Governance

AI governance for regulated industries: a practical framework

Effective AI governance rests on four pillars: accountability, transparency, security and human oversight. In regulated industries it cannot be a document that sits beside the system; it has to be built into how the system is designed, deployed and monitored.

Published October 27, 2026 | For chief information officers, chief data officers, chief operating officers and compliance heads

Governance built into the stackAccountability: Named owners, Risk committee, Board reporting. Controls in delivery: Inventory, Risk tiers, Validation, Monitoring. Platform: Catalog and lineage, Access control, Model registry, Audit logs. Cloud foundation: AWS, Microsoft Azure, Google Cloud, Infrastructure as code.Accountability01Named ownersRisk committeeBoard reportingControls indelivery02InventoryRisk tiersValidationMonitoringPlatform03Catalog and lineageAccess controlModel registryAudit logsCloud foundation04AWSMicrosoft AzureGoogle CloudInfrastructure as codeGovernance built into the stackAccountability: Named owners, Risk committee, Board reporting. Controls in delivery: Inventory, Risk tiers, Validation, Monitoring. Platform: Catalog and lineage, Access control, Model registry, Audit logs. Cloud foundation: AWS, Microsoft Azure, Google Cloud, Infrastructure as code.AccountabilityNamed ownersRisk committeeBoard reportingControls in deliveryInventoryRisk tiersValidationMonitoringPlatformCatalog and lineageAccess controlModel registryAudit logsCloud foundationAWSMicrosoft AzureGoogle CloudInfrastructure as code
Fig. 1 Accountability sits at the top, and each lower layer produces the evidence the layer above relies on.

What are the four pillars?

  • Accountability. Every model and AI-supported decision has a named owner who answers for its performance and risks.
  • Transparency. Outputs can be explained to the people who rely on them, and to reviewers and regulators, at a level suited to the decision.
  • Security and privacy. Data and systems are protected, access is controlled, and personal information is handled according to the privacy law that applies, such as the Personal Information Protection and Electronic Documents Act (PIPEDA) in Canada, the General Data Protection Regulation (GDPR) in Europe or state laws in the United States.
  • Human oversight. A person remains responsible for consequential decisions and can review, override and stop the system.

Why are documents not enough?

Policies describe intent. Supervisors and auditors increasingly look for evidence that controls operate in practice: access logs, validation records, monitoring results and records of human review. Governance that is engineered into the platform produces that evidence as a by-product of normal work, rather than as a scramble before an examination.

What does the platform need to provide?

Engineering the controls means the data and AI platform itself carries them. A data catalog and lineage show where training and input data came from, policy-based access control and encryption with managed keys limit who can see sensitive data, and data contracts and quality checks stop bad inputs early. For models, a registry records versions and approvals, and evaluation suites test accuracy and bias and, for generative AI, whether retrieval-augmented generation (RAG) answers match their source documents. MLOps and LLMOps monitoring watch for drift and misuse. Infrastructure as code and continuous integration and delivery (CI/CD) pipelines apply the same controls on AWS, Microsoft Azure or Google Cloud, so evidence is produced the same way in each environment. The ISO/IEC 42001 AI management system standard, from the International Organization for Standardization (ISO) and the International Electrotechnical Commission (IEC), and the U.S. National Institute of Standards and Technology (NIST) AI Risk Management Framework are useful references when structuring this work; adopting them is a choice, not a requirement or a certification.

How should you tier AI use cases by risk?

Not every use case needs the same control. Tiering lets you apply heavier oversight where harm is greater and keep low-risk work moving.

TierExampleTypical controls
LowerInternal document search, drafting aidsApproved tools, data access limits, user guidance
MediumTriage or prioritization with human decisionValidation, monitoring, review sampling, logging
HigherOutputs affecting clients, credit, trading or compliance outcomesIndependent validation, explainability, human approval, formal change control

What controls belong in the checklist?

  • Maintain an inventory of all AI systems, including vendor-supplied ones, with owners and purpose.
  • Apply risk tiering and an approval path before deployment.
  • Record data lineage, run quality checks and document permitted uses.
  • Validate independently before release and after material change.
  • Monitor continuously for drift, errors, bias and misuse.
  • Use role-based access, logging and retention of prompts, outputs and decisions where appropriate.
  • Define human review points and escalation routes.
  • Prepare incident response, including a way to switch a system off.
  • Perform third-party due diligence and agree contract terms covering data use.
  • Train the people who use and supervise the systems.

What do regulators generally expect?

Expectations differ by regulator and activity, but common themes recur: sound model risk management, protection of personal information, control of third-party and outsourcing risk, and clear accountability. Securities regulators, prudential supervisors and privacy authorities in many jurisdictions (for example Canada, the United States, the United Kingdom and the European Union, where the European Union's AI Act also applies) have each published or are developing views on AI. Because these evolve, confirm current requirements with the regulator and counsel, and treat this article as orientation only.

How do you make governance practical?

  1. Start early. Involve risk and compliance in design, not in a final review.
  2. Automate evidence. Capture logs, test results and approvals in the delivery pipeline.
  3. Standardize. Use common templates for model documentation, testing and sign-off.
  4. Review regularly. Reassess tiering and controls as use cases and regulation change.

Who should be responsible for what?

RoleResponsibility
Board and executivesSet risk appetite and receive regular reporting
Business ownerAccountable for each AI use case and its outcomes
Technology and data teamsBuild, secure and operate the systems
Risk and complianceSet standards, challenge and monitor
Independent validation or auditTest models and controls
Legal and privacyAdvise on obligations, consent and contracts

Keep the lines clear between those who build and use a system, those who set standards and challenge, and those who independently assure.

What changes with generative AI?

Generative tools add questions about prompts, outputs and data leaving the firm. Controls to consider include approved tools only, restrictions on confidential data, review of outputs before external use, and logging for investigation. Treat them as part of the same framework, not a separate one.

What is a sensible first sequence?

  1. Take inventory of what is already in use, including tools staff adopted themselves.
  2. Agree a short acceptable use policy so staff know which tools and data are permitted.
  3. Tier the inventory and review the highest-tier items first.
  4. Choose one or two use cases as pilots for the full control set, and refine the process from what you learn.
  5. Report to senior management and the board on a regular schedule.

Resist writing a complete framework before any system is reviewed. Controls improve faster when tested on real use cases. Give staff a route to raise concerns about AI behavior, and treat each reported problem as input to improve the controls.

Where ARCHR fits

ARCHR designs governance, security and human oversight into data and AI work from the start. It can help organizations build an AI inventory, define risk tiers, set up validation and monitoring, and prepare evidence for internal and external review. Work is shaped around the client's stack and the regulatory context they operate in.

This article is for general information only and is not legal, financial or investment advice.

FAQ

Frequently asked questions

What regulations apply to AI in financial services?

It varies by jurisdiction and activity. Common themes include model risk management, privacy and data protection, outsourcing and third-party risk, and accountability. Confirm current requirements with your regulators and counsel.

Does governance slow projects down?

Poorly designed governance can. Governance built into delivery tends to reduce rework, incidents and audit findings, because requirements are known from the start.

Who should own AI governance?

Accountability should sit with named senior owners, supported by risk, compliance, legal, security and technology. A cross-functional committee is common, but it should not replace individual ownership.

Do we need governance for AI we buy from vendors?

Yes. Third-party AI still affects your clients and your obligations, so it needs due diligence, access controls, monitoring and a way to explain outcomes.

Let's talk about your data and AI priorities

Tell us what you are trying to solve. We aim to reply within one business day.