It brings AI and machine learning systems inside each institution's cyber risk framework.
Does CISD 2026 apply to AI systems?
Yes. The directive includes AI and ML among all categories of cyber risk in Section 16(2), sets security requirements for AI systems in Section 100, requires AI incident notification in Section 115, mandates model-specific testing in Section 122, and adds AI and ML governance guidelines in Annexure E.
Which institutions does CISD 2026 cover?
Institutions licensed under Acts 930, 1032, 774, 987, 726, 732 and 1154, and any other institution licensed or regulated by the Bank of Ghana, including banks, specialised deposit-taking institutions and payment service providers.
Can a Ghanaian bank use an AI model hosted outside Ghana?
CISD 2026 requires sensitive financial and personal data to remain within Ghana. An AI system that processes such data runs on infrastructure inside the jurisdiction, and external hosting is limited to non-sensitive services with regulatory approval.
What AI testing does CISD 2026 require?
Section 122 adds AI and ML model-specific testing to the directive's testing regime, covering vulnerabilities particular to models and their training data, alongside vulnerability assessment, penetration testing and red team exercises.
How does AdwumaTech help institutions meet CISD 2026 for AI?
AdwumaTech maps each CISD 2026 AI obligation to a control, an owner and a record before architecture is fixed, deploys systems inside the institution's perimeter in Ghana, and delivers model-specific test results and its Evidence Chain record with each system. AdwumaTech is ISO/IEC 27001 certified and aligned to ISO/IEC 42001.
What records does AdwumaTech deliver for CISD 2026 compliance?
A deployment architecture and data flow record, an AI system register entry, a security control mapping, test results by model version, an incident classification and response runbook, a vendor due diligence file, training records, and an Evidence Chain record covering impact assessment, data lineage, disaggregated evaluation, decision logs, oversight records and change control.
AdwumaTech AI publishes operational diagnostics, systems research, and implementation insight on enterprise and government AI in Africa and beyond.
Tags
Explore Our Solutions
Discover how we build high-quality data for frontier AI models.
View our AI solutions