Most AI governance advice is written for organisations with a risk committee. It arrives as frameworks, registers and multi-stage approvals, and it’s largely irrelevant if you employ nine people.
That doesn’t mean the underlying concerns are wrong. They’re just wrapped in machinery you don’t have and don’t need.
I spent years in compliance and government ICT work, writing guardrails, templates and solution governance for exactly that kind of machinery. Here’s the short version, which is what a small business actually needs.
The risk that actually bites
Ask people about AI risk and they’ll usually say data breach. That’s a real concern and it’s not the one that goes wrong most often.
The common failure is confidently wrong output, sent without anyone checking.
An AI drafting a customer reply will produce a delivery timeframe that reads exactly like every other delivery timeframe you’ve ever written. It may be invented. If it goes out, you’ve committed to it, and the customer has it in writing.
The same applies to quoted prices, warranty terms, availability and anything else that sounds like a fact. Nothing looks wrong. That’s precisely the problem: there’s no visible signal separating a retrieved fact from a fluent guess.
Everything below follows from that.
The one-page version
Five things. If you write down nothing else, write these.
1. What can go out without a human reading it
For most small businesses the answer starts at “nothing external”. Internal drafts, summaries and notes are lower stakes and can flow freely. Anything reaching a customer gets read first.
You can relax this later for specific categories once a system has proved itself. Start tight, because the cost of starting tight is small and the cost of starting loose is a sent email you can’t retract.
2. What must never be automated at all
Complaints. Anything involving a dispute. Anything legal. Anything about someone’s health, money or personal circumstances where getting it wrong causes real harm.
These aren’t technical limits, they’re a judgement about where your business needs to be human. Make it deliberately rather than discovering it during an incident.
3. What staff may put into AI tools
The decision most often made by default, which means made badly.
Some tools train on what you submit. Some don’t. The setting frequently differs between free and paid tiers of the same product, which is how well-intentioned people end up pasting client details into something that retains them.
Decide what categories are acceptable: customer names, financial details, health information, anything under a confidentiality agreement. Then check your tools’ terms against that list. It’s a twenty-minute job and it’s the highest-value twenty minutes in this whole article.
4. Where answers come from
Any AI system giving factual answers should be grounded in your own material and able to show its source.
This isn’t a technical nicety. It’s what lets a person check an answer in seconds rather than reconstructing it from scratch. A system that says “here’s the answer, from this document” can be verified. One that just produces the answer has to be trusted or fully re-checked, and in practice busy people trust it.
5. Who owns it
One name. Whose job is it to notice when the system starts behaving oddly, and who decides when a new use is acceptable?
In a small business this is usually the owner, and that’s fine. What’s not fine is nobody, which is the default when it isn’t written down.
What you can safely skip
Plenty, and it’s worth saying so, because the volume of available advice makes people freeze.
You don’t need a formal risk register, a model inventory, an ethics board, quarterly audit cycles, or a policy document with version control. Those exist to coordinate large numbers of people who can’t all talk to each other. You can all talk to each other.
You also don’t need to understand how the models work. You need to know what your system is allowed to do and how you’d notice if it stopped doing it properly.
Checking it still behaves
Guardrails written once and never revisited are decoration.
The practical version is sampling. Read a handful of the system’s outputs each week at first. Not all of them, just enough to notice a pattern. As it proves itself, drop to monthly.
What you’re looking for is drift: answers that were fine in March quietly getting less accurate because your prices changed, or a policy was updated, or the underlying source material moved on and nobody told the system.
That last one is the most common maintenance failure I see. The automation is working exactly as built. The business changed around it.
Why this matters more than it sounds
The reason to write a page like this isn’t compliance. It’s that it makes the system usable.
People either trust automation completely or not at all, and both are wrong. Explicit boundaries let you trust it exactly as far as it’s earned, which is the only sustainable position. Staff know what to check. You know what’s been decided. Nobody has to guess when something unusual comes up.
It takes about an hour, and it’s the difference between AI you can safely lean on and AI you keep having second thoughts about.
If you’d like a second opinion on where your boundaries should sit, describe what you’re planning to automate and I’ll tell you which parts I’d keep a human on.
Frequently asked questions
Does a small business need an AI policy?
You need something, but it can be a page rather than a document. The point is agreeing what staff may put into AI tools, what has to be checked before it goes out, and who decides when something new comes up. Most of the risk in small businesses comes from having none of that written down at all.
What are AI guardrails?
Practical limits on what an AI system may do without a human. Common ones: never send external email unreviewed, never quote a price without approval, never touch customer records, always cite the source of an answer. They're business rules, not technical settings, though they get enforced technically.
What is the biggest AI risk for a small business?
Confidently wrong output going out unchecked. A made-up delivery date or an invented policy detail reads exactly like a real one, and once it's sent you're committed to it. That risk is much more common in practice than the data-breach scenarios that get discussed more.
Can staff paste customer information into AI tools?
That's the decision to make explicitly rather than by default. Some tools train on what you submit and some don't, and the setting often differs between free and paid tiers. Decide what categories of information are allowed, write it down, and pick tools whose terms match.
How do I check whether an AI system is behaving?
Sample its output regularly rather than trusting it wholesale. Read a handful of drafts each week at first, then less often as it proves itself. Log what it does so there's a record to check, and make sure someone owns that job by name rather than in principle.
Wondering what this would look like in your business? A short chat is usually enough to tell.
Let’s chat