How to build an AI governance framework for small firms
An AI governance framework for small firms is a short operating agreement for how the firm uses AI. It names the inputs and tools the team may use, the actions that require human sign-off, and the evidence the firm retains. For a professional-services firm, the core can be three written rules, one named decision owner, clear approval thresholds, and an approval record. This is an operating framework, not a certification or legal opinion.
Written by Loïc Guyon (Tileo), operator of Pitstop.
On this page
- What should your framework contain?
- How can a small firm apply NIST without copying an enterprise program?
- Which work needs which approval threshold?
- How do you put the framework into use?
- What does exact-text approval look like in practice?
- Where do tools help, and where is a written rule enough?
- How should you review the framework?
- FAQ about a small-firm AI governance framework
What should your framework contain?
Start with decisions that appear in daily work. A partner or operations lead should be able to read the framework and tell whether an agent may draft a meeting summary, what happens before candidate outreach leaves the firm, and who handles an exception.
The table below is a compact starting point. Replace each role label with a person inside your firm.
| Decision | Default rule | Named owner | Evidence retained |
|---|---|---|---|
| Approved tools and data | Use only the tools and input types listed in the firm's written rule. | Operations lead | Current rule and the approval for each exception |
| External communications | A person approves the exact text before it leaves the firm. | Mandate or account partner | Approved text, approver, and approval record |
| High-impact client or candidate work | A partner reviews the output before the team relies on it or sends it. | Responsible partner | Reviewed output and sign-off |
| Exceptions | The named owner decides whether the exception is allowed. | Decision owner | Request, decision, and decision owner |
| Periodic review | Revisit the rules on the schedule chosen by the firm. | Managing partner or operations lead | Review decision and any revised rule |
The framework needs three written rules:
- State which tools and input types the firm approves.
- State which actions need human sign-off.
- State what evidence the firm retains after approval.
Then name the person who owns decisions when the rules do not settle a case. Record that person's decision with the exception.
Your acceptable-use rule can stay separate and feed this framework. Use this AI acceptable use policy template to define allowed inputs and uses. If unapproved tools already appear in team workflows, the guide to shadow AI in professional-services firms helps frame that operating problem.
How can a small firm apply NIST without copying an enterprise program?
The NIST AI Risk Management Framework is voluntary, rights-preserving, non-sector-specific, and use-case agnostic. NIST says that the framework helps organizations manage AI risks and promote trustworthy and responsible development and use.
The NIST AI RMF Playbook groups suggested actions under Govern, Map, Measure, and Manage. Use the Playbook as reference vocabulary, not as a legal mandate or certification.
Apply that vocabulary to the decisions in your table:
- Govern: Write the rules, name the owner, and set the approval thresholds.
- Map: Place the firm's work into the agreed risk tiers.
- Measure: Check whether each tier has an approval threshold and specified evidence.
- Manage: Approve, revise, reject, or document an exception, then retain the decision evidence.
For a small firm, each term can describe an operating step. It does not require an enterprise committee, a model inventory, or a compliance suite.
Which work needs which approval threshold?
The threshold should follow the consequence of the work. An internal draft and a message to a candidate do not call for the same decision.
| Tier | Professional-services examples | Default threshold | Evidence retained |
|---|---|---|---|
| Lower impact | Drafting an internal meeting summary or organizing internal notes | Team member checks the output before relying on it | Checked draft when the work feeds a later decision |
| Client or candidate facing | Candidate outreach, client updates, and mandate briefs | Responsible person approves the exact text before it leaves | Exact approved text and approver |
| High impact | A proposal, a candidate assessment, or work that shapes a client decision | Partner approves the output | Approved output and partner sign-off |
| Exception | Any use outside the approved tool, data, or action rules | Named decision owner decides before the use proceeds | Exception request and decision |
These tiers are decision labels, not declarations that any use is safe. They tell the team who decides and what evidence to keep.
Write the threshold in observable terms. "Use judgment" does not tell an associate whether a client update needs approval. "The account partner approves the exact text before it leaves" does.
How do you put the framework into use?
Use a short implementation checklist:
- List the AI tools and input types that the firm approves.
- Sort the covered work into the risk tiers in the table.
- Assign one named owner to each decision.
- Define the point at which a person must approve the work.
- Record the exact output that the person approved.
- Define how the owner handles and records exceptions.
- Choose a review schedule and record each review decision.
Keep the wording close to the work. "Candidate outreach requires exact-text approval from the mandate partner" gives the team a decision rule. A broad statement about responsible AI does not settle that case.
What does exact-text approval look like in practice?
Consider an illustrative client update. An agent drafts the message. The draft does not leave the firm.
The account partner reads and approves the exact text. Approval applies to the final text that the partner sees.
The firm retains the final approved text, the approver, and the approval record. If the sent content can be compared with the approved version, the journal can show whether the content differs from the approved fingerprint.
That sequence gives the firm evidence for the decision it made. It does not turn the tool into the decision owner. The partner still owns the approval.
Where do tools help, and where is a written rule enough?
A written rule plus a named owner is enough to define approved tools and data, assign approval responsibility, set thresholds, and decide exceptions. Software can preserve the exact approved text, the approver, and the journal entry in one workflow.
Pitstop is one implementation option. Agents ask before anything leaves, and a person approves the exact text. The journal can show whether the sent content differs from the approved fingerprint. Pitstop does not currently block or send the message itself.
Keep the distinction clear. Governance is the firm's decision system. A tool can support the approval record, but it does not replace the written rule or the named owner. The same distinction matters when firms assess autonomous AI agents: the key question is which actions the agent may take and where a person must decide.
How should you review the framework?
Use the review schedule chosen by your firm. At each review, work through the decision table and record whether each rule stays, changes, or needs an exception path.
Review these questions:
- Do the approved tools and input types match current practice?
- Does each external action have a clear approval threshold?
- Does every threshold name a person who can decide?
- Can the firm retrieve the approved output and its approval record?
- Have exceptions exposed a rule that needs revision?
Do not turn the review into a search for paperwork. The useful output is a current decision rule that a team member can follow during client or candidate work.
FAQ about a small-firm AI governance framework
Is there a sample AI governance framework to copy?
Yes. Copy the first table in this article, replace the role labels with named people, and adapt the default rules to your work. Keep the rows for approved tools and data, external communications, high-impact work, exceptions, and periodic review.
Do we need AI governance tools before we start?
No. Start with written rules, named owners, approval thresholds, and the evidence to retain. Add a tool when the firm needs support for applying those decisions and preserving approval records.
Can this become an AI governance framework PDF?
Yes. Once the owners and rules are agreed, export the decision table and supporting rules as a PDF for distribution. The decision owner retains the current version and its review record.
What is the difference between an AI governance policy and this framework?
A policy states the firm's rules. This operating framework connects those rules to decisions, owners, thresholds, and evidence in daily work. The policy can say which input types are approved. The framework says who decides when a proposed input falls outside that rule.
An AI governance framework for small firms does not have to begin with enterprise machinery. Begin with the decisions your team already faces. Write the rule, name the owner, set the threshold, and retain what the person approved.