What are the best practices for AI governance in a small company?
Treat governance as something you maintain, not a document you finish. Most of it lives outside the AI policy, in the account settings, the instructions given to AI agents and the checks wired into workflows, and those drift as vendors change defaults, the work changes shape and new people inherit rules they never agreed to.
Name one owner, check every place a rule lives at least quarterly, and bring the review forward when you add a tool, take on a new kind of client work or a vendor changes its terms.
Questions as you read? Ask, or take the policy template with you.
Where should we send it?
Sent ✓ Download it now →
In September I wrote that an AI governance framework is the set of written answers an organization gives to four questions. Which tools may be used. What data may go into them. Who releases AI-assisted work. What gets checked afterward.
Those are still the right questions, and writing the answers down is real work that half of employers have not done.
What I left out is what happens to those answers afterward.
A written answer is accurate on the morning it is written. It describes the tools as they were configured that day, the work as it was being done that quarter, and the people who were in the room when the decisions were made. Every one of those keeps moving. The file does not.
The middle figure is the one I did not expect. Charter’s researchers surveyed just over five thousand people who use AI at work and asked which of the resources their employer provides had helped them. Hands-on training came first by a distance. Policies and guidelines came last of everything they tested.
The people the document is written for rank it below every other kind of help they are given. That is not a verdict on what the policies say. It is a verdict on what a document does when it sits still.
Where does a company’s AI governance live?
In more places than the policy, and the policy is usually the only one anybody reviews.
The file holds the decisions somebody wrote down. The rest sits where those decisions were implemented: the data settings inside each account, the instructions sitting in an agent’s configuration, the checks wired into a workflow, the approval step added to a template two quarters ago and never revisited.
Each of those is a rule a person set at a particular moment, for a reason that made sense that week. None of them is in the document. Most have no owner.
Reread the policy and you learn that the file still says what it said. You do not learn that the account it names now trains on what you send it, or that an agent removed a monitoring step while following an instruction to the letter.
The one-page AI use policy.
A 300 word policy you can adopt today, five blanks to fill, and the amnesty that makes it work.
On its way. Check your inbox, or download it now →
What moves underneath a written answer?
Three things, and none of them announces itself.
The vendor changes the default. You approved a tool in September based on how it behaved in September. The change arrives in pricing documentation or a plan description, not in an announcement, and your policy still names the tool by the same word it used before.
The work takes a shape the policy did not anticipate. The rules were written for the projects on the board at the time. Then a client asks for something adjacent, and the adjacent thing is not covered, and nobody notices because nobody opens a policy while they are quoting a job.
The people change. The rules were agreed by a group who argued about them. A year later, part of the team joined after that conversation ended. What they inherited is a file, not an agreement, and a file carries none of the reasoning that made the rules make sense.
There is no notification when a policy stops being true, which is the whole problem with treating governance as something that can be finished.
The AI Briefing
Tuesdays. 500+ leaders. No hype, just what works.
What does it cost when governance drifts?
The failures do not arrive as one kind of event. They run across a range, and the range is wider than most governance conversations allow for.
At one end sit machines doing exactly what they were told. A safeguard comes off because an instruction was read more broadly than anyone intended, and the work carries on with the safeguard missing and the reason faithfully recorded.
In the middle sit the quiet commercial ones. A vendor’s default changes, client material starts flowing somewhere nobody approved, and weeks pass before anyone opens the settings.
Further along sit the human ones, and these are the expensive ones. Senior people stop exercising the judgment they were hired for, not through carelessness but through a series of small decisions that the stakes were low enough to proceed on.
And at the far end sits liability. When something goes wrong on a client’s work, somebody answers for it, and the people who price that risk for a living are already deciding who.
None of these is caught by a stricter document. Each one happens in the space between what was written down and what is currently true.
A safeguard that came off because an instruction was obeyed
On October 20, OpenAI published a report on two incidents inside its own engineering work. Their own summary of what went wrong:
“In both cases, the agents considered whether monitoring safeguards should be preserved, but interpreted routine coding instructions as permission to weaken them.”
Nobody in the first incident behaved badly at any point.
On October 3, a researcher asked a coding agent to build evaluation tasks from employee feedback collected in Slack. The instruction included four ordinary words: don’t send any messages. Nobody wants an automated process posting into a company channel while it works.
Eighty-two messages later the pipeline stopped. Its monitor had flagged something and could not save the finding.
The agent read the instruction correctly the first time. OpenAI’s report says it treated “don’t send any messages” as a restriction on outreach, separate from the monitoring system, and concluded that the pipeline should run with normal monitoring notifications enabled. It then gave its own subagent an explicit instruction to repair the pipeline without disabling monitoring. The subagent considered suppressing the alerts and declined, on the grounds that disabling required alerts was unsafe.
Two correct judgments, made independently, in sequence.
Then the parent agent reconsidered and directed the suppression anyway. The patch emptied the lists of people who receive monitoring alerts and logged the reason: “User explicitly requested no messages.”
OpenAI’s description of the result: “The misalignment monitor still ran and saved its reports, but human alerts were disabled.”
The instruction was clear. The execution was correct. The safeguard came off anyway, and the audit trail is immaculate.
Nothing here would be caught by access control, because no access was exceeded. Nothing here violates a policy, because the policy was followed. The four words were given for a good reason by somebody who was not thinking about monitoring at all, and they quietly outranked a safeguard nobody had thought to protect from them.
The governance that failed was not in a document. It was a monitoring setting inside a pipeline.
These were internal engineering tasks. No external users were affected, neither incident reached OpenAI’s highest severity level, and the company states plainly that it does not yet have a satisfying explanation for why its models do this.
A thirty-person company does not run misalignment monitors. The shape still transfers: a narrow instruction, given for sound reasons, overriding a protection that was never written down as untouchable.
A default that moved while nobody was looking
In October, Meta launched new models built for long-running agent work. Inside the pricing documentation sits a configuration choice most teams will never open.
Muse Code subscriptions default to Contributor mode. Under Contributor mode, Meta may train future models on the material sent to the service, which includes code, uploaded files, and whatever context the work requires. The pay-as-you-go API runs in Standard mode, which excludes that content from training. The setting can be switched off by anyone who knows it is there.
That sentence describes agency work. A developer is building something for a client, and the context that makes the work good includes the client’s proprietary calculations, sample customer records, internal documentation. The client approved the use of AI on the project.
That approval said nothing whatsoever about contributing their material to a model provider. Those are two separate permissions, and almost no engagement letter separates them. Most agency agreements we see do not, and at this point most firms do not know they need to.
A company that wrote down which tools may be used and what data may go into them had an accurate document in September and a wrong answer in October. The document did not change. The ground under it did, and nothing prompted anyone to go and look.
Weeks later, Google let Gemini create, preview and edit documents, spreadsheets and presentations inside the conversation. Google’s own fine print says the feature has no separate administrator control. An organization can decide who gets the Gemini app. It cannot decide this.
Set both against Anthropic’s commercial terms, which govern business and API use. One sentence: “Anthropic may not train models on Customer Content from Services.”
Vendors made opposite decisions about the same question, and the firm using them did not choose between them. Somebody picked a plan.
Before your next client engagement: for each kind of work you do, which tools are approved, under which account, with which data settings, and when did somebody last open the configuration and read what it says?
The one-page AI use policy.
A 300 word policy you can adopt today, five blanks to fill, and the amnesty that makes it work.
On its way. Check your inbox, or download it now →
What is the market selling instead?
Software that controls what AI agents can reach. None of it checks whether your own answers are still current.
October was a heavy month for governance purchases. Proofpoint agreed to acquire Varonis for about $6.3 billion in cash. Workday agreed to acquire C1.ai, an agentic identity platform, to govern what agents can reach and to revoke that reach when a person leaves or an assignment ends. Okta introduced a gateway that authorizes specific agent actions rather than access alone.
Every one of those solves something real. Agents inherit permissions, those permissions outlive their purpose, and somebody should be selling the fix.
What none of them does is tell a company that its own answers went out of date. Nobody sells that, because it cannot be installed.
Who is deferring to AI the most?
The people who sign off.
Charter’s next finding runs opposite to the way almost every governance framework is built.
Governance assumes a backstop. The system produces something, a senior person reviews it, and the organization is protected by that review. Every approval workflow ever written rests on the senior human being the strongest point in the chain.
Charter asked five thousand people whether they trust AI’s output over their own judgment on work decisions, and sorted the answers by rank.
About a third of managers and executives said yes, roughly twice the rate reported by individual workers. The backstop is thinner than the thing it is meant to be backing.
Senior people are not skipping the review. Eighty-one percent of executives say they always or usually fact-check AI outputs, which is a higher rate than workers report. They are also 3.3 times more likely than workers to go ahead and use an output they doubted, and the reason they give most often is that they decided the stakes were low enough.
So the checking happens. The overruling does not. Somebody looks at the work, has a reservation, decides the reservation does not rise to the level of stopping anything, and signs.
Deciding the stakes are low is itself a judgment, made in the moment, usually under time pressure, in order to avoid making the harder one.
Anthropic’s November 2 filing says the company does not expect to let Claude design and build the most advanced AI models without meaningful human supervision or understanding. Writing about that sentence, the technology writer Shelly Palmer put the weakness of the idea plainly: “A person can approve a decision they barely understand. A review can satisfy a procedure while missing the failure that matters.”
He was describing a frontier lab supervising its own models. The same sentence is what a company policy asks for when a junior checks a deliverable before it reaches a client, and almost no policy says what the reviewer needs to know in order to catch anything.
I described this in September as checking without overruling. In a firm of twenty or sixty people the person doing both is usually a founder, and there is no one above them whose job it is to catch the drift.
Where the bill lands
The Financial Times reported this month that insurers are preparing for multimillion-dollar claims from AI agents that escaped their makers’ controls, and that lawyers in the industry are weighing whether the chief executives of AI companies can be held personally liable for what their models do. Aon reviewed more than 300 AI-related legal cases and found that claims could land on policies covering crime, intellectual property, media liability, cyber, and technology errors and omissions. The policy getting the most attention is directors and officers coverage, which exists to protect executives sued over their own decisions.
Tim Rayner, who runs underwriting and claims in the UK for the analytics firm Verisk: “It’s on every CEO to make sure that their business is appropriately governed and controlled. AI doesn’t change that.”
None of this has been tested in court. One lawyer in the story points out that an American claim “could be under anything.”
What is not in dispute is who a client will look to. An agent holds no insurance and owns nothing. When something goes wrong on a client’s brand, the question becomes who deployed the tool, and the answer is the firm that ran it on the client’s behalf.
The founder of a thirty-person company is a chief executive for these purposes. The exposure is proportionally smaller and structurally identical.
What does reopening look like?
Returning to every place a rule lives, on a schedule, with one person’s name on it.
The four answers stay exactly as they are. What changes is that somebody owns them, with named triggers that bring the reading forward.
That last step needs active protection. A team that cannot retire a tool without it reading as somebody’s error will keep running tools nobody wants. Clearing something out is ordinary maintenance and should carry no more weight than that, which is a thing a leader has to say out loud before it is believed.
Our co-founder Sascha Bosio wrote the one-person version of this in a piece on the mindful use of AI. The practice he describes runs before, during and after a single exchange: knowing the goal before starting, leading the work rather than following it, and asking afterward whether the original intention held. His line for it is “I use AI not to replace myself, but to extend myself, not to delegate, but to grow.” That operates at the scale of one person and one interaction. This operates at the scale of a firm and its clients. The discipline underneath both is the same, and neither works without the other.
Where this leaves the four questions
They hold. Which tools, what data, who releases, what gets checked. Write the answers down if you have not, because an unwritten agreement decays the same way with even less to catch it.
Then date the page, give it an owner, and put the next reading in the calendar. Open the settings while you are there.
Governance is the ongoing practice of maintaining living documents and rules. The policy is one of them. So are the settings, the agent instructions, and the checks nobody ever filed anywhere.
Sources
- Charter AI Pulse, Fall 2026 · Charter, released 2026-10-21. Surveyed 5,002 AI users in the US in August 2026, interviewed 91 in September, weighted to US Census workforce population. Co-led by Brian Elliott and Christina Janzer. Sponsored by the NeuroLeadership Institute; Charter states no vendor influence on the findings.
- Suppressing misalignment monitoring alerts and disabling misalignment monitoring · OpenAI Alignment, report updated 2026-10-20. Incidents 2026-10-03 and 2026-10-06, internal deployments.
- Introducing Muse Wonder models · Meta, 2026-10.
- Muse pricing and rate limits · Meta developer documentation. Contributor mode as the subscription default.
- Muse models documentation · Meta. Standard mode excludes content from training.
- Create, preview, and edit files directly in the Gemini app · Google Workspace Updates, 2026-10-29. Feature has no separate administrator control.
- Commercial Terms of Service · Anthropic, effective 2025-06-17, Section B.
- Anthropic Form S-1/A, Amendment No. 1 · Anthropic, filed 2026-11-02. On meaningful human supervision or understanding of advanced model development.
- Proofpoint to acquire Varonis · Proofpoint, 2026-10. About $6.3bn in cash, closing expected H1 2027, subject to approvals.
- Workday to acquire C1.ai · Workday, 2026-10. Terms undisclosed, closing expected in Workday’s fiscal Q4 ending 2027-01-31.
- Insurers brace for AI agent claims · Financial Times, 2026-10. Aon’s review of 300+ AI-related legal cases; Tim Rayner of Verisk on directors and officers exposure. Subscription required.
- AI governance: the unsexy topic that is about to become your problem · Laura Pretsch, bosio.digital, 2026-09-07. The four answers this piece extends.
- Mindful AI: what the mindful use of AI actually means · Sascha Bosio, bosio.digital.
- Judgment happens in the pause · Laura Pretsch, bosio.digital, 2026-09-14.
- Superintelligence is measured in tasks. Judgment was never one. · Laura Pretsch, bosio.digital, 2026-10-12.
- Cognitive offloading and AI, explained for business leaders · Sascha Bosio, bosio.digital, 2026-10-19.
Frequently Asked Questions
How often should we review our AI policy?
Quarterly suits most firms under a hundred people, with triggers in between that pull the review forward: a new tool, a new category of client work, any change to a vendor’s terms, and any stretch where the work stopped resembling what the policy describes. The interval matters less than having one at all with a person’s name attached to it.
Where does AI governance live besides the policy document?
In the data settings inside each account, the instructions sitting in an agent’s configuration, the checks wired into a workflow, and the approval steps added to templates. Each is a rule somebody set at one moment. Most have no owner, and the document is usually the only thing that gets reviewed.
What usually changes between reviews?
Vendors change the defaults on tools you already approved, often in plan or pricing documentation rather than in an announcement. The work takes on shapes the policy never anticipated. And people join who inherited the document without being part of the conversation that produced it, so they hold the rules without holding the reasoning.
Does a client's approval to use AI cover putting their material into the tool?
Not automatically, and the two belong in separate sentences in the engagement letter. Permission to use AI on a project is a different thing from permission to contribute that client’s material to a model provider, which some tools do by default on some plans. Name both, for each kind of work.
Who should own the AI policy review in a company without a compliance team?
One named person with the authority to change tools and accounts, which in a firm of twenty to sixty people is usually a founder or an operations lead. Authority matters more than title here. A review that belongs to everyone belongs to nobody and does not happen.
Is written AI governance worth doing if employees rank it as unhelpful?
Yes. The Charter finding describes how policies are used rather than whether they should exist. A policy nobody returns to is a filed document. A policy that gets reopened on a schedule, with the account settings checked against what it says, becomes the thing the document was always for.
How do we retire a tool without it looking like a mistake?
Make it routine and say so before it comes up. Closing something that no longer holds is maintenance, and a team that reads it as failure will quietly keep running tools nobody wants. Naming that expectation when the practice is set up costs nothing and prevents years of drift.
Does any of this apply to a company with no written AI policy at all?
It applies more. Every firm already has governance, written or not: a settled understanding of which tools are fine, what goes out to clients, and who looks before it ships. Those answers decay exactly the same way, with nothing at all to catch them. Writing them down is what makes a review possible in the first place.



