Most writing about AI is about what it can do. This post is about the opposite: the work we recommend small businesses keep with people, at least for now.
That isn't caution for its own sake. Knowing where not to use AI is what makes it safe to use everywhere else. A team that trusts the boundaries is far more willing to hand over the work that sits inside them.
The test we use
Before handing any task to an AI agent, ask three questions:
- If it's wrong, who notices, and how quickly? If a mistake would reach a customer, a regulator, or your bank account before anyone sees it, the task isn't ready.
- Can a person easily check the result? AI is at its best when its work is quick to review. If checking takes as long as doing, the gain disappears.
- Would you be comfortable explaining the decision to the person it affects? "Our system decided" is not an answer most customers or employees will accept.
If any answer is uncomfortable, keep a person in charge. Here is where that usually lands.
1. Final decisions about people
Hiring, firing, pay, performance reviews, and disciplinary matters should stay with people. AI can help organize information (summarizing applications against criteria you set, for example), but the decision and the conversation belong to a manager.
These decisions carry legal obligations, and they shape how your team feels about working for you. Neither is something to delegate.
2. Anything that commits money
Approving payments, issuing refunds, changing prices, or agreeing terms with a supplier should require a person's sign-off. An agent can prepare the paperwork, match an invoice to a purchase order, or flag something that looks unusual. It shouldn't press the button.
The rule is simple: AI drafts, people approve.
3. Advice your clients rely on
If your business gives professional judgment (tax positions, financial guidance, clinical information, legal interpretation), that judgment should come from a qualified person.
AI can help a professional work faster: finding the relevant document, summarizing a file, drafting a first version of a letter. But the client is paying for someone accountable to stand behind the answer.
4. Difficult conversations
Complaints, bad news, apologies, and anything emotionally charged deserve a human voice. An agent can route the message to the right person quickly and pull together the background they'll need. A person should write the reply.
Customers are remarkably forgiving of a slow, sincere response. They are not forgiving of a fast, tone-deaf one.
5. Sensitive information you haven't mapped
Before AI touches health records, financial details, or personal data, you need to know exactly where that information will go, who can see it, and how long it's kept. If you can't answer those questions yet, the task waits.
This is especially important in healthcare and finance, where professional and legal obligations apply. Check them with your advisers before any AI system sees regulated information.
6. Processes you haven't written down
This one surprises people. If a process lives only in someone's head, AI can't do it reliably, and you won't be able to tell whether it's doing it well.
Write the process down first. Often that step alone reveals the simple improvements, and makes it clear which parts genuinely need AI.
What AI is good for right now
None of this means waiting. There is plenty of valuable work that passes all three tests:
- Sorting and routing incoming messages and documents
- Drafting first versions of routine replies for a person to review
- Summarizing long documents, threads, or meeting notes
- Pulling information from forms into your existing systems
- Answering staff questions from your own approved procedures
- Preparing reports from data you already have
These tasks are repetitive, easy to check, and low-risk if something goes wrong. They're also where most of the time savings are.
"Yet" is the important word
The boundary isn't permanent. As an agent proves itself on a task and your team gets comfortable reviewing its work, some actions can move from "person decides" to "agent acts, person reviews." That move should be a deliberate decision, made one task at a time, not something that happens by drift.
If you'd like help deciding where your own boundaries should sit, that's exactly what we cover in a discovery call. You can also read how we build AI workforces with approval points designed in from the start.
