Yes, you need a written AI policy, and the size of your company is not a defense. Fifty-nine percent of workers at businesses with fewer than ten employees say their employer has no clear AI policy, which means the tools are already in daily use, unsupervised, on your data. Writing the policy does not require a legal department or a compliance budget. It requires one afternoon and the willingness to make five decisions and put them in writing.
Most owners assume governance is something that kicks in once you clear a certain headcount or a certain revenue line. It is not. The exposure exists the moment an employee pastes a customer list into a chatbot to draft a follow-up email, and that moment already happened this week whether you sanctioned it or not.
/ 01The policy already exists, it’s just unwritten
Sixty-six percent of office professionals say they have used an AI tool at work that they believed was not approved, according to a 2026 survey from PagerDuty and Wakefield Research. More than a third admitted doing it with customer data. Your team is not waiting for permission. They are already running the experiment, and the results are sitting in whatever chatbot history nobody thought to check.
This is not a rebellion. It is a rational response to an unclear signal. If leadership has not said what is allowed, employees default to whatever gets the job done fastest, and the tools are good enough now that fastest usually wins. Verizon’s 2026 Data Breach Investigations Report found that unauthorized AI use, what security teams call shadow AI, is now the third most common non-malicious insider action detected inside breached organizations, with detections rising roughly fourfold in a single year. That is not a hypothetical risk. That is a trend line with your company’s name not yet on it.
The cost of skipping this is not only regulatory. MIT’s Project NANDA found that ninety-five percent of enterprise generative AI pilots show no measurable financial return, and the gap between the companies that see returns and the companies that do not traces back to exactly this kind of operational discipline, not model quality. A policy is not a brake pedal. It is the first piece of the operating system that turns experimentation into an actual result.
None of this means AI is dangerous. It means AI without a stated boundary behaves like water. It finds every gap in the floor.
/ 02What regulators actually expect from a company your size
The regulatory story sounds like it is written for enterprises, and mostly it is. The EU AI Act reached full effect in August 2026 and requires AI literacy training for staff at any organization deploying AI systems that touch EU customers, regardless of headcount. Colorado’s AI Act, effective February 2026, is stricter still for anyone building or materially customizing tools that make consequential decisions about employment, housing, or lending. Most small and mid-market companies are not building those systems. They are renting them, which lowers the bar but does not erase it.
What regulators, and more practically what your own insurer and your biggest customer’s procurement team, actually want to see is documentation. Not sophistication. A one page policy that names your approved tools, states what data never goes into them, and names a human who owns the exceptions will satisfy nearly every review a company your size will actually face. The forty page governance framework a consultant tries to sell you is built for a company with a general counsel on staff. You probably do not have one, and you do not need one to do this correctly.
The traffic light system beats the manual
The single best structural choice for a small team is a traffic light system: green, yellow, red. Green covers low risk tasks any employee can do with an approved tool without asking, drafting a blog outline, summarizing a public article, brainstorming subject lines. Yellow requires a second set of eyes, anything touching a real customer or a real number. Red is a hard no: no customer data, no financials, no unreleased strategy, ever, in any tool that trains on what you feed it.
This works because it asks nothing of an employee that they cannot apply in the fifteen seconds before they paste something into a text box. A compliance binder nobody has time to consult during a Tuesday afternoon deadline will fail. A three color system posted next to the espresso machine will not.
A policy nobody wrote down is just a habit waiting to become a headline.
What you can actually do this quarter
None of this requires a hire. It requires an afternoon, a shared document, and a willingness to say a few things out loud that have so far only been implied.
Write the one page version
List the approved tools, the banned data categories, and the name of one person who reviews exceptions. Ship it as a single page, not a manual.
Install the traffic light
Green, yellow, red, posted somewhere people actually look, not buried in an onboarding folder from 2019.
Name one owner, not a committee
Governance dies in committee. Pick the person, tell the company who it is, and give them the authority to say no.
Run one live session, not a memo
Forty-five minutes, real examples from your own industry, questions allowed. A PDF nobody opens teaches nobody anything.
Put a review date on the calendar now
The tools you are writing rules for in August will not be the tools your team is using by spring. Revisit the policy on purpose.
Why the policy still won’t get written
Here is the uncomfortable part. Almost every owner reading this already agrees with everything above. The reason the policy still is not written by Friday is not disagreement. It is that writing it competes with payroll, with the client who is upset, with the hire you are three weeks behind on, and none of those things lose to “AI governance” in a fair fight for your Tuesday.
Strategy work in a small company belongs to whoever has ten free hours, and there is no one on your team who has ten free hours. That is not a character flaw. It is an org chart problem. The policy that would take an outsider an afternoon takes an owner three months of good intentions and a folder called “later.”
The honest fix is not more willpower. It is putting the decision in front of someone who is not already at capacity, for exactly long enough to make the five calls above and write them down.
I’ve written this exact policy for companies smaller than yours. Happy to write it for one more.
Where the policy usually gets written: as the first deliverable of an AI Opportunity Audit, or in month one of a fractional CMO engagement. Most of the companies that ask are in Connecticut, Massachusetts or Rhode Island, and the five calls above are the same in all three.
For a Connecticut insurer or bank, the governance expectations are written down; see the questions insurance and financial-services companies in Hartford actually ask.
The policy is one step on a longer path; Getting AI to run inside a mid-market company lists the Signal’s thirty-six AI-adoption questions in the order they come up, including Connecticut’s new law.