An AI governance policy is not a statement that the company "uses AI responsibly". It is a set of decisions about which systems are allowed, who approves them, what data they can use, who reviews the output and what happens when something goes wrong. The policy should be short enough to use and specific enough to audit.
The regulatory direction is clear. The EU AI Act applies most high-risk obligations from 2 August 2026 and reaches full application in 2027. ISO/IEC 42001 provides a management-system standard. NIST's AI Risk Management Framework provides a practical risk vocabulary. Bangladesh's Personal Data Protection Act, 2026 adds privacy obligations for personal data used in AI systems. Even if an SME is not directly in scope of the EU rules, customers and vendors will increasingly ask for evidence.
Where things stand in 2026
Most SMEs have shadow AI: staff using public chatbots for drafting, coding, customer replies and document summaries. The business may have no inventory, no data rule and no review process. The risk is not only a bad answer. It is confidential data pasted into a tool, a generated claim published without review or an automated action that cannot be explained.
The first control is visibility. Ask teams which AI tools they use, for what and with what data. You will find more than expected. Then decide which tools are approved, which are allowed with conditions and which are prohibited.
2027: inventory, classification and human oversight
2027 should be the year of the AI register. For each system, record the owner, purpose, vendor, model, data used, affected people, decision type, human oversight and review date. Classify by consequence: assist, recommend, decide. An assistant that drafts a reply is different from an agent that approves a refund or scores a loan applicant.
Write the policy around three questions: What may the system do? What may it not do? Who is accountable when it is wrong? The answers should be specific. "The AI will be monitored" is not a control. "A named manager reviews all refunds above BDT 5,000 before release" is.
- Data rules: what personal, confidential or regulated data may be used.
- Human oversight: which decisions require approval and who can approve.
- Vendor terms: training use, retention, subprocessors, model changes, audit.
- Incident process: how to report, contain and review an AI failure.
2028: monitoring, incidents and training
By 2028, the policy should be operating. Monitor accuracy, bias, drift, cost and incidents. Keep a log of model changes and vendor changes. When a model is updated, re-run the evaluation set before it reaches production.
Train staff on the policy in plain language. They need to know what data they can paste, which tools are approved, how to check an answer and how to report a problem. A policy that only the IT team understands will be ignored.
Incident reporting should include AI-specific failures: a wrong answer that reached a customer, a data leak through a prompt, an automated action that could not be reversed, a model that behaved differently after an update. The response should be proportionate to the consequence.
2029: board reporting and third-party assurance
2029 is the year AI governance becomes a board topic. Report the AI register, the high-consequence systems, incidents, audit findings, training completion and vendor concentration. Use plain language. The board does not need a model architecture diagram; it needs to know where the business is exposed.
Third-party assurance becomes more common. Customers may ask for an AI management certificate, a completed questionnaire or evidence of risk assessment. ISO/IEC 42001 certification is one option; a well-run internal program with evidence may be enough for an SME. The important thing is that the evidence exists before the customer asks.
Contract renewals should include AI-specific terms: notification of material model changes, data use, subprocessors, audit rights, incident notification and exit. A vendor that cannot support those terms is a risk, regardless of the demo.
2030: embedded governance
By 2030, governance should be embedded in the way systems are designed and procured. New AI use cases go through a short review: purpose, data, risk, oversight, evaluation and owner. The review is proportionate. A low-risk writing assistant gets a checklist; a high-consequence decision system gets a deeper assessment.
The mature position is not "we have a policy". It is "we can show which AI systems we use, why, who owns them, how they are tested and what happened when they failed". That is the evidence that customers, auditors and regulators will ask for.
What to do in the next 90 days
- Inventory every AI tool in use, including personal accounts and browser extensions.
- Classify each use as assist, recommend or decide.
- Write a six-page policy with data rules, approval, oversight and incident response.
- Assign an owner to every system and a review date.
- Run the policy past one real use case and fix what is unclear.
How to measure progress
- Register coverage: the share of AI systems with an owner and a risk class.
- Risk assessment completion: before production, not after.
- Human oversight: the share of high-consequence decisions with a named approver.
- Incidents and near misses: reported, reviewed and closed.
- Training completion: the share of staff who completed the policy briefing.
AI governance is a management system, not a document. The businesses that do it well will move faster because they know which risks they have accepted.
Where this fits in our work
If you want help applying this to your business, start with Managed Technology & Cybersecurity. We publish scope and starting prices on the pricing page. For the next step in this series, read AI-ready data foundations: 2027, 2028, 2029, 2030.
Sources
- EU Artificial Intelligence Act, Implementation Timeline, accessed 16 November 2026.
- NIST, AI Risk Management Framework, accessed 16 November 2026.
- ISO/IEC 42001, AI management systems, accessed 16 November 2026.
- DataGuidance, "Bangladesh: Personal Data Protection Act 2026", accessed 16 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 *.