AI Governance Framework (2026): Roles, Controls, NIST AI RMF, ISO 42001 and EU AI Act Dates

TL;DR

• An AI governance framework is the set of roles, rules and records a company uses to decide which AI systems and agents it runs, what data they may touch, who approves their actions and how problems get caught and fixed.
• Build it on NIST AI RMF (Govern, Map, Measure, Manage) and certify the management system against ISO/IEC 42001 if customers or regulators ask for proof.
• The EU AI Act timeline moved in 2026. Prohibitions and AI literacy have applied since 2 February 2025, transparency duties since 2 August 2026, and the Annex III high-risk rules now apply from 2 December 2027, per the European Commission.
• Agents are where most frameworks break. Sentra is the governed organization memory under your agents: permission-aware context, a named approver before anything is sent, and a record of what is true now.

What is an AI governance framework?

An AI governance framework is the operating system a company uses to control its AI: who owns each system, which data and actions it is allowed, how risk is assessed before launch, and how it is monitored after. A good AI governance framework fits on a few pages, names a person for every decision, and produces evidence an auditor or a customer can check.

AI governance is broader than model safety. It covers vendor chatbots that staff use every day, AI features switched on inside existing SaaS, internal models, and agents that read and write in company systems. Most failures happen in the boring parts: nobody knew a tool was in use, an agent had more access than the person it worked for, or a decision had no record.

The cost of skipping it is now measured. IBM's Cost of a Data Breach Report 2025 found that 13% of organizations reported breaches of AI models or applications, and 97% of those lacked AI access controls. 63% of breached organizations had no AI governance policy or were still writing one.

The three reference standards, and how they fit together

Use one framework as your spine and map the others to it. Most companies pick NIST AI RMF because it is free, voluntary and written in plain terms.

NIST AI RMF. The AI RMF 1.0 organizes the work into four functions. Govern is cross-cutting and sets culture, roles and policy. Map establishes the context and risks of a specific system. Measure analyzes and tracks those risks. Manage allocates resources to treat them, including incident response. NIST says the framework is being revised under the White House AI Action Plan, so check the NIST page for updates.

ISO/IEC 42001. ISO/IEC 42001 specifies requirements for establishing, running and improving an AI management system. It follows the same clause structure as ISO 27001, so a company with an existing security program can extend its audit cycle rather than start a new one. It was published in December 2023, per the IEC webstore.

EU AI Act. The Act sorts AI by risk and assigns duties to providers (who build or brand AI) and deployers (who use it). Most companies buying AI are deployers. The dates below reflect the Digital Omnibus on AI, Regulation (EU) 2026/1744, which entered into force on 27 July 2026.

EU AI Act obligations and dates (verified October 2026)

Source: European Commission. The Omnibus also rewrote the AI literacy duty in Article 4: providers and deployers must take proportionate measures to support staff AI literacy, without having to guarantee a set level for each person, per the AI Act Explorer.

Two deployer duties are worth building now, before the high-risk date. Under Article 26, deployers of high-risk systems must assign human oversight to people with the competence and authority to act, monitor the system, and keep its automatic logs for at least six months. Employers must also inform workers before using a high-risk system at work.

Fines under Article 99 reach EUR 35 million or 7% of worldwide annual turnover for prohibited practices, and EUR 15 million or 3% for most other breaches, whichever is higher.

The AI governance framework: seven components

Each component below lists what to build, who owns it, and where it maps in NIST AI RMF and ISO/IEC 42001. Use it as your AI governance checklist.

1. Roles and accountability

Name an executive owner (Head of AI, CIO or Chief AI Officer), an AI risk committee, and a named owner for every system. Legal, security, privacy and a business lead sit on the committee. Write down who can approve a new system, who can approve a high-risk one, and who can switch one off.

• NIST: GOVERN 2.1, roles and lines of communication documented.
• ISO 42001: Clause 5, leadership, roles and responsibilities.
• Evidence: A one-page RACI and committee minutes.

2. Inventory of AI systems and agents

You cannot govern what you have not listed. Keep one register of every AI system: vendor tools, AI features in existing SaaS, internal models and agents. Each row records the owner, purpose, data it touches, actions it can take, risk tier and review date.

• NIST: GOVERN 1.6, mechanisms to inventory AI systems.
• ISO 42001: Clause 4 scope and the Annex A controls on AI system life cycle.
• Tip: Find what is already in use before you write rules. Our guide to shadow AI covers discovery.

3. Data access and permissions

Set the rule once: an AI system or agent sees only what the person it serves could already open at the source. Classify data (public, internal, confidential, restricted) and say which tiers each approved tool may receive. Block restricted data from any tool whose contract allows training on your inputs.

• NIST: MAP 4, risks of third-party data and software components.
• ISO 42001: Annex A controls on data for AI systems.
• Detail: See AI agent permissions for inheritance patterns.

4. Human approval for agent actions

Split agent actions into tiers. Reading and drafting can run freely. Sending to a customer, changing a record, moving money or deleting anything needs a named approver. Keep the approval and its reason in the log. The same principle sits behind segregation of duties in finance.

• NIST: MANAGE 2.4, responsibilities to supersede, disengage or deactivate systems.
• EU AI Act: Article 26 human oversight for high-risk deployers.

5. Risk assessment before launch

Tier each system by impact: low (internal drafting), medium (customer-facing drafts, internal decisions), high (hiring, credit, health, safety, legal effect). High-tier systems need an impact assessment, testing on realistic data, and committee sign-off.

• NIST: MAP and MEASURE 2.3, performance measured in conditions like deployment.
• ISO 42001: Clause 6.1, AI risk assessment, treatment and impact assessment.

6. Monitoring, audit and incident response

Monitor accuracy, drift, cost and misuse in production. Log every agent action with the inputs it relied on. Define what counts as an AI incident, who gets paged, and how a system is paused. Review incidents at the risk committee monthly.

• NIST: MEASURE 2.4 production monitoring; MANAGE 4.1 and 4.3 on incident response and communication.
• ISO 42001: Clause 9, performance evaluation, and Clause 10, improvement.
• Detail: See AI agent observability.

7. Vendor management

Most of your AI is bought. For each vendor, record training terms, data retention, deployment options, certifications (SOC 2, ISO 27001, ISO 42001), sub-processors and exit terms. Re-review when the vendor ships a major model or feature change.

• NIST: GOVERN 6.1 and MANAGE 3.1, third-party risk.
• ISO 42001: Annex A controls on third-party and customer relationships.

AI agent governance: where most frameworks fall short

AI agent governance is the part of the framework that controls software that takes actions in company systems. A chatbot that gives a wrong answer wastes a minute. An agent that acts on a wrong answer sends the email, closes the ticket or updates the record.

Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. In May 2026 it added that applying one uniform standard to every agent will fail, and recommended governance proportional to each agent's autonomy.

Agent governance needs three things that model governance does not.

• Governed context. An agent acts on what it is told is true. If it reads last quarter's price list or a superseded policy, it acts on stale facts with full confidence. Govern the source of an agent's context the way you govern its permissions. See context engineering.
• Permissions inherited from sources. An agent connected with a service account often reaches far more than any one employee could. The safe default is query-time permission checks against each source system, where a failed check counts as no access.
• Approvals and records. Tie each action tier to an approver, and keep a record of what the agent did, for whom, and on which facts.

Where Sentra fits in your AI governance framework

Sentra is not a GRC tool and does not replace one. It is the governed organization memory that your assistants and agents read from, so the controls in components 3, 4 and 6 hold in practice.

Sentra connects email, chat, meetings, documents, ERP, CRM, HRIS and BI through 200+ integrations. Every answer and action respects each source's existing permissions, checked at query time against synced access lists, and a failed check counts as no access. An agent working for a sales rep sees what that rep could open, and nothing more.

Sentra knows when a fact became true and when it stopped being true. That matters for governance: an agent asking for the current refund policy gets the current one, with its source, and not the version replaced in March. A named person approves anything Sentra raises or sends before it reaches anyone else.

Any agent or model can use the same governed memory through REST and MCP, including ChatGPT, Claude, Copilot, Gemini and internal agents. That gives the risk committee one place to set the rules for context instead of one per tool. Sentra holds SOC 2 Type II and ISO 27001 certifications and deploys in the cloud, in an isolated VPC or air-gapped on premises.

To see how governed memory fits your framework, book a demo.

How to start: a 30-day plan

1. Week 1: Name owners. Appoint the executive owner and the risk committee. Publish the RACI.

2. Week 1 to 2: Build the inventory. Survey teams, pull SaaS admin logs and expense reports, and list every AI system and agent with an owner.

3. Week 2: Set data rules. Publish data classes and which approved tools may receive each. Pair it with a short AI acceptable use policy.

4. Week 3: Tier the systems. Assign risk tiers. Flag anything that may fall under Annex III before 2 December 2027.

5. Week 3 to 4: Fix agent access. Move agents off broad service accounts onto inherited permissions, and add approvers for external actions.

6. Week 4: Turn on logging and the incident process. Agree what an incident is and run one tabletop exercise.

7. Quarter 2: Decide on certification. If enterprise customers ask for proof, scope an ISO/IEC 42001 audit alongside your ISO 27001 cycle.

FAQ