Automation and AI
What Should a Small Business Automate First?
Start with a frequent task that follows clear rules and costs little when it goes wrong. Here's a grid to choose, with the checkpoints to keep.
An enquiry arrives by email, someone copies it into a spreadsheet, and another person writes back to confirm receipt. Nothing dramatic, but it happens every day. The temptation is to automate everything at once, or to hand the whole thing to an AI tool.
Our suggestion is more modest: pick one task to start with, one that repeats often, follows clear rules and costs little if it goes wrong. This article offers a grid to find it, illustrative examples and the checkpoints to plan for. It promises no specific gain: the outcome depends on how your business works.
Start from the real work, not the tool
Before talking about software, describe the task the way a person does it today: what triggers it, what information is used, what decisions are made, what comes out and who is told. Five to ten lines are enough.
This exercise often reveals two things. First, some steps rely on habits nobody has written down. Second, the task has special cases that nobody mentions unprompted: a customer who writes twice, a missing attachment, a request that has nothing to do with your services.
A task you can't describe clearly isn't ready to be automated. Clarifying it is already a useful result.
Four criteria for choosing the first task
We suggest comparing your candidates with four questions.
1. Does the task come up often? A step repeated every day deserves more attention than a once-a-year chore. Frequency is something you observe, not something you estimate from memory.
2. Are the rules clear? If two people on your team do it the same way, a rule can probably describe it. If each does it differently, standardize first.
3. What happens when it goes wrong? A misfiled email is fixed in a minute. A quote sent to the wrong customer or an incorrect invoice leaves a mark. The costlier the mistake, the more the human check matters.
4. Where does the human checkpoint go? A person can review a draft before it's sent, approve an action or handle the exceptions set aside. Decide where before building anything.
Common automation tools do provide this kind of step: an approval request that pauses the workflow until a person responds, as described in Microsoft's documentation on Power Automate approvals (page dated April 29, 2026). It's one category of feature among others; citing it is not a tool recommendation.
A prioritization grid for your candidates
The table below is illustrative. The ratings are examples of reasoning, not findings about your business or measured results. Redo it with your own tasks.
A prioritization grid for your candidates
| Task (example) | Frequency | Clear rules | Risk of error | Human checkpoint | Reading |
|---|---|---|---|---|---|
| Sort incoming enquiries by type | High | Yes, if types are defined | Low | Review of enquiries flagged "to check" | Good candidate to start with |
| Create a follow-up record from a form | High | Yes | Low to medium (duplicates) | Periodic check of records created | Good candidate; plan for duplicates |
| Prepare a draft reply to an enquiry | High | Partly | Medium (tone, inaccurate facts) | Review and approval before sending | Possible, with mandatory approval |
| Remind the team about an overdue task | Medium | Yes | Low | None, or review of reminders sent | Simple, often low risk |
| Send an invoice or a quote | Medium | Varies | High | Approval of each document | Hold off until the rules are stable |
| Reply alone to a complaint or sensitive case | Low | No | High | Handled by a person | Do not automate |
Read the whole table before choosing. A frequent but risky task is better served by a checkpoint than by being left to run alone. A rare, ambiguous task is rarely a good starting point.
Plan for exceptions and human oversight
An automated process works well in the cases it was built for. Trouble comes from the others: an empty field, an invalid address, a message that looks like nothing it has seen.
For each task you select, write down three things:
- What is handled automatically, and under what conditions.
- What is set aside for a person to handle, and how that person is notified.
- What requires approval before any action a customer can see.
The principle: keep a person in the loop wherever a mistake shows from the outside. A personalized reply, a proposal or a price commitment should be reviewed before it goes out, while a standard acknowledgment can be sent without individual review.
If you add AI, for example to summarize an enquiry or suggest a draft, the same precautions apply, plus one more: limit the information passed to the tool. Canada's privacy regulators note that accountability for decisions rests with the organization, not with an automated system (principles of December 7, 2023). AI doesn't fix everything: it suggests, a person decides.
How do you know the first automation works?
Before you start, decide what you want to observe, then compare with where you began. A few possible markers: the number of cases set aside, the errors corrected by hand, the enquiries handled without being missed. Choose the ones that make sense for you; none is universal.
Run a limited trial, for instance on one category of enquiries or over a few weeks, before extending it. If your task involves forms or enquiries from your website, measuring useful actions is a topic of its own, covered in our article on the roles of GA4 and Google Tag Manager.
Frequently asked questions
Should we start with AI?
No. A simple rule or a connection between two tools is often enough. AI becomes relevant when the task involves reading free-form text, and it then calls for human oversight in proportion to the risk.
Do we need to change tools first?
Not necessarily. Look first at what you already use: a CRM, a form, an inbox. If the difficulty lies in how information moves between those tools, our CRM and integrations page describes how we approach that kind of need.
How much time will it save?
We can't say for a business we haven't observed, and a general figure would be misleading. Measure the current situation, then the effect of the trial.
Next step
Pick one task, describe it in about ten lines and fill in the grid above. If you'd like help scoping it, AETHER can review the task with you and define the exceptions, human checks and limitations before anything is set up. See our Automation and AI page.