What should you automate first in a small business?

Find the repeated interruption

For a few days, write down the work that keeps interrupting your team. Look for the same customer question, details copied from one screen into another, or a follow-up that depends on someone remembering it. You are looking for a repeatable process, not the most frustrating one-off problem.

Then describe the task without naming any technology. For example: when a quote request arrives, collect the contact details, put them in the customer system, and assign a person to respond. That description is much more useful than deciding you need an AI agent.

Compare the value with the cost of getting it wrong

Some tasks take a lot of time but are easy to check. Others take seconds but carry consequences if they go wrong. Drafting a record from an order document is a different commitment from automatically approving an order.

Write down how often the task happens, how long it takes, what an error would cost, and who would notice. A simple routing rule may solve the problem. AI is worth considering when the input varies enough to need interpretation, such as a customer's description of a job.

• How many times does this happen in a typical week?
• What is the next action, and who owns it?
• Can we check the result before it affects a customer?
• What happens if the connection or the software is unavailable?

Count the whole job, including the checking

Consider an illustrative process with 40 requests a week that each take five minutes to retype. That is 200 minutes of entry work. If the new process needs one minute of review per request, the potential reduction is 160 minutes—not the full 200. Setup, exceptions, and ongoing maintenance also take time.

Use your own numbers and compare them with the project cost and recurring software charges. Time recovered is useful, but it is not automatically cash saved. It becomes more valuable when the team uses it for paid work, faster responses, or avoiding an additional expense.

Keep the first version small enough to prove

Pick one source, one destination, and one owner for the first release. For example, handle one kind of website inquiry and put it in one inbox. Test incomplete requests, duplicates, and failures before connecting every form and channel in the business.

Agree on what success means before building. It might be fewer minutes spent entering records, fewer unassigned inquiries, or less scheduling back-and-forth. Compare the result with the old process over a representative period, then decide whether to expand.

Bring a real example to the first conversation

You do not need a specification. Bring a description of the task, the tools involved, and a made-up or safely redacted example of the information that moves between them. Leave out passwords and private customer details.

The useful starting question is: what would have to happen differently for this to be worth paying for? If we can answer that clearly, we can discuss a project that solves the problem instead of just adding another tool.