How to Build an AI Governance Framework From Scratch

Shubham S.
October 13, 2026
•
20
mins

A prospect's security questionnaire arrives with a new line: "Please provide your AI governance framework." You ship two LLM-powered features and your team uses four AI tools nobody formally approved. What you have is a Slack thread and good intentions.

An AI governance framework is the set of roles, inventory, risk rules, policies, review gates and monitoring that decides how a company builds, buys and uses AI. You build it in seven steps, in a fixed order, because each step produces the input the next one needs.

Quick take For a team of 50 or fewer, a framework that holds up is small: one named owner, a spreadsheet of every AI system, three risk tiers, five short documents, a pre-launch review question list, and a quarterly review with notes. You can have all of it in place in weeks, not quarters.

By the end of this guide, you will be able to build that framework yourself and hand a customer, an investor or an auditor a clear answer.

Here's what I'll cover:

  • What an AI governance framework is, and how it differs from a policy and from ISO 42001
  • Why building one before a customer asks costs less
  • The seven steps in build order: roles, AI inventory, risk tiers, policies, a review gate, monitoring and a review cycle
  • How the result maps to ISO 42001, the NIST AI RMF and the EU AI Act
  • The mistakes that quietly break these programs

What an AI Governance Framework Is, and What It Is Not

An AI governance framework is the operating system for AI inside your company: it names who is accountable, lists every AI system, sorts them by risk, sets the rules, checks new AI before launch and watches it afterward. It covers AI you build, AI you buy and AI your employees pick up on their own.

That last group matters most for early-stage SaaS teams. Most of the AI risk in a ten-person company sits in the tools and vendor models it uses, not in a model it trained.

A framework is not a single document. It is a program with a few moving parts, and a policy is one of them.

AI Governance Framework vs AI Governance Policy vs ISO 42001

These three get mixed up constantly, and they answer different questions.

  AI governance framework AI governance policy ISO 42001
What it is The whole program: roles, inventory, tiers, gates, monitoring One document stating the rules A certifiable management system standard
Question it answers How do we run AI responsibly day to day? What are the rules, and who must follow them? Can an outside auditor confirm our system works?
Proof you hold it Records, reviews, evidence The signed, versioned document A certificate from an accredited body
Where to go deeper This guide ComplyJet's AI governance policy guide What is ISO 42001

So when this guide says "write the policy," it points you to the policy guide for the document itself and spends its words on the program around it. If a buyer eventually asks for a certificate, you can build this framework toward ISO 42001 without starting over.

ISO 42001
Build the framework once, certify against it later
ComplyJet's ISO 42001 workspace includes AIMS policy and procedure templates, an AI risk assessment framework, AI impact assessments and automated evidence collection.
See ISO 42001 support

Why Build an AI Governance Framework Before a Customer Asks

The cheapest time to build an AI governance framework is before anyone asks for it. After a questionnaire lands, you are reconstructing history under a deadline. Before, you are writing down decisions you are already making.

Three pressures are converging on SaaS teams at once:

  • Customer diligence. Enterprise buyers now add AI questions beside the SOC 2 ones: which models you use, where prompts go, who approved it.
  • A gap your audit report leaves open. SOC 2 was not built to test fairness, safety or model behavior, as ComplyJet's guide to whether SOC 2 covers AI explains.
  • Regulation with dates attached. The EU AI Act applies to any company whose AI is placed on the EU market or whose output is used there, wherever it is based.

On dates, be careful which ones you quote. Under Regulation (EU) 2026/1744, the Digital Omnibus on AI, in force since 27 July 2026, high-risk obligations for Annex III systems now start on 2 December 2027, and for AI embedded in regulated products on 2 August 2028.

Prohibited practices have applied since February 2025, and transparency duties since 2 August 2026. The full timeline is in our EU AI Act compliance guide, and the European Commission's AI Act Service Desk timeline tracks the official dates.

Watch out The Omnibus delay moved high-risk dates only. The prohibited-practice bans and transparency duties were not pushed back, and AI literacy expectations were softened rather than delayed. A framework that waits for 2027 starts behind.

None of this requires a large program. It requires an order of operations.

How to Build an AI Governance Framework in Seven Steps

Here is how to build an AI governance framework without stalling: follow these AI governance framework steps in order, and let each one produce a concrete artifact you can show someone.

Seven building blocks of an AI governance framework in build order: roles, AI inventory, risk tiers, policies, review gate, monitoring and review cycle.
  1. Assign roles. One accountable owner and clear backups.
  2. Build the AI inventory. Every system, vendor model and AI feature, with an owner.
  3. Set risk tiers. A rule that sorts the inventory into three levels.
  4. Write the policies. A short document set tied to the tiers.
  5. Add a review gate. A checklist new AI passes before it ships.
  6. Set up monitoring. Watch quality, drift and incidents per tier.
  7. Review and keep evidence. A quarterly cycle with records.

AI Governance Framework Steps, in Build Order and Why That Order

The order is not arbitrary. You cannot tier systems you have not listed, and you cannot list them without someone accountable for the list. Policies written before tiers end up generic, and monitoring written before a review gate watches things nobody approved.

Step What you produce Typical owner Rough effort on a small team
1. Roles One-page responsibility table Founder or CTO Half a day
2. AI inventory Spreadsheet of AI systems AI owner 1 to 3 days
3. Risk tiers Tiering rule plus tiered inventory AI owner with security 1 day
4. Policies Five short documents AI owner with legal 3 to 5 days
5. Review gate Intake form and approval path Product and security 1 day
6. Monitoring Checks per tier, incident path Engineering 2 to 5 days
7. Review cycle Calendar, minutes, evidence folder AI owner 1 day, then quarterly

The effort column is my own estimate for a team of 50 or fewer, not a benchmark. Your first inventory will take longer if people have used AI tools you did not know about.

Step 1: Assign AI Governance Roles

AI governance roles come first because every later step needs an owner, and a missing owner is the usual reason these programs die. Name one accountable person, not a committee. A committee can come later, once you have enough AI systems to justify one.

The owner does not need to be an AI expert. They need authority to say no to a launch and the time to run a quarterly review. At a startup that is usually the CTO, head of security or a technical co-founder.

Note ISO 42001 requires top management to commit to its AI management system and assign roles and responsibilities (Clause 5), and the NIST AI RMF's GOVERN function expects executive leadership to take responsibility for AI risk decisions. A named owner satisfies the spirit of both.

AI Governance Roles for a Team of Fifty or Fewer

Small teams double up. The point is that every row has a name next to it, even if one person fills three rows.

Role Accountable for Typical holder
Executive sponsor Approves the framework and its budget; signs off on top-tier AI CEO or CTO
AI governance owner Runs the program, the inventory and the quarterly review CTO, head of security
Security lead Vendor and data-handling review; incident response Head of security or IT
Legal and privacy Regulatory mapping, customer terms, data processing terms Counsel or outsourced DPO
System owners Intended use, testing and monitoring of one AI system each Product or engineering lead
Everyone else Follow the acceptable-use rules; report problems All staff

Verdict: one name beats a five-person committee. If you cannot say who would be called when an AI feature misbehaves at 2 a.m., you do not have a role structure yet.

Step 2: Build the AI Inventory Your Framework Runs On

An AI inventory is a living list of every AI system your company builds, buys or allows. Everything downstream, from risk tiers to monitoring, runs off it. If a system is not on the list, it is not governed.

Include three kinds of entries:

  • AI you ship: features in your product that call a model or run your own
  • AI you embed from vendors: the AI features switched on inside tools you already pay for
  • AI staff use on their own: chat assistants, browser extensions, note-takers

The third kind is the one teams miss. A Hacker News commenter wrote in May 2026 that the model card "quickly became a standard component of AI governance frameworks," which is a reminder that documentation habits, not just policies, are part of what an inventory feeds.

AI Inventory Fields That Customers and Auditors Ask For

Keep the schema small enough that people will actually fill it in. These fields cover what a customer or auditor is likely to ask:

Field Why it exists
System name and owner Someone is accountable by name
Purpose and intended use The basis for judging whether it is used as designed
Out-of-scope uses What it must never be used for
Vendor and underlying model You inherit your model provider's risks
Data it touches Customer data, personal data, secrets, public data
Who it affects Employees, customers, applicants, the public
Risk tier Set in step 3
Approval date and approver Evidence the review gate ran
Last review date Evidence the quarterly cycle runs
Try this yourself Pull your SSO app list, last quarter's software expense lines and the browser extensions installed on company laptops. Search each for "AI," "assistant," "copilot" or "GPT." Whatever turns up that is not already on your list is your first shadow AI finding.

The NIST AI RMF names this artifact directly: its GOVERN function includes keeping an inventory of AI systems. We cover the framework in our NIST AI Risk Management Framework guide.

Step 3: Set AI Risk Tiers for Your AI Governance Framework

AI risk tiers let you spend effort where it matters. A summarizer for internal meeting notes and a model that screens job applicants should not face the same review, and a framework that treats them alike will be either too heavy to survive or too light to defend.

Three tiers are enough for most small teams. Use a rule anyone can apply, not a judgment call:

Tier Test (any one is enough) Minimum controls
High Influences decisions about people (hiring, credit, access, pricing), or handles sensitive personal data, or acts without human review in a customer-facing flow Executive sign-off, documented evaluation, human oversight, ongoing monitoring, quarterly review
Medium Customer-facing output with a human in the loop, or touches customer data, or depends on a third-party model Owner sign-off, vendor review, data-handling check, periodic spot checks
Low Internal use, no customer or personal data, easy to reverse Acceptable-use rules and an inventory entry
Three-tier ladder for AI risk tiers: low, medium and high, with the review effort rising at each level.

Your internal tiers do not need to match the EU AI Act's legal categories. They need to be at least as cautious. The legal classification of a feature, including which Annex III areas count as high-risk, is covered in our EU AI Act risk classification guide. Run that test on any system you rate High, and record the result in the inventory.

Verdict: when in doubt, tier up. Moving a system down later with a written reason is easy. Explaining an under-tiered one after an incident is not.

Step 4: Write the Policies Your AI Governance Framework Needs

Policies turn the tiers into rules. Resist the urge to write one 40-page document. A short set, each answering one question, is easier to keep current and easier for staff to read.

AI Governance Framework Template: The Minimum Document Set

Treat this AI governance framework template as the paper layer of your program. Five documents cover most small-team needs:

Document The question it answers Typical length
AI policy What is our position, scope and the principles we hold to? 1 to 2 pages
Acceptable use rules What may staff do with AI tools, and with which data? 1 page
Third-party AI rule How do we approve and review vendor AI and model providers? 1 page
AI change and release procedure What review applies when a model, prompt or data source changes? 1 to 2 pages
AI incident path Who is called, and what happens, when an AI system fails? 1 page, or a section of your existing incident plan

For the policy itself, ComplyJet's AI governance policy guide includes a full template and sample text. Use it rather than starting from a blank page.

ComplyJet customers
How an early-stage AI company runs compliance day to day
Raghuram Vadapally, Managing Director and Head of Software Engineering at Synexar AI, calls ComplyJet "an indispensable part of our compliance workflow." Read their SOC 2 and HIPAA story.
Read the Synexar AI story
Pro tip Put the version and date in each document's header, and name which tier each rule applies to. When a customer asks for the "AI policy," you can send a dated, versioned file instead of a link to a wiki page that changes weekly.

Step 5: Add a Review Gate for New AI Before It Ships

A review gate is the checklist every new AI system or major change passes before it goes live. It is where the framework stops being paperwork and starts changing decisions. Without it, the inventory is a record of what already happened.

Review gate flow: a new AI system is proposed, tiered, reviewed against intake questions, then approved, approved with conditions or rejected, and added to the inventory.

Keep the intake to questions the owner can answer in ten minutes:

  1. What is the system for, and what must it never be used for?
  2. Which model and vendor sit underneath it, and under what data terms?
  3. What data goes in, and what comes out to whom?
  4. Who is affected if it is wrong, and how badly?
  5. How was it tested, and what would you rerun after a model or prompt change?
  6. Can a human override it, and can you switch it off in minutes?
  7. What tier does the rule in step 3 assign?

Then route by tier: Low tier is approved by the system owner and logged, Medium by the AI governance owner, and High by the executive sponsor. A decision is one of three outcomes: approve, approve with conditions, or reject. Log which, with a date.

Note Question 5 matters because the system you approved is not the system running six months later. A model version bump or a prompt rewrite can change behavior without a code release, so changes to either should re-trigger the gate for Medium and High tier systems.

Step 6: Set Up AI Governance Monitoring and Incident Response

AI governance monitoring is how you notice that an approved system has started behaving differently. A review gate checks a system once. Monitoring checks it for as long as it runs, which is where most real failures show up.

Scale the effort to the tier:

What to watch High Medium Low
Output quality against your test set Every release plus scheduled reruns After model or prompt changes Not required
Vendor or model changes Tracked and reviewed Tracked Noted at quarterly review
Unusual usage or data flows Alerting Periodic review Not required
User complaints and overrides Logged and reviewed Logged Logged
Kill switch Required, rehearsed Required Not required

Two parts deserve more attention than they usually get. First, a kill switch: a feature flag or config that turns the AI behavior off in minutes without a deploy. Second, an incident path. You likely already have one for security; extend it to cover an AI failure, such as a leaked prompt, a harmful output or a vendor outage. ComplyJet's incident management policy guide covers building that foundation.

Pro tip Rehearse the kill switch once. A switch that exists in a runbook but has never been flipped tends to fail the first time someone needs it.

Step 7: Review the AI Governance Framework and Keep the Evidence

A framework that is never reviewed decays into a document. Put a quarterly review on the calendar, chaired by the owner, with three standing questions: what AI did we add or change, what did monitoring show, and what should we tier up or down?

Also define triggers for an off-cycle review: a new high-tier system, a vendor switching its model or terms, an AI incident, or a change in law that touches you. Write down the decisions. Minutes are evidence.

What you keep is what a customer or auditor will eventually ask for. A short evidence folder covers it:

  • The current role table and the dated policy set
  • The inventory with tiers, owners and last-review dates
  • Approved review-gate forms for Medium and High systems
  • Test results and monitoring summaries for High systems
  • Quarterly review minutes and any incident records
  • Staff training records on acceptable use

Verdict: if you can produce those six items in an afternoon, you can answer almost any AI governance questionnaire honestly.

Which AI Governance Frameworks to Build On: ISO 42001, NIST AI RMF and the EU AI Act

You do not need to invent your structure. The established AI governance frameworks give you vocabulary, and your seven steps line up with all three. The table maps each building block to where it shows up.

Building block ISO 42001 NIST AI RMF EU AI Act
Roles Clause 5 Leadership; Annex A.3 Internal organization GOVERN Provider and deployer role duties; AI literacy expectations
AI inventory Clause 4 context and scope; Annex A.4 Resources GOVERN (inventory of AI systems); MAP Needed to apply any classification
Risk tiers Clause 6 Planning; Clause 8 risk assessment MAP, MEASURE Four-tier classification; Annex III
Policies Annex A.2 Policies related to AI GOVERN Documentation duties for high-risk systems
Review gate Clause 8 impact assessment; Annex A.5 MAP, MEASURE Risk management system for high-risk systems
Monitoring Annex A.6 and A.9; Clause 9 MANAGE Post-market monitoring, logging, human oversight
Review cycle Clauses 9 and 10 GOVERN Ongoing for high-risk systems

The mapping is my own reading, not an official crosswalk, and a certification auditor will test against the standard's exact text.

Which to choose comes down to what you need to prove:

  • Build toward ISO 42001 if buyers ask for a certificate. What is ISO 42001 covers the standard, and the ISO 42001 checklist lists the controls. Already running ISO 27001? ISO 42001 vs ISO 27001 shows what carries over, and ISO 27001 and AI compliance covers the security side.
  • Align to the NIST AI RMF if customers want a vocabulary and no certificate is needed.
  • Add EU AI Act checks if EU customers or users are in scope.
  • Otherwise, finish steps 1 to 3 this month. They cost almost nothing.

Mistakes That Break an AI Governance Framework

These are the failure modes I would watch for first:

  • Starting with the policy. A policy written before the inventory describes AI you imagine, not the AI you run.
  • Governing only what you build. Vendor AI and staff-adopted tools carry most small-team risk.
  • A committee with no owner. Shared accountability means nobody gets called.
  • One tier for everything. Uniform heavy review gets bypassed; uniform light review proves nothing.
  • Approving once and never again. Model and prompt changes alter behavior after sign-off.
  • No evidence trail. A program you cannot show is a program you cannot defend.

Each of these shows up as a weak answer to a customer question. Fixing them early is cheaper than fixing them in front of a buyer.

FAQs

What Is an AI Governance Framework?

It is the set of roles, inventory, risk rules, policies, review gates and monitoring that decides how a company builds, buys and uses AI. A policy is one document inside it. The framework is the program that makes the policy real.

How Do You Build an AI Governance Framework?

In seven steps, in order: assign roles, build the AI inventory, set risk tiers, write the policies, add a review gate, set up monitoring and run a quarterly review with evidence. Each step produces the input for the next.

What Are the Components of an AI Governance Framework?

Accountable roles, an AI inventory, risk tiers, a short policy set, a pre-launch review gate, monitoring with an incident path, and a review cycle with records. Larger organizations add committees and tooling on top of these.

Who Is Responsible for AI Governance in a Company?

One named owner, backed by an executive sponsor. At a startup that is typically the CTO or head of security. System owners are responsible for individual AI systems, and legal and privacy advise on regulation and contracts.

What Is the Difference Between an AI Governance Framework and a Policy?

The policy states the rules. The framework is everything that applies them: owners, inventory, risk tiers, reviews and monitoring. You need both, and the framework is what produces the evidence behind the policy.

Do Small Companies Need an AI Governance Framework?

Not as a legal matter in most cases, but any company shipping AI or using vendor AI on customer data will meet customer questions about it. A small framework answers them honestly and is quick to build.

How Long Does It Take to Build an AI Governance Framework?

My estimate for a team of 50 or fewer is a few weeks of part-time effort for the first working version, assuming no surprises in the inventory. Treat the quarterly review as the ongoing cost.

Which Framework Should You Base Yours On?

Base it on what you need to prove. ISO 42001 gives you a certificate, the NIST AI RMF gives you a shared vocabulary, and the EU AI Act sets legal duties if EU customers or users are in scope. Many teams use the NIST language and certify to ISO 42001 when asked.

Book a demo
Put your AI governance framework on solid ground
Talk through your AI inventory, your existing SOC 2 work and whether ISO 42001 makes sense for a team your size.
Book a demo

Related Reading