Most advice about AI in a small company is about where to start. Just as useful is a short list of where not to start, or where not to go at all. A tool that drafts a friendly reply in seconds can also send a wrong price, a wrong date or a private detail to the wrong person just as quickly.
This article is for an owner or operations lead who is already hearing "we could automate that" and wants a quick, consistent way to answer. It does not need technical knowledge. It needs about ten minutes per task and an honest answer to five questions.
Why a check helps
Without a check, decisions follow enthusiasm. A task gets automated because it is annoying, or because a colleague saw a demo, not because it is safe. A check does two things. It makes the reasoning visible, so a "no" is not taken personally. And it moves attention from "can AI do this?", which is nearly always "sort of", to "what happens when it is wrong?", which is the question that matters.
If you have not yet chosen a first use case, start with choosing a first AI use case and then apply this check to your shortlist.
The five questions
Take one task at a time and answer each question with yes, no or unsure. Treat "unsure" as a "no" until you find out.
1. What does a mistake cost?
Think of a realistic bad day, not the worst case. If the output is wrong once in twenty times, what is the damage? A slightly clumsy internal summary costs a minute. A wrong quote costs money and trust. A wrong instruction to a customer may cost both.
If the cost of one mistake is higher than the time saved by a hundred correct runs, keep the task with people, or let AI draft and have a person decide.
2. What data does it touch?
List the information the task needs: names, contracts, prices, health details, passwords, employee records. Then ask whether you are comfortable with that information leaving your systems to be processed by a tool, and whether the supplier's terms say what happens to it.
Personal data about customers or staff, and anything covered by a confidentiality promise, needs a clear yes before it goes near an AI tool. If you are not sure what your staff already paste into public tools, read about shadow AI first.
3. Does it carry legal or contractual weight?
Some outputs are more than text. A contract clause, a dismissal letter, a tax figure, a regulatory filing or a safety instruction creates obligations. AI may help a person prepare these, for example by summarising a long document, but the final wording should be written or checked by someone qualified.
4. Can it be undone?
A draft that sits in a folder can be deleted. An email that has been sent, a payment that has been released, a record that has been overwritten or a message posted publicly cannot be pulled back so easily. The harder it is to reverse, the more reason to put a human step before the action. The article on choosing the first AI agent process goes deeper into this point.
5. Who answers for it?
Name one person who is responsible for the result of this task. If nobody can be named, or the answer is "the AI", the task is not ready. Responsibility does not move to a tool. The named person should also have the time and knowledge to check the output, otherwise the check is a formality.
Reading the answers
You do not need a score. Use a simple rule:
- Five clear yes answers: a good candidate. Start small, with a person reviewing early results.
- One or two "no" or "unsure": limit AI to a draft or a suggestion that a person approves before anything leaves the company. See human approval and audit trail for how to record that approval.
- Three or more "no", or a "no" on legal weight or sensitive data: keep the task with people for now.
Examples
These are illustrations, not findings from any company.
| Task | Likely result | Why |
|---|---|---|
| Summarising a long internal meeting note | Good candidate | Low cost of error, no outside recipient, easy to discard |
| Drafting a reply to a routine customer question | Draft with approval | A person reads it before it is sent |
| Setting a price for a custom order | Keep with people | Money and trust are involved, and it is hard to undo |
| Writing an employee warning letter | Keep with people | Personal data and legal weight |
| Deciding which supplier invoice to pay | Keep with people, or approval only | Payment cannot easily be reversed |
Put it on one page
Once you have gone through five or ten tasks, write the outcome on a single page with three headings: allowed, allowed with approval and not for AI. Add the date and the name of the person who maintains it. Share it with staff together with your usage policy for AI tools. People are more likely to follow a boundary they can see.
Review it
Tools improve and your processes change. Look at the list every quarter and when something goes wrong. A task in "not for AI" can move once you have better controls, better data handling or a lower-risk way to use the tool. A task in "allowed" can move the other way after a near miss. That is the check working, not failing.
Common mistakes
- Judging a task by how often AI gets it right rather than by what the failures cost
- Treating "unsure" as "probably fine"
- Letting the person who built the automation also decide whether it is safe
- Naming no owner for the result
- Writing the list once and never reviewing it
Checklist
- The five questions answered for each candidate task
- "Unsure" treated as "no" until resolved
- A named person responsible for each automated task
- A human approval step wherever an action cannot be undone
- A one-page list of allowed, allowed with approval and not for AI
- A quarterly review date
Sources
- Practical guidance based on common risk-management practice for small businesses, offered as a decision aid. The questions and thresholds are starting points to adapt to your own company. This is not legal advice; check requirements for personal data and regulated activity with a qualified adviser. No client results or market figures are claimed in this article.