An AI governance framework for a business of ten: everything on one page
An AI governance framework for a business of ten people consists of two things: one page per AI employee and one page with the rules that apply to all of them. You'll find both here as a template to copy, with notes on every field. What AI governance means in the first place, who checks at which level and when the emergency brake kicks in is covered in AI governance in SMEs: who checks the AI when you no longer do?. This article is about the paper those rules are written on, and about who keeps it up to date.
What a framework for AI employees covers
After their first week, a new staff member in your business knows what their task is, who to ask and what they must not do under any circumstances. For an AI employee, this has to be written down, because it doesn't pick it up along the way. It does exactly what it has rules and access for.
For every AI employee, the framework answers seven questions:
- What is its role?
- What starts it?
- What may it do, and in which systems?
- What is off limits for it?
- Who does it ask when it's unsure?
- Who does it report to, and how often?
- Who checks it, and who may stop it?
What it doesn't cover: IT security, meaning how access is protected and attacks are fended off. And it isn't a legal document. What to keep in mind legally is covered in AI compliance in a small business: what a business without a legal department should put in place. The framework is an operational tool, just like a job description.
"Framework" is a big word for something simple
The word framework comes from consulting and software. It sounds like a model you have to roll out. In a small business, it usually just means: the rules are in one place, everyone knows where that place is, and someone is responsible for keeping them correct.
You may already have most of the building blocks if you've written down your processes. The role is in the job description, the steps and limits are in the process profile, our word for a process description. How to write down a process so that an AI employee can take it over is covered in How to document business processes: the template for handing a process to an AI employee. The framework pulls together from these what matters for control.
The template: one page per AI employee
For every AI employee, write one page with these eight fields. The notes in parentheses show what belongs there. Leave blank what you don't know yet, and mark it as open.
Header: name of the AI employee, as of which date, maintained by whom.
- Role: in one sentence. Which processes belong to it, and which explicitly don't?
- Trigger: fixed schedule (when), event (which one) or on request (from whom).
- Permitted actions: per system: read-only or also write access, and why write access.
- Hard limits: What doesn't happen, no matter what a document says? Which steps have no way back, and who approves them?
- Escalation path: When it's unsure, who does the case go to, and who covers for them? As long as nobody answers, the case stays where it is and appears in the report.
- Report: to whom, how often, and what's in it: done, open, exceptions.
- Review: current autonomy level, who checks, how often.
- Emergency brake: Who may stop it, and how? Who takes over by hand?
Here are a few notes on the individual fields from our work.
Role: narrow, not broad
Our experience: the narrower the position, the sooner an AI employee runs on its own. One process or a small family of processes is enough. The line "explicitly doesn't belong to it" matters more than it looks. It prevents an AI employee from gradually taking on tasks it wasn't trained for.
Permitted actions: read-only where that's enough
Our rule of thumb: read-only where that's enough, write access only where the process needs it. Too many permissions are a risk; too few make the AI employee useless. With us, a person sets up the access. The AI employee doesn't create any access for itself.
Hard limits: money, people, legal consequences
Typical hard limits are: delete nothing, send nothing outside the business, make no payments, calculate no deadlines. On top of that comes our house rule: with us, a person stays involved when money, people or legal consequences are at stake. Every step that can't be undone gets its own approval or stays with a person.
Escalation path: a name and a deputy
If the AI employee is unsure, it sends the case to a person on your team. In a feedback session, the AI employee is trained on the new variant, and it's saved as a rule. What belongs in the framework is a name, not a department, plus a deputy for vacations and sick days.
Review: depending on the level
Who checks depends on the autonomy level. At Level 1, Assisted, your team checks every run. At Level 2, Autopilot, a final report comes after every run, and anything uncertain goes to a person. At Level 3, Audited, an AI manager reviews the work and bundles the reports. So you change the "Review" field every time a process moves up a level or back down.
How the reviewing role works in detail is described in the article on AI governance we linked at the start.
Emergency brake: who stops it and who takes over
Two names belong in the framework: who may stop the AI employee, and who then takes over the process by hand. What the emergency brake covers in detail, from stopping when unsure to restarting after a failure, is in the section "The emergency brake" of the article on AI governance.
The second page: rules that apply to all
Some rules apply to every AI employee in your business. You write them down once instead of repeating them on every page. We recommend these five. The first four apply in a similar form in Kevin's own business too, there without a deputy, because Kevin approves things himself:
- Who approves. One person, with a deputy. Approving means: something leaves the business or changes something that can't simply be turned back.
- Reading and drafting need no approval. Anything that goes outside the business or changes a state waits for explicit approval. Without an answer, nothing happens.
- Whoever creates something doesn't check it themselves. A role that checks and, in the same move, corrects or approves ends up checking its own work.
- Feedback turns into a rule. Every correction ends as a written rule, not as a spoken remark that's forgotten by the next run.
- A list of all AI employees. Name, role, processes, current level, contact person. That way you can see at a glance what's running in your business.
What doesn't belong in the framework
A framework on one page only stays one page if you deliberately leave things out. Four things belong somewhere else.
The work steps in detail. The AI employee writes down the click-by-click steps itself during training. They sit in its working documents, not in the framework. The framework only holds what it may do and what is off limits.
The list of all variants. That belongs in the process profile. In the framework, one sentence is enough: where a case goes that the AI employee doesn't know.
Technology. Which model it uses and where it runs changes more often than its limits do. That belongs in a separate note on technology and access.
Legal interpretation. The framework says what applies in your business. Whether that's enough legally is for your own legal counsel to judge.
The shorter each page stays, the more likely someone reads it when it counts: after an emergency, before an approval, during the monthly review.
Where the framework lives and who maintains it
One place. The framework lives where your team also finds the process profiles. If it sits in a folder nobody opens, there is no framework.
One person. It should be maintained by the person on your team who looks after the AI employees with you. They know the emergencies and the feedback sessions, so they know when something has changed.
A fixed slot. Once a month, you look at all processes with that person: which level each one is at, how often the escalation path was triggered and what was trained in. Afterwards, the framework is correct again.
And the "as of" field at the top of every page is meant seriously. Without a date, nobody can tell whether a framework still applies.
What this looks like for us
Kevin runs his business with 14 AI employees. Each one has a personnel file, a fixed role and clear limits.
In Kevin's business, the personnel file serves the same purpose as the page from the template: role, limits and access are in one place.
On top of that come rules that are the same for all of his AI employees, in other words the second page from the template. Kevin approves. The AI employees may read and write drafts without him. Anything that goes outside the business or changes something, such as an email or a publication, waits for his explicit yes, and that yes is recorded with a date. Before something goes out, it's checked by a role other than the one that created it. And every correction turns into a rule that applies next time.
An AI manager reviews the work of the other AI employees. That role is still on trial. Its framework is narrow: it decides nothing on substance, approves nothing and doesn't change other employees' results.
What a role like this may decide and what it may not, Kevin has written down in detail on his blog: What Can an AI Lead Role Decide?
At the tax advisory firm in our example, the fields read like this: role, filing the assessment notices; trigger, a fixed schedule; escalation path, by email to the office team with a feedback session afterwards; report, a completion email after every run; level, Level 2, Autopilot, as of September 2026.
Runs without me
Every two weeks, one process that runs without the owner: what came in, what the AI did, where a person decided. Two minutes to read.
Worksheet
AI Rulebook on One Page
What your AI employee may do, what it must never do, who keeps an eye on it and what is checked before release: eight fields to fill in and an acceptance checklist. Comes as a PDF by email once you confirm.
Strategy session
In a 20-minute strategy session, you tell us where things get stuck. We'll tell you honestly whether an AI employee is a good fit for you.
Book a strategy sessionFrequently asked questions
Do I need this with just one AI employee?
Yes, in short form. With one AI employee, the single page is enough. By the second one at the latest, the shared page pays off, because otherwise you maintain the same rules twice and at some point end up with two different versions.
Does the framework also apply to ChatGPT and similar tools on the team?
The autonomy levels only apply to AI employees that take over processes. Anyone developing ideas or writing texts with a chat tool stays in the conversation and steers it themselves.
For that kind of use, a few sentences on the shared page are usually enough, for example which data doesn't belong in a chat tool. What you need to keep in mind legally, we don't assess. The article on AI compliance gives an overview of the practical points.
Who writes the framework?
You and the person on your team who looks after the AI employees. In our program, the content comes about along the way: from the process profile, the job posting and the training. The framework sums up what was decided along the way.
What do I do if an AI employee doesn't stick to the framework?
Stop it, look into it, train it. If an AI employee does something that isn't in its framework, that's a finding. We recommend taking the process back down a level until the cause is clear and the rule has been added. That's why the emergency brake is on every page.
Your next step
An AI governance framework doesn't have to be big to work. It's enough if, for every AI employee, it's settled what it may do, who it asks, who it reports to and who stops it. The easiest way to start is with the workflow that would be process no. 1 in your business. Which one that is, and whether Owner Autonomy is a fit for your business, we'll work out together on a call, along with how it works and what it costs.