AI tools are moving from drafting things for you to doing things for you: sending emails, updating records, booking, ordering, paying. Tools that take actions on their own are often called AI agents. They can save real time. They can also make a mistake at speed, and some mistakes cannot be undone.
So how much should you let AI do on its own? This guide gives a small business a simple way to decide, and a handful of guardrails that do not need a technical team.
Key takeaways
- Decide by consequence: the harder a mistake is to undo, the more a person must approve it.
- Anything irreversible, such as paying, deleting or messaging customers, needs a named person's sign-off.
- Give AI tools the least access they need, and start them read-only.
- Treat AI output as a draft until someone has checked it.
- Make sure you can switch it off and roll back, and keep your own backups.
How do you decide what AI can do on its own?
Ask one question of every action: if the AI gets this wrong, can we put it right easily?
Not "how clever is the tool?" A very capable tool can still make an unrecoverable error on an input nobody thought to test. One bad payment or one deleted folder can outweigh months of good work.
| Type of action | Examples | Who decides |
|---|---|---|
| Low impact, easy to undo | Sorting emails into folders, tagging records, drafting internal notes | AI can act; spot-check now and then |
| Medium impact, can be undone with effort | Updating customer records, scheduling, drafting replies to customers | AI prepares; a person reviews before it goes out, at least at first |
| High impact or irreversible | Paying or refunding money, changing supplier bank details, deleting data, sending messages to customers, making commitments or discounts | AI may suggest; a named person approves each one |
Start everything one row more cautious than you think it needs. You can relax the rules later, once you have seen how the tool behaves on your real work.
Why a person must approve anything irreversible
Testing only covers what you thought to test. Real life sends in the odd case: the unusual invoice, the angry customer, the email that looks like it is from a supplier but is not. For actions that cannot be reversed, the safe pattern is simple: the AI proposes, a person approves.
One large finance team took exactly this approach after a single wrong adjustment had rippled through its billing and reports. Its AI tools can now only propose adjustments. A separate step checks that the person approving has the authority to do so, allows that one posting only, and keeps a record. A small business does not need anything as elaborate, but the principle is the same: keep the final say outside the AI.
Give AI tools the least access they need
When you connect an AI tool to your email, accounts or customer system, it often asks for broad access. Be stingy.
- Start read-only. Let it see and suggest before it can change anything.
- Give it its own login where the software allows, rather than borrowing a member of staff's. That way you can see what it did, and switch it off without locking anyone out.
- Limit it to the task. A tool that drafts replies to enquiries does not need access to your bank account or your whole file store.
- Take access away when a trial ends or the tool is no longer used.
Make sure "human review" is real
Having a person check the output only helps if they really check it. Under time pressure, people drift from asking "is this right?" to "does this look about right?", and AI is very good at looking right. It writes with confident detail even when it is wrong.
Ways to keep review honest:
- Make it clear what was written by AI, so reviewers know to look closely.
- Keep the amount to review manageable. Nobody reliably spots an error buried in pages of plausible text.
- Ask for a one-line reason when someone approves something important.
- Now and then, pull a few approved items and check them properly.
- Watch the override rate. If people never change anything the AI produces, ask whether that is because it is perfect, or because nobody is looking. Our guide to telling whether AI is paying off explains how to track this.
Customer-facing AI needs an easy way out
If customers deal with an AI chat or email assistant, give them a clear, quick route to a person. Decide when it should hand over: after a couple of failed attempts, straight away for anything urgent or involving money, and whenever the customer asks. Pass on what has already been said, so the customer does not have to repeat themselves. Being stuck in a loop with a bot does more damage to trust than having no bot.
Plan for the day it goes wrong
Something will, eventually. Prepare now.
- Name one person responsible for each AI tool, and for dealing with problems it causes.
- Know how to switch each tool off quickly, and who is allowed to.
- Keep your own backups of anything the AI changes or builds up, not only the supplier's.
- Keep a simple record of what AI tools do and what people approved.
- Write down the actions each tool must never take, before listing what it may do.
- Review access and rules when a tool gains new features or connects to something new.
The backups point is easy to overlook. One small technology firm relied on an AI agent that had built up a useful memory of its work over time. The firm had not kept its own copy. When that memory was lost, so was a large part of what the agent had learned, and there was nothing to restore. Recoverability is the safeguard you control yourself.
Earn trust in stages
The safest way to introduce an AI agent is to let it shadow the existing process first. Have it suggest what it would do while people carry on as normal, and compare its suggestions with what they actually did. When it agrees with your team reliably on the simple cases, let it act on those, with limits. Widen what it may do step by step, as the evidence builds. Nobody should be granting an AI tool full freedom on its first day.
If you want help setting this up, a good adviser will ask what could go wrong and how you would undo it before they talk about what the tool can do. Our guide to bringing your team with you covers the people side of introducing these rules.