US small-business field guide · Checked August 13, 2026
AI governance for a small business should fit on the desk, not fill a department
You do not need a committee to decide whether an employee may use an AI writing assistant. You do need a clear data boundary, an accountable owner, a way to test important output, and a response plan when the tool is wrong. This guide turns those ideas into a workable US small-business process.
Michael Okeje
Small-business AI governance research · Last updated August 13, 2026
The real starting point is discovering shadow AI
Most small businesses do not begin with a formal AI program. A team member tries a chatbot, a marketer uses an image tool, a sales representative pastes a prospect email into a writing assistant, or an accounting platform quietly adds an AI feature. Banning everything after the fact can drive the use underground. Pretending nothing is happening leaves the business unable to see what information is leaving, what claims are being published, or what decisions are being influenced.
I would begin with a friendly inventory question: “Which AI tools or AI features are you using for company work, and what do you put into them?” Ask across email, documents, meetings, marketing, customer service, recruiting, finance, design, and operations. Record the answer without blaming the person. The purpose is to establish a truthful map before creating rules.
The SBA's small-business guidance says to start small and test whether a tool adds value, while also identifying intellectual-property, security, customer-trust, and ethical risks. That balance is the right tone. Governance should help the business use AI deliberately, not make responsible employees afraid to disclose what is already happening.
The four-page governance starter kit
A small business can create a useful first version without buying a governance platform. Keep these records together, review them with the people doing the work, and improve them as the business learns.
Page 1: AI-use register
Tool or feature, vendor, purpose, owner, users, data entered, output used, connected systems, risk tier, approval status, and next review date.
Page 2: employee policy
What may be entered, what needs review, what is prohibited, which tools are approved, how public claims are checked, and how incidents are reported.
Page 3: use-case approval
The intended and non-intended use, affected people, baseline, quality bar, data path, human fallback, vendor terms, and evidence supporting approval.
Page 4: incident log
Date, system, input or action, impact, people affected, containment, correction, notification decision, root cause, and change made to prevent repetition.
Set risk tiers that people can actually use
A single rule for every AI use is either too weak for high-impact work or so burdensome that people ignore it. Use a short tiering system and explain it with examples. The tier belongs to the use case, not only the vendor brand.
Green: assistive
Examples: Brainstorming, public-information summaries, formatting, or internal drafts that a person checks before use.
Minimum control: Approved tool list, no sensitive data, human review, and a simple report route.
Amber: business process
Examples: Customer replies, sales research, internal knowledge, marketing, or document workflows using company information.
Minimum control: Named owner, approved data path, representative tests, access controls, and periodic review.
Red: consequential
Examples: Employment recommendations, financial or insurance decisions, health-related work, legal conclusions, or autonomous actions that change records or send commitments.
Minimum control: Specialist review, stronger evidence, explicit human decision authority, incident plan, and documented approval.
An AI tool can move between tiers when the use changes. A writing assistant used for a public blog draft is different from the same assistant used to summarize a confidential dispute. Record the change rather than relying on memory.
A policy people can follow
The best policy is not the most impressive document. It is the one a busy employee can understand before pasting a file into a tool. Use direct examples and put the strictest rules near the top.
| Rule | What it means in daily work |
|---|---|
| Protect confidential information | Do not enter customer, employee, financial, health, legal, credential, or proprietary information into an unapproved tool. Use redaction and the minimum necessary context. |
| Keep people accountable | AI may draft, classify, summarize, or suggest. A named person remains responsible for public communications, commitments, decisions, and records. |
| Use approved tools | The owner maintains a short list of approved tools and records what each is permitted to handle. Free trials are not automatically approved for business data. |
| Verify material output | Check numbers, names, dates, citations, claims, legal language, customer facts, and policy requirements against the original source before sending or publishing. |
| Disclose when appropriate | Do not imply that a human personally wrote or reviewed something when the context or law requires a different disclosure. Keep a customer-trust standard, not only a legal minimum. |
| Report and learn | Report privacy exposure, harmful output, fabricated facts, unauthorized action, security concern, or repeated quality failure. Retire or change the workflow when evidence says it is unsafe or unhelpful. |
The SBA specifically highlights protecting sensitive information, checking intellectual-property concerns, reviewing AI content, and maintaining customer trust. The FTC's business guidance adds a practical vendor lesson: put security expectations in writing, control access, verify compliance, and update controls as threats change.
What counts as an AI incident?
Do not wait for a dramatic breach before creating an incident route. An incident can be a confidential document entered into an unapproved tool, an AI-generated claim published without verification, a candidate or customer treated unfairly because of an automated recommendation, an agent taking an unauthorized action, a fabricated number in a financial document, or a vendor changing a data practice without the owner noticing.
The first response is containment. Stop the workflow or revoke access if harm can continue. Preserve the relevant evidence without spreading the sensitive information further. Tell the owner what happened, who may be affected, and what has already been corrected. Seek legal, security, privacy, insurance, or domain advice when the situation requires it. Do not make a public promise about an investigation before the facts are known.
After the immediate response, classify the cause. Was the policy unclear? Was the tool unapproved? Did a user misunderstand the output? Did the vendor retain data differently than expected? Did the system fail a known test? The answer should change a control, a training example, a vendor decision, or the workflow itself.
A 30-day rollout for a five-person business
- Days 1-5: ask every team member which AI tools and features they use, then record the use cases without assigning blame.
- Days 6-10: classify each use as assistive, business-process, or consequential. Pause only the uses that create immediate, material risk while you review them.
- Days 11-15: publish the six-rule policy with examples of prohibited data, required review, approved tools, public claims, and reporting.
- Days 16-20: appoint one accountable leader and one owner for each active workflow. Write the intended use, non-purpose, data path, and fallback.
- Days 21-25: test the important workflows on ordinary and difficult cases. Record the result, review burden, and failure conditions.
- Days 26-30: approve, restrict, pilot, redesign, or stop each use. Set a date or trigger for review when the vendor, model, data, users, or decision changes.
For the broader framework, read our AI governance guide. For the test design, use AI evaluation; for a live experiment, use the 30-day AI pilot plan; and before choosing a vendor, complete the vendor evaluation checklist.
A small-business governance scorecard
The aim is not perfect prediction. It is a business that can explain how it uses AI, protect the information entrusted to it, catch mistakes, and improve its process as the technology changes.
Frequently asked questions
What is AI governance for a small business?
It is the practical system a small business uses to decide which AI uses are permitted, what data may be entered, who is accountable, how outputs are checked, how vendors are reviewed, and what happens when something goes wrong. It can begin with a short policy, an AI-use register, named owners, a few risk tiers, and an incident route rather than a large committee.
Does a small business need an AI policy?
A short policy is useful whenever employees or contractors use AI with company work, customer information, employee information, intellectual property, or public-facing communications. The policy should be specific enough to change behavior: explain what may not be entered, what requires review, what cannot be delegated, which tools are approved, and how a concern is reported.
What information should employees not put into public AI tools?
Unless an approved product and agreement specifically permit it, avoid entering passwords, payment information, Social Security numbers, health information, private customer records, confidential contracts, trade secrets, unreleased financial information, and identifying employee or candidate data. Use the minimum necessary information and ask the owner or adviser when the classification is unclear.
Who is responsible for AI in a small business?
The owner or an accountable executive should own the overall program. Each use case should have a business owner who is responsible for the outcome, and a technical or vendor owner where implementation needs one. Bring in a lawyer, accountant, security specialist, or domain expert when the use affects regulated data, employment, finance, health, safety, or important customer decisions.
How should a small business respond to an AI mistake?
Stop or limit the workflow when the mistake could cause continuing harm, preserve the relevant prompt, input, output, action, and version information, notify the owner, correct the customer or operational record, assess who was affected, and decide whether a vendor, regulator, insurer, or professional adviser must be contacted. Turn the confirmed failure into a new test or policy rule.
Can a small business use the NIST AI RMF?
Yes. NIST's AI Risk Management Framework is voluntary and can be scaled down. Its Govern, Map, Measure, and Manage functions are useful prompts for a small business: assign accountability, describe the use and affected people, test the system, and monitor and improve it. Do not copy enterprise paperwork without adapting it to the actual risk.