An AI acceptable use policy for small business now has to cover agents that act, not just people who paste. ISACA's May 2026 poll found only 38 percent of organizations had a formal, comprehensive AI policy. Seven parts to write, from approval lines to Massachusetts law.
An AI acceptable use policy for small business is a short written rule set that says who may use AI for what, which data may go to which model, what an AI agent may do on its own and what it is logged doing. It is overdue for most companies: in ISACA's May 2026 poll, only 38 percent of organizations had a formal, comprehensive AI policy.
The same poll found that 90 percent of respondents believe employees are already using AI, and a quarter of organizations have no active policy at all. People use AI whether or not there are rules, which is why I argue for governing shadow AI by enabling it rather than banning it. The gap is wider now than when most small firms wrote their first AI rules, because those rules were written for people typing into a chat window. Agents change the question from "what may I paste" to "what may this software do while nobody is watching."
I wrote how to build an AI policy for your team before you need one as the one-page starter: approved tools, three data tiers, review, disclosure and a prohibited list. If you have nothing on paper, start there. This piece covers the parts that page leaves out and that agents make necessary, written for a Massachusetts business and drawn from a year and a half of running LTFI, our agentic business operating system, in live production.
It should cover seven things: who and what the policy applies to, which actions need a person's yes, which data may go to which model, what gets logged, the company rules every AI follows, spending and pause controls, and what happens when a model or vendor changes. The first three carry most of the risk.
The reason the list grew is simple. A chat assistant waits for a person to copy its answer somewhere. An agent can send the email, update the record or book the meeting itself. ISACA's March 2026 poll found that only 36 percent of respondents said humans approve most AI-generated actions before execution, and 20 percent did not know where responsibility would lie if an AI system caused harm. A policy is where you answer both questions before they come up.
For context on where Central Massachusetts businesses sit: the Census Bureau's Business Trends and Outlook Survey put AI use at 19.8% of US businesses as of May 3, 2026, and "less than 20% of firms with four or fewer employees reported using AI." If those national figures hold here, most firms around Worcester and Westborough are early, and early is the cheapest time to write the rules.
An AI acceptable use policy for small business starts by naming every kind of actor it applies to: employees, contractors, and each AI agent as its own entry with a named human owner. An agent that runs on someone's personal login is invisible in your records.
Write one line per agent: its job in one sentence, its owner, the systems it can reach and how to switch it off. Give each agent its own identity rather than a borrowed one, the idea I explain in what a non-human identity for an AI agent is. If you do not yet know what AI is already running, an inventory comes first, and an AI audit for your business covers how to run one.
Scope people the same way. In the October 23 Kief Studio webinar, Brian Gagne asked the question every small business should put in its policy: does the person doing your social media need your financials? "Probably not." A contractor can be given the files without being given the AI system.
Sort every action an AI can take into three lines: what goes ahead on its own, what waits for a person's yes, and what only a person may ever do. Then name who is allowed to give the yes for each.
Here are the default lines we use in LTFI, which each company then sets for itself. They make a sound starting draft for any business:
| Goes ahead | Waits for a person's yes | Person only |
|---|---|---|
| Briefs and summaries | Sending email | Rules and permissions |
| Research | Posting and publishing | Who approves what |
| Drafts | Messages to clients | Passwords and keys |
| Keeping records current | Spending and scheduling | Deleting and money |
The middle column is the one that protects the business. Brian summed it up in the webinar as "drafts go ahead, anything that leaves waits." The right column matters just as much, because an agent that can change its own permissions or approvers has no real limits. Write in the policy that only a named person, usually the owner or administrator, can change who approves what.
Keep the number of approvals small enough that people actually read them. A yes requested on every click turns into a reflex, so save it for the actions that leave the business or cannot be undone.
Map each class of data to the models allowed to see it. A three-tier label (public, internal, restricted) is not enough on its own once you use more than one AI provider, because the real question is where the data ends up.
A simple routing table does the job. Public material can go to any approved model. Internal business data can go only to approved providers on business terms that do not train on your inputs. Restricted data (customer personal information, health and financial records, employee files, credentials) goes only where the policy names, which may mean a local model on your own hardware or no AI at all. Name the providers in the policy and date the list.
Massachusetts makes this part more than good practice. Under 201 CMR 17.00, "every person that owns or licenses personal information about a resident of the Commonwealth" "shall develop, implement, and maintain a comprehensive information security program" that is written down. The same section requires you to oversee service providers by selecting ones "capable of maintaining appropriate security measures" and "requiring such third-party service providers by contract" to do so. An AI provider that receives resident personal information is a service provider in that sense, so your AI routing table and your written information security program should agree. Massachusetts WISP requirements under 201 CMR 17 covers the program itself.
Data that is a mess will leak through any routing table, which is why data governance is the real prerequisite for AI and belongs before this section in practice.
Log enough to answer four questions about any AI action later: who asked, which agent or model did the work, what it touched or produced, and who approved it. Then set how long you keep those records.
Logs are the part most policies skip and the part that pays back first. IBM's 2026 Cost of a Data Breach report found that among organizations with AI-related breaches, "the vast majority (92%) lacked proper AI access controls", and only 40% of organizations reported using access controls on AI models and data at all. Access control without a log only tells you what should have happened.
Logs also do day-to-day work. Brian described this in the webinar: if you did not like an output, "you've got the logs to go figure out exactly where to tweak it." That is how we tune LTFI's AI staff, and it is how you notice an agent changing behavior over time, the subject of AI agent drift and how to catch it.
Write your standing rules once, in one place, and require every AI tool and agent to work from them. Rules scattered across private prompts drift apart, and nobody can review them.
These are the business rules a good employee learns in the first month: what every press release must include (the example from the webinar), which claims you never make, how you address clients, which sources count as facts. The policy should say where the rules live, who may change them and that changes are dated.
This section is also where Massachusetts consumer protection law comes in. When Attorney General Andrea Campbell issued her AI advisory on April 16, 2024, WBUR reported its core message: "existing state consumer protection, anti-discrimination and data security laws still apply to emerging technologies". A rule such as "no AI-written claim about what our product does goes out unchecked" belongs here. Massachusetts AI regulations for a business covers the law in more depth. This is general information, not legal advice.
Set a spending cap per person and per agent, and name who can stop every agent at once. Both belong in writing, because neither can be improvised during a problem.
ISACA's March 2026 poll found that 56 percent of respondents did not know how quickly they could halt an AI system after a security incident. For a small business the answer should be one action by a named person, practiced once while nothing is wrong. Write who that person is, who backs them up and how to turn agents back on.
Say what happens when the AI changes underneath you, and when a person leaves. Both events are routine, and both quietly break policies that did not plan for them.
Models change on the provider's schedule. Researchers at Stanford and Berkeley found that one major model identified prime numbers with 84% accuracy in March 2023 and 51% in June 2023 on the same questions. Your policy should require a quick recheck of important AI tasks whenever a model or provider changes, and a named person to sign it off.
When an employee or contractor leaves, their AI access, their agents and any keys they created should go on the same offboarding checklist as their email. When you leave a vendor, the policy should say your data, rules and logs come with you in a usable format.
LTFI turns most of the policy above into settings rather than reminders. We built it from scratch at Kief Studio as one system, and we run it for each company on its own dedicated server, with no shared hosting.
That is the gap between a governed system and the lookalikes now appearing, many of which are a chat window or a wrapper around a free open-source agent framework. A wrapper can show a policy document. It rarely enforces approval lines, scoped seats, written rules and a log in the same place, and that combination is what makes a policy real.
If you would like to see those settings working, the webinar replay shows them live, and ltfi.ai has a walkthrough. Either way, write the policy first. A team that has argued out its approval lines on paper will get more from any AI system it chooses. We have worked remotely with clients around the world for more than 13 years, so LTFI can be set up wherever you are, and for businesses in the Worcester and Shrewsbury area we are glad to meet in person.
It is a short written policy that says which AI tools staff may use, which data may go to which model, which AI actions need a person's approval, and what gets logged. With agents in use, it also names each agent, its owner and how to switch it off.
Two or three pages is enough for most firms: the seven parts above, each a short paragraph or table. If people cannot read it in one sitting, they will not follow it.
Yes, but they should point at each other. The written information security program required by 201 CMR 17.00 covers protecting personal information, including oversight of service providers. The AI policy adds which models may see that information and what agents may do with it.
Name a person for each kind of action that leaves the business, such as email, posting, payments and scheduling. Changing approvers, permissions, keys and money should stay with the owner or a named administrator only.
Review it every quarter, and also whenever you add an AI provider, change a model or give an agent a new job. Date every change so you can tell which version was in force.
Massachusetts AI regulations are existing laws, not one AI statute. The Attorney General's April 16, 2024 advisory applies Chapter 93A, the anti-discrimination law and 201 CMR 17.00 to AI, and a 2025 settlement over AI lending models reached $2.5 million. What each law means for an AI agent.
Only 17% of U.S. firms with fewer than 20 employees used AI in spring 2026. An owner's guide to AI agents for small business: a readiness check, the first jobs and the never list, who owns each agent, day-one governance and the first 90 days.
AI agent drift is the slow change in what an agent does after launch. Researchers found one model's prime number accuracy fell from 84% in March 2023 to 51% in June 2023 on the same questions. The four causes, five checks that catch it, and how we handle it in LTFI.
Work With Us
Kief Studio builds, protects, automates, and supports full-stack systems for businesses up to $50M ARR.
Newsletter
Strategy, psychology, AI adoption, and the patterns that actually compound. No spam, easy to leave.
Subscribe
