Before you write an AI policy, find out what your tools can already see.
You do not need a 40-page AI policy to start. You need a clear record of what your tools can reach, what they may use, and who decides when something changes.
Sunnyworks ·
Last month your scheduling tool added an AI assistant that reads meeting notes to suggest follow-ups. You didn't turn it on. It arrived on by default, the way most AI features do, and unless someone specifically went looking, it's still reading them.
That's not a hypothetical. It's the ordinary way AI enters a small business, not as a decision, but as a setting nobody unchecked.
You do not need a 40-page AI policy to catch up to it. You need a visible answer to four questions:
Which AI features are turned on?
What information can each feature reach?
What is allowed to go there?
Who decides when something changes?
That is the beginning of an operating rule, small enough to use, specific enough to keep a business from re-deciding the same question from scratch every time a tool adds a transcription feature or a new way to share data.
This is operational guidance for founders who want to think clearly about their own tools. It is not legal, compliance, privacy, cybersecurity, or regulatory advice, and it does not replace a conversation with someone qualified to give that advice for your specific business.
Why AI enters through a tool you already use, not a new project
The first AI decision in a small business is rarely a formal project. It's a setting inside a tool you already pay for.
A meeting tool begins producing transcripts. An inbox offers to draft replies. A design application suggests images from what you've uploaded. A project system summarizes client notes. The feature arrives as part of an update, and the business inherits a decision about information before anyone writes the decision down.
That inheritance is already widespread. A vendor-sponsored survey of 402 U.S. small-business leaders, fielded in mid-2026, found 61% were already actively using AI, with the rest experimenting, and the top reason for holding back was not disinterest but a lack of internal expertise, followed by unclear cost or ROI and security or privacy concerns. Treat the exact numbers as directional, since the survey was commissioned by a vendor with AI tools to sell, but the shape of it matches what shows up inside most small businesses: the tools are already in use, and the thinking about them is still catching up.
The answer is not to turn every feature off. It's to know what's on, what it can reach, and what rule governs it.
What framework should a small business actually use for AI governance?
You don't need to invent one. NIST's voluntary AI Risk Management Framework organizes AI risk work around four functions: Govern, Map, Measure, and Manage. The four questions above are the plain-language version. Here's how a founder-led business translates the framework into four practical moves.
- 1. Govern: name the decision owner
Who is allowed to approve a new AI tool or feature? Who can make an exception? When will the rule be reviewed?
For a small business, the answer may simply be the founder. That's still useful. A named owner is better than an invisible assumption, and it gives a future employee or contractor somewhere to take a question.
- 2. Map: list the tool, the feature, and the information, not just the tool name
Do not inventory only the tool name. Record the specific AI feature and the information it can access.
For each feature, write down:
What it does
- Which files, messages, records, or meetings it can reach
- Whether it creates, summarizes, classifies, or sends information
- Where the output goes next
- What the vendor says about retention, access, and use of inputs
The Federal Trade Commission's small-business guidance starts with the same basic discipline for sensitive information: take stock of what you have, where it is, who can access it, and how it moves.
- 3. Measure: test the boundary before you trust it
A setting that looks safe is not the same as a workflow that's been checked.
Run a small test with non-sensitive material. Ask:
Did the feature use the information I expected?
Did the output include anything it shouldn't have included?
Can another person see the result?
Is the result accurate enough for the job?
What human review is required before anyone acts on it?
Keep the test result with the decision record. The goal isn't to build a lab. It's to stop relying on an assumption you never checked.
- 4. Manage: write the rule and the exception path
Your one-page rule can be simple. Here's what it looks like filled in with three real kinds of features:
- Tool and feature
- Information reached
- Allowed information tier
- Human checkpoint
- Owner
- Last reviewed
- Meeting transcription
- Internal planning calls
- Internal, not restricted
- Founder reviews summary before sharing
- Founder
- 2026-08-01
- Inbox reply drafts
Client email threads
- Restricted — drafts only, never auto-sent
- Founder or team lead reads and edits before sending
- Founder
- 2026-08-01
- Design tool image suggestions
- Uploaded brand assets, no client data
Open
- None required before internal use
- Founder
- 2026-08-01
Use your own categories, but make them concrete. One workable starting point:
- Open: material already intended for public use
- Internal: your operational information, pricing, drafts, and planning
- Restricted: client information, personal information, credentials, confidential agreements, and anything you are not free to hand to a third party
These are a starting point, not a universal legal classification. The right boundary depends on the information, the tool, your contracts, and the obligations that apply to your business.
The question a client will eventually ask you
Could you answer this plainly, today?
“Was any of my information processed by an AI feature?”
A useful answer isn't “we take privacy seriously.” It names the tool, the feature, the kind of information involved, the human review, and what happens next.
If you can't answer that yet, it's not necessarily a policy failure, it's a visibility problem, and it's exactly what the inventory above is for. Clients, partners, and insurers are starting to ask this question as a matter of course, not an edge case. Better to have the answer ready than to be building it live on a call.
The rule isn't the point. The record is.
AI governance doesn't have to mean a compliance program. For a founder-led business, it can start as one current page: a decision owner, a tool inventory, a few information categories, a human checkpoint, and a review date.
That page gets more valuable as the business grows. Someone else can follow it. A new tool can be checked against it. A client question can be answered from it. And when the founder is out of the business for a week, or permanently, the rule doesn't disappear with her memory, which is the same reason this page belongs next to every other operating record a business keeps, not off in a separate “AI policy” no one opens.
The goal isn't less AI. The goal is fewer invisible decisions.
Sources
NIST AI Risk Management Framework — a voluntary framework, organized around Govern, Map, Measure, Manage.
NIST AI RMF Core
FTC: Protecting Personal Information — A Guide for Business — supports the “take stock” inventory principle.
Pax8 SMB AI Pulse, July 13, 2026 — vendor-sponsored survey (n=402 U.S. small-business leaders, ±4.9pp margin of error); cited as a directional signal only, not an authoritative figure.
- Where to go next