AI Governance: A One-Page Framework for Small Companies

Start AI governance with 5 controls: an inventory of AI use, data rules, human review, vendor checks and an incident path. For a company without a compliance team, the useful minimum is a named owner and working evidence for each control, so you can answer a customer questionnaire without promising controls you cannot demonstrate.

Use the National Institute of Standards and Technology's AI Risk Management Framework, or NIST AI RMF, to organize the work. Consider ISO/IEC 42001, the international AI management system standard, when an external certification has a business purpose. Check the European Union AI Act separately for the legal obligations that apply to your role and use case. They solve different problems.

The framework descriptions and EU dates below were verified against official NIST, ISO and European Commission pages on 20 October 2026. This is operational guidance, not legal advice; get qualified advice on your specific legal obligations.

What Is AI Governance?

AI governance decides who may use AI, for which work, with which data, and who remains accountable when it produces an answer or takes an action. Policies set the rules; approvals, technical restrictions and records make those rules usable.

Think of a shared garden. An approved plant list does not tell you who can open the gate, change the irrigation or deal with a damaged bed. You need boundaries, a responsible person and a way to spot and repair problems. Buying better gardening equipment does not settle those questions.

For a business, the distinction becomes concrete when a support assistant moves from drafting replies to issuing refunds. The model may be unchanged, but its authority has changed. Governance should catch that change before the new permission reaches customers.

Data governance covers the information: access, quality, permitted use and retention. AI governance adds the behavior: what the system may infer, recommend or do, how you check it, and who handles the consequences.

AI Governance Framework: The One-Page Version

For a company with 20 to 500 people, size the process to the consequences of each use. Give each control an owner, a rule and a record. The following is an original operating template for a small company, not a certification checklist or a substitute for a system-specific assessment.

Scope: Staff AI use and AI features in sales, support and engineering, including features embedded in existing software.
Accountable executive: The COO or founder. Functional leads own their workflows; the IT or security lead owns access and technical controls.
Approval trigger: A new purpose, sensitive data source, external audience or action permission requires review before use.

Review cadence: Review the register quarterly and whenever purpose, permissions, vendor behavior or data changes. This is a suggested starting cadence; increase it where the consequences justify it.
Exceptions: Name the approver, reason, restrictions and expiry. An undocumented exception is an unknown exposure.
Board/customer reporting: Show what is in scope, what is controlled, what remains unresolved and who owns the next decision.

Keep this page short. Put the inventory and evidence behind it, with restricted access where records contain customer or employee information. The page is the index to a working process.

AI Governance Best Practices for Sales, Support and Code

Start with the workflow that has the greatest ability to expose data or affect a customer. Counting subscriptions alone will miss the permissions that matter.

Inventory Uses, Not Just Tool Names

Create a separate entry when the same product serves a different purpose or has different permissions. A sales drafting assistant and a customer-record enrichment workflow deserve separate entries even if they use the same vendor.

For each entry, record the business purpose; accountable owner; service and account; data categories; connected systems; read and write access; people affected; review method; approval status; and last review date. For a built system, add the deployed model or service configuration and a link to its change record.

Ask functional leads to include browser extensions, meeting assistants, coding assistants, automations and AI features switched on inside existing software. Reconcile their answers with available billing and access records. Describe the coverage honestly: a staff survey plus account review is a discovery method, not proof that every unapproved use has been found.

For a hypothetical 30-person company, the founder can maintain the register while sales, support and engineering leads confirm their entries. The benefit is a clear escalation path without creating a permanent committee.

AI Governance Policy: Specify the Data Boundary

Write the rule at the level an employee can follow: which account, which data and which task. “Use AI responsibly” does not tell a salesperson whether a customer contract can go into a chat window.

Set a conservative default: public material may go into approved tools; customer records, confidential code and internal commercial information require a specifically approved service and use case. Keep credentials and secrets out of prompts. Handle sensitive personal information through an explicit assessment rather than an informal exception.

Record what the approval depends on: retention settings, access permissions, contractual data use and the relevant integration. A vendor's statement about model training does not, by itself, answer where prompts are stored or who can access them. Those are separate checks.

Use the live AI acceptable use policy template for the employee-facing rules. Attach it to this governance process so an exception has an owner and a review date.

Put Human Review Before the Consequence

A reviewer needs the source information, clear criteria, enough time and permission to reject the output. Someone clicking an approval button without those things adds little protection.

In sales, require a person to check promises about pricing, security and contractual terms before they reach a prospect. In support, let the assistant draft from approved material, while a named employee checks refunds, account changes and sensitive disclosures. In engineering, route generated changes through the same code review, testing and release controls you expect of other contributions.

For each workflow, write the failure conditions that stop release. A support answer that exposes another customer's details fails regardless of how fluent it sounds. A code change that passes a unit test but bypasses authorization is still unacceptable.

An AI agent, software that can use tools to take actions, needs a permission boundary as well as output review. Limit the accounts, records and actions it can reach. Test that boundary before granting write access, and keep a way to revoke it.

Check the Vendor and the Configuration You Use

Assess the service you are buying, including the plan, account type and integration. A general security page is supporting material, not a description of your configuration.

Keep a short record of these checks:

• Data handling: What inputs, outputs and metadata are retained? Are they used for training? What deletion controls exist?
• Contract and processing: What does the agreement permit, which other providers process the data, and what commitments apply to incidents?
• Identity and permissions: Can you control access, remove a departing employee and restrict connected systems?
• Operational evidence: What logs, change notices and support routes can you obtain for this workflow?
• Exit and failure: Can you disable the feature, revoke credentials and continue the essential task without it?

Mark an unanswered question as unresolved. If the missing answer concerns customer data or an irreversible action, reduce the scope or keep that use paused until someone accepts a specific, documented risk.

Use the Incident Process You Already Have

Give staff an obvious route to report an unsafe answer, data exposure, unauthorized action or suspected misuse. The first responder should know who can disable the integration and who decides whether it can restart.

Record the affected workflow, time, data involved, relevant output or action, permissions and available logs. Restrict access to that evidence. Contain the problem, assess who was affected, and assign responsibility for any notification duties under your contracts and applicable rules.

Then fix the control that failed. If a support assistant exposed a restricted document, deleting the answer does not repair its document access. Check the permission change, test the failure case again and record the restart decision.

For a hypothetical 150-person business, the existing security incident lead can coordinate this with the support owner. If you have no dedicated security role, name the person who can revoke access and the executive who can make the escalation decision.

AI Governance Standards: NIST, ISO and the EU AI Act

Use NIST to structure risk decisions, ISO to organize a management system, and the EU AI Act to determine applicable legal duties. Treating the three as interchangeable creates misleading questionnaire answers.

Apply NIST to a Workflow You Can Inspect

The NIST functions are an ongoing process, not four boxes you complete once. The official AI RMF Core explicitly allows tailoring and describes governance as running across the other functions.

Applied to a support assistant, the mapping could look like this:

• Govern: The support lead owns the use; the COO approves the risk boundary; IT controls access.
• Map: Document the customer information the assistant reads, who receives its answers and what harm an incorrect answer could cause.
• Measure: Check representative and difficult cases, including attempts to retrieve restricted information. Record the results and unresolved limitations.
• Manage: Restrict access, require review where needed, monitor failures and disable the workflow if its limits are breached.

This is an illustrative mapping. Describe it as “using NIST AI RMF to structure risk management” when that is what you have done. Keep the evidence behind the statement.

AI Governance Certification: Make It a Scoped Business Decision

ISO/IEC 42001 concerns the organization's AI management system. Its public description uses a Plan-Do-Check-Act approach: establish the process, operate it, assess it and improve it. A one-page policy is only a small part of that work. ISO's standard overview.

Certification is a separate external assessment. ISO develops standards but does not certify organizations itself; external certification bodies perform that work. ISO's certification guidance.

Consider certification when a customer requires it, your procurement market values it, or your AI operations justify a formal management system. Confirm which entities, locations and activities a certificate covers. A staff member's course certificate does not demonstrate that your company's workflows operate under an assessed management system.

EU AI Act Dates: Separate Existing Duties From Later Deadlines

The high-risk timetable has changed, but that does not postpone the whole Act. The Commission says the AI Omnibus entered into force on 27 July 2026, extending the main high-risk application dates and simplifying the previous company AI-literacy requirement. Commission announcement.

The official implementation timeline records these milestones:

• 2 February 2025: Original prohibitions and AI-literacy provisions started applying.
• 2 August 2025: General-purpose AI model obligations and governance provisions started applying.
• 2 August 2026: General application, including Article 50 transparency rules.
• 2 December 2026: A specific Article 50(2) transition for certain synthetic-content system providers already on the market before 2 August 2026; this is not a blanket transparency extension.
• 2 December 2027: High-risk systems in Annex III.
• 2 August 2028: High-risk AI embedded in regulated products covered by Annex I.

These are selected milestones; the Commission's implementation timeline identifies the associated provisions and transitions.

For a founder, the first question is your role and intended use. The Commission describes a scope that reaches public and private actors inside and outside the EU when they place AI on the EU market or put it into service or use it in the EU. It also identifies certain employment and creditworthiness uses as high-risk examples. Headcount alone does not answer either question. Commission scope and classification FAQ.

For an operator running a customer-facing chatbot, review the applicable disclosure requirements now. For a builder considering hiring-related automation, assess the intended use before treating it like an ordinary drafting assistant. Keep the date and source of that assessment in the inventory.

The site's EU AI Act tooling guide is the adjacent buying resource. Use the official sources above for the dates and requirements summarized here.

What to Send With a Customer Security Questionnaire

Send evidence that matches the customer's question and the service you provide. There is no universal smallest pack that guarantees a pass: a customer assessing internal drafting has different concerns from one buying an autonomous system with access to production data.

For a first review of ordinary business AI use, assemble these six evidence items:

1. Approved policy and owner: A current policy, its approval date, scope and accountable executive.
2. Relevant inventory extract: The AI uses involved in the customer's service, including vendors, data categories, permissions and owners.
3. Vendor evidence: The assessment and supporting terms or settings for those uses. Share what your agreements allow you to share.
4. Review and test sample: A redacted record showing an actual output or change was checked against defined criteria, including what happened when it failed.
5. Staff briefing record: Who received the applicable rules and how they learned the reporting and approval route.
6. Incident route and exercise: Contacts, stop authority and a completed incident record or a clearly labeled rehearsal.

Do not send sensitive prompts or customer data just to make the pack look thorough. Provide a scoped summary and offer restricted access to appropriate supporting records.

A useful questionnaire answer names the boundary. Only if these controls are implemented, wording like this is defensible:

AI-assisted support drafting is included in our AI-use register. The support lead owns the workflow. Approved staff review responses before sending, the integration cannot issue refunds, and incidents enter our existing security process. We can provide the relevant policy, vendor assessment and a redacted review record.

If a control is planned, call it planned and give the accountable owner and target date. If you have assessed only support, do not imply the statement covers sales and engineering. Replace broad claims such as “all AI is compliant” with the specific scope and evidence you can support.

For the board, report decisions rather than document volume: which uses are approved, which can act on external systems, what material failures occurred, which exceptions remain open and what needs executive approval. Use counts from your own register; an incomplete inventory should be labeled incomplete.

Who Should Act Now, and What Can Wait?

Act now wherever AI handles customer information, makes commitments or can change a live system. Begin with permissions and ownership before expanding its authority.

Founders: Name the accountable executive and stop approving uses through informal chat alone. At a small company, one person can hold several roles, but each decision still needs a record.

Operators: Separate drafting from consequential actions in sales and support. A senior operator at a hypothetical 400-person company should be able to identify the owner of each workflow without tracing a chain of forwarded messages.

Builders: Reassess when you add a tool connection, change data access or deploy a materially different configuration. Existing approval for text generation does not establish approval for production writes.

Buyers: Ask for vendor evidence before granting sensitive access. Consider governance software when manual records and approvals become unreliable across the actual portfolio. The AI agent governance platforms comparison covers that buying decision.

Formal certification and a dedicated governance platform can wait if neither is required and your present controls remain manageable. A harmless internal draft does not need the same approval ceremony as an account-changing agent. Existing controls can remain in place when the purpose, data, permissions and risk remain unchanged, subject to their scheduled review.

AI Governance Training Should Follow the Job

Train people on the decisions they must make. A generic awareness presentation will not show a support employee how to stop an unsafe integration.

Give sales staff a draft containing an unsupported product promise. Give support staff an answer that mixes information from different customers. Give developers a plausible change that weakens an access check. Ask them to identify the problem, reject the output and use the correct reporting route.

Record the briefing, attendance and follow-up. Repeat the relevant part when the workflow changes or an incident exposes a gap. Treat this as an operating recommendation, not a claim that every company must buy a specific course or complete a fixed number of training hours.

What's Overhyped About AI Governance?

The most expensive mistake is confusing a governance artifact with a working control.

A policy cannot restrict a credential. If the assistant has broad write access, a sentence saying “seek approval” does not remove that access. Align the rule with the technical permission.

Human review is not automatic assurance. It needs criteria, source material, authority and a response when the reviewer finds a problem. Track missed issues and fix the process.

A vendor's assurance does not describe your workflow. Check the service and scope covered, then assess your own data access, integration and operating decisions.

A certification does not make every output correct. Use it as evidence about the assessed management system. Continue to check the behavior that can affect your customers.

A dashboard does not own a decision. Someone still has to accept a risk, narrow the use or stop it. If that person is unclear, buying another reporting layer will not resolve the gap.

Frequently Asked Questions

Will data governance be replaced by AI?

AI can help with governance tasks, but responsibility for data quality, access, permitted use and retention still needs an owner. Keep those controls and add rules for how AI uses the data and acts on its outputs.

What does GovAI do?

GovAI is a think tank that supports government and industry decision-makers through research on advanced AI and programs that develop governance talent. It is a distinct organization, not the name of a company compliance standard. GovAI's mission.

What is one of the main challenges of AI in governance?

Keeping the approved use aligned with the actual use. A system can gain a new data source or action permission while an old assessment still describes it as a drafting assistant. Make those changes trigger a review.

What is AI security governance?

It is the ownership and oversight of AI security decisions: information handling, access, testing, deployment and incident response. Connect it to the security process you already operate and name who can disable each AI workflow.

Can you give me an example of AI governance?

A support assistant drafts replies from approved documents, while a named employee checks the response before sending. It cannot issue refunds. The business records the vendor review, checks for inappropriate data access and gives staff a route to report failures.

Is AI governance a good career?

It is worth exploring if you enjoy connecting technical behavior with policy, risk decisions and evidence. Assess specific roles by their responsibilities and the skills they require; a job title alone says little about the work.

What is the AI governance salary?

There is no useful single salary for that label. Compare current posted ranges for the same location, seniority and scope, distinguishing operational risk roles, technical assurance and policy research. A blended number would hide those differences.

Which 5 jobs will survive AI?

No fixed list of five jobs can be guaranteed to survive unchanged. For a governance career, focus on skills used across roles: investigating failures, checking evidence, understanding permissions and explaining decisions to people accountable for the outcome.

How do I get into AI governance?

Start with a bounded workflow you have permission to assess. Build its inventory entry, identify possible failures, propose controls and produce evidence from a review or clearly labeled exercise. That shows the connection between a policy and its operation.

The Monday Move

Ask sales, support and engineering to bring their active AI uses to a short working session. Name the executive owner, fill the register, and select the workflow with the greatest customer or data consequence.

Trace that workflow from approved data to output, review and any action. Confirm who can stop it. Rehearse a failure, record the gaps and assign each correction to a person. Use those findings for the next customer answer or board update.

For practical guidance on running AI inside a business, join the newsletter.