Most companies already use AI, even without a policy. Staff paste customer emails into a chatbot, a vendor adds a summarisation feature, a developer calls a model API, and a marketing team generates images. Individually these look harmless. Together they create a set of decisions nobody has authorised, documented or reviewed.
An AI governance policy is the written answer to four questions: which AI uses are allowed, who approves them, what data they may touch, and how you will detect and correct harm. It should be short enough to use and specific enough to audit. This checklist gives you the sections, the evidence to keep and the questions a board should ask.
1. Scope and inventory
Owner: chief operating officer or equivalent. Evidence: AI register with review dates.
- Define what counts as AI for the policy, including embedded features in software you already buy.
- List every use case, the business owner, the vendor or model, the data used and the people affected.
- Mark whether each use is customer-facing, employee-facing or internal-only.
- Record the status: proposed, pilot, live, paused or retired.
- Include shadow AI discovered through expense claims, network logs or staff interviews.
2. Roles and accountability
Owner: executive sponsor. Evidence: RACI or named role descriptions.
- Name one accountable executive for AI risk, even if the work is delegated.
- Define who approves a new use case, who reviews data protection and who can stop a deployment.
- Give legal, security, operations and customer service a defined input, not an optional comment.
- Set a rule for conflicts of interest when the same team builds and approves a high-risk use.
- Record decisions and dissent in a log that survives staff turnover.
3. Risk classification
Owner: risk or compliance lead. Evidence: risk-tier template applied to each use case.
- Define tiers based on who is affected, the consequence of a wrong output and the sensitivity of the data.
- Treat employment, credit, education, health, safety and access to essential services as higher risk.
- Require a higher level of testing, human review and documentation for high-risk uses.
- Include reputational and legal risk, not only financial loss.
- Review the tier when the use case, model or data changes.
4. Data rules
Owner: data protection lead. Evidence: data-flow map and retention schedule.
- State which data categories may never be entered into external AI tools.
- Require a lawful basis and a documented purpose for personal data used in AI.
- Minimise data: send the minimum needed for the task and prefer identifiers over full records.
- Set retention for prompts, outputs, embeddings and logs.
- Check cross-border transfer and vendor training terms before approving a tool.
- Record how a customer or employee can request access, correction or deletion.
5. Vendor and model due diligence
Owner: procurement with security. Evidence: completed vendor questionnaire and contract clauses.
- Ask what model or service is used, where it runs and whether inputs are used for training.
- Request evaluation results, known limitations and the vendor's change-notification process.
- List subprocessors and any cross-border data flows.
- Require breach notification, audit rights and deletion on exit.
- Agree what happens if the model is deprecated or the price changes materially.
- Check that the vendor's security posture matches the data you plan to send.
6. Human oversight
Owner: business owner of each use case. Evidence: oversight procedure and sampled approvals.
- Define which decisions must have a human review before they take effect.
- Give reviewers enough information to challenge the output, not just a button to approve it.
- Set escalation for disagreements between the model and the reviewer.
- Provide an appeal route for customers affected by an automated decision.
- Monitor rubber-stamping: if approval rates are near 100%, the control is not working.
7. Transparency and customer communication
Owner: customer experience lead. Evidence: published notices and scripts.
- Tell customers when they are interacting with an AI assistant and how to reach a person.
- Explain in plain language what data is used and what the output can and cannot do.
- Avoid claiming accuracy or capability you cannot evidence.
- Keep a public or contractual AI notice aligned with the internal policy.
- Review marketing claims about AI with legal and product before publication.
8. Testing and evaluation
Owner: product or QA lead. Evidence: test plan, results and release sign-off.
- Test on representative local data, including Bangla and mixed-language text where relevant.
- Measure accuracy, false positives, false negatives and failure modes, not just demo quality.
- Run adversarial tests for prompt injection, data leakage and unsafe outputs.
- Check performance across user groups where the use case affects people differently.
- Require a rollback plan and a defined acceptance threshold before go-live.
9. Monitoring and incident response
Owner: operations and security. Evidence: dashboard, incident log and post-incident reviews.
- Monitor output quality, drift, refusal rates, latency and cost.
- Log enough to investigate an incident without retaining unnecessary personal data.
- Define what counts as an AI incident: harmful output, data leak, discriminatory result or unsafe action.
- Set notification paths for customers, regulators and partners where required.
- Review incidents for root cause and update the register and policy.
10. Employee use and training
Owner: HR with security. Evidence: training records and acknowledgement.
- Publish a short list of approved tools and the data that may be used with each.
- Explain that confidential client data, credentials and personal data do not belong in public chatbots.
- Train managers to recognise over-reliance and to review AI-assisted work.
- Require new staff to acknowledge the policy and refresh training annually.
11. Review and evidence
Owner: governance lead. Evidence: annual review minutes and action log.
- Review the policy at least annually and after a material change in law, vendor or use case.
- Keep the AI register, risk assessments, approvals, test results and incident logs together.
- Report a small set of metrics to the board: active use cases, high-risk approvals, incidents and overdue reviews.
Board questions
- Which AI systems can make or influence decisions about customers or employees?
- What would we do if a model produced a harmful output tomorrow?
- Which vendors hold our data, and what did we agree about training and deletion?
- Who can stop an AI deployment, and have they ever used that authority?
- Are we relying on a vendor's assurance instead of our own testing?
Our managed technology team helps write and operationalise these controls, and our pricing page explains how advisory and retained support are scoped. For the year-by-year view, read AI governance policy year by year.
Metrics
- Number of AI use cases with a named owner and current risk tier.
- Percentage of high-risk use cases with documented human review.
- Incidents per quarter and time to resolution.
- Overdue policy reviews and vendor reassessments.
- Training completion rate for staff using approved tools.
Sources
- National Institute of Standards and Technology. AI Risk Management Framework. Accessed 17 November 2026.
- International Organization for Standardization. ISO/IEC 42001: Information technology — Artificial intelligence — Management system. Accessed 17 November 2026.
- EU Artificial Intelligence Act. Implementation Timeline. Accessed 17 November 2026.
- Gartner. Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure. Accessed 17 November 2026.
- DataGuidance. Bangladesh: Data Protection Overview. Accessed 17 November 2026.
Responses (0)
No comments yet. Add the first one below.
Leave a response
Your email address will not be published. Required fields are marked *.