StackLift
All playbooks

Playbook

AI Governance Policy Checklist

Use this checklist to write a short AI governance policy your team will follow. It covers what to inventory, who owns what, which uses need review, and how the policy stays current.

Best reader

Founders, CTOs, legal and operations leads introducing AI tools or AI features

Outcome

A one-to-three page AI policy with named owners, risk tiers and a review rhythm.

Use this sequence

1

Inventory every AI tool and AI feature, including ones inside software you already use.

2

Sort each use into a risk tier and note which laws or standards you may need to check.

3

Assign an owner and an approver for every use.

4

Write the data rules: what may and may not be sent to a model.

5

Define human review points and an incident path.

6

Publish the policy, train the team and set a review date.

Start with an inventory

You cannot govern what you have not listed. Governance starts with a plain list of where AI is used today.

Tools staff use on their own

AI features inside existing software

Models called by your own products

Who owns each and what data it sees

Tier the risk

Not every use needs the same scrutiny. A tiered approach keeps low-risk uses fast and puts review where it matters.

Low: drafting, summarizing internal notes

Medium: customer-facing answers

High: decisions about people, money, health or legal status

Keep the policy short and living

A policy nobody reads protects nobody. Write it in plain language and review it on a schedule.

One owner for the policy

A review date on the page

A simple way to request a new AI use

Start in 60 seconds

Turn this checklist into a scoped plan.

Answer a few questions and get an AI-assisted, architect-reviewed scope, cost range, and timeline for your software project.

Frequently asked questions

How long should an AI governance policy be?

Short enough that people read it. Most teams do well with a few pages that state the rules, the owners and the process for approving new uses, with detail kept in linked procedures.

Does a small company need one?

Yes, a brief one. Even a few people using AI tools with customer data benefit from written rules on what may be shared and who approves new uses.

The complete guide

AI Governance Policy Checklist: the complete guide

Updated November 18, 2026 ยท Published by , Founder, StackLift AI. Written with AI assistance; how we make content.

Short answer

An AI governance policy is a short written set of rules for how your organization buys, builds and uses AI. A usable one lists every AI use, sorts them by risk, names an owner and approver for each, sets data rules and human review points, and has a review date. Start with an inventory, not a template.

What is an AI governance policy?

An AI governance policy sets out who may use AI, for what, with which data, and who answers for the result. It is a working document for staff and engineers, not a statement of values. If people cannot find the rule that applies to their task in a few minutes, the policy is too long.

Good policies are built on a risk-based idea that appears in most frameworks: the more a system can affect people, money, safety or legal rights, the more review it needs. Frameworks such as the NIST AI Risk Management Framework, ISO/IEC 42001 and the EU AI Act all express this in different words. You may need to check which of them apply to you. This guide describes them in general terms only and is not legal advice.

What are the steps to write an AI governance policy?

  1. Inventory every AI use. Output: a register of tools, features and models, with the data each touches. Involve IT, product and department heads.
  2. Tier the risk. Output: each use marked low, medium or high with a one-line reason. Involve legal and the business owner.
  3. Assign owners. Output: a named owner and approver per use. Involve the leadership team.
  4. Write the data rules. Output: what may, may not and must never be sent to a model. Involve security and legal.
  5. Define human review and incident handling. Output: review points and a route for reporting problems. Involve operations and support.
  6. Publish, train and schedule the review. Output: a dated policy, a short briefing for staff and a calendar entry. Involve the policy owner.

The sections below turn each step into items you can copy. If you are earlier in the journey, the enterprise AI readiness checklist covers data, people and process first.

How do you inventory AI tools and features?

Start by asking, not by assuming. Staff often use AI tools that IT never approved, and software you already pay for may have added AI features. Ask each team what they use, then check billing and browser extension lists for confirmation.

AI register: one row per use
FieldWhat to recordWhy it matters
Use and ownerWhat the AI does and who is responsibleSomeone must answer for it
Tool or modelProduct, provider and versionNeeded for vendor review and change tracking
Data touchedTypes of data sent or retrievedDrives data rules and privacy checks
Users affectedStaff, customers or the publicDrives the risk tier
Decision impactDrafts only, suggestions or automated actionsShows whether a human reviews the output
Approved by and dateApprover name and review dateGives you an audit trail

Include AI you build as well as AI you buy. For products with agents or retrieval, the agentic workflow design playbook helps you describe what each agent may do.

How do you sort AI uses into risk tiers?

Use three tiers and keep the criteria simple. The goal is to let low-risk uses move fast and to focus attention on the few that deserve it. Write the criteria into the policy so the decision is repeatable.

A simple three-tier model
TierTypical usesControls to apply
LowDrafting, summarizing internal notes, brainstorming without sensitive dataBasic data rules, no approval beyond the manager
MediumCustomer-facing answers, internal search over company documents, code generationOwner and approver, testing, logging, human review of exceptions
HighDecisions about hiring, credit, health, legal status or safety; automated actions with money or recordsFormal assessment, documented testing, human decision, audit trail, legal review

Some regulations classify AI systems by use, and certain uses may be restricted or carry extra duties. The EU AI Act, for example, sorts systems into risk categories. Ask counsel whether any of your uses fall into a regulated category in the places you operate. Do not rely on a template to answer that.

What data rules and vendor checks belong in the policy?

Most AI incidents begin with data going somewhere it should not. Write the rules in plain words and give examples. Then check each vendor against the same rules before approving it.

  • Never send: passwords, API keys, secrets, and any data a contract or law forbids sharing.
  • Send only with approval: customer personal data, financial records, health data and unreleased product plans.
  • Free to use: public information and your own non-sensitive drafts.
  • Prefer enterprise or API terms that state how prompts and outputs are stored, and whether they are used to train models.
  • Record where data is processed and which subprocessors are involved.
  • Keep a way to delete data and to export logs if a customer or regulator asks.
Vendor questions before approval
QuestionWhat to look for
How is our data used and kept?Clear terms on retention and on training with your inputs
Where is it processed?Regions and subprocessors you can accept
What security evidence exists?Reports or attestations the vendor provides; you check what they cover
Can we turn features off?Admin controls and audit logs
What happens when the model changes?Notice of changes and a way to test them

Our security overview describes how StackLift handles delivery. For model providers you may call through an API, see the OpenAI integration guide and the Claude integration guide.

Where do humans stay in the loop?

Policy should say which outputs a person must check before they take effect. The test is simple: if the AI is wrong, who gets hurt and how fast? The higher the harm, the earlier a person should review. The human-in-the-loop glossary entry explains common patterns.

Review points by impact
Impact if wrongReview patternEvidence to keep
Minor and easy to undoSpot checksSample results
Customer sees itReview exceptions and low-confidence answersLogs of flagged cases
Money, records or legal rightsA person approves each actionApprover, time and decision
People's access to jobs, credit or careA person decides, using AI only as inputDocumented reasoning and an appeal route

Also name the owner of each AI system, and make clear that responsibility does not move to the vendor. Staff should know how to report a bad output, a data leak or a suspicious request, and who acts on the report.

How do you test, monitor and handle incidents?

Governance without testing is paperwork. Before an AI use goes live, test it on realistic examples, including the awkward ones. After launch, watch for drift: models change, data changes, and users find new ways to use the system.

  • Keep a fixed test set for each medium or high use and rerun it when the model, prompt or data changes.
  • Log inputs, outputs and decisions where privacy rules allow, so you can investigate.
  • Track complaints, overrides and escalations as signals of quality.
  • Define what counts as an AI incident: harmful output, data exposure, wrong automated action, or a regulatory complaint.
  • Name an incident owner and write the first three steps: contain, assess, notify.
  • Hold a blameless review after each incident and update the policy.

If your AI features are part of a product you ship, the production readiness checklist covers monitoring and rollback in more detail.

Which standards and laws should you check?

Which rules apply depends on where you operate, what you build and who your users are. Treat this table as a list of things to check with your own advisers, not a statement of what applies to you.

Frameworks and rules to check
SourceWhat it isWhere it helps
NIST AI Risk Management FrameworkA voluntary framework organized around governing, mapping, measuring and managing AI riskStructuring your risk process
ISO/IEC 42001A management system standard for AI that organizations can be assessed againstBuilding a formal program
EU AI ActEuropean regulation that sorts AI systems by risk and sets duties by categoryChecking whether a use is regulated
GDPR and other privacy lawsRules on personal data, including automated decisions in some casesData rules and rights requests
Sector rulesRequirements in finance, health, employment and other fieldsHigh-impact uses

Dates and details in this area change, so confirm current requirements from the official sources or with counsel. StackLift does not claim any certification against these; it builds delivery controls into the plan, and you decide which obligations apply to you.

How do you roll the policy out and keep it current?

Publish the policy where people work, not only in a shared folder. Brief each team for fifteen minutes with examples from their own tasks. Give people an easy way to ask for approval of a new AI use, and answer quickly. If approval is slow, people will go around it.

  • Put a review date and an owner on the first page.
  • Review at least once a year and after any incident, regulation change or major new use.
  • Keep a change log of what changed and why.
  • Add new tools to the register as part of procurement.
  • Report to leadership on uses by tier, incidents and exceptions.

If you are planning an AI build and want governance designed in from the start, run the free estimate or contact us to talk through scope.

Questions people ask

Who should own the AI policy?

One named person with authority across teams, often a CTO, head of operations or a risk or legal lead. Legal, security and the business owners contribute, but a single owner keeps the policy current.

Do we need a policy if we only use ChatGPT-style tools?

Yes. The risk is often data leaving the company through everyday use. A short policy on what may be pasted into which tools, and who approves new tools, covers most of it.

What is the difference between AI governance and AI ethics?

Ethics states principles such as fairness and transparency. Governance turns them into owners, rules, tests and records that you can check. You need both, but governance is what people can follow day to day.

How do we handle shadow AI?

Make the approved route easier than the workaround. Offer sanctioned tools, publish a quick approval path and explain the data rules. Then ask teams what they use and add it to the register without blame.

Does the EU AI Act apply to companies outside Europe?

It can, depending on where the system is placed on the market or where its output is used. Check with counsel whether your products or customers bring you into scope.

How often should the policy be reviewed?

At least annually, and after any incident, new regulation or significant new AI use. A fixed date on the first page and a named owner keep the review from slipping.

Sources

Sources and further reading

The standards, rules, research and official documentation this page draws on.