Skip to content

AI agents · August 2026

AI agents in a small business: where to start

Published in August 2026

The question is not whether AI works in a small business. It does. The question is which of your processes to touch first, because choosing badly is the fastest way to burn both budget and internal credibility at once.

The first process has three characteristics

  • It is frequent. If it happens twice a month, the saving does not justify the effort of automating it.
  • It is describable. If nobody on the team can explain the rules it follows, an agent will not guess them either.
  • Failure is cheap. To start with, avoid processes where a mistake costs a customer or a fine.

The typical candidate in a small business is processing documents that arrive by email: supplier invoices, delivery notes, reports. High volume, clear rules, and a mistake that gets caught and corrected without serious consequences.

What you need in place first

Less than people usually claim, but not nothing. The information has to live somewhere reachable, even if that is an inbox or a shared folder, and someone inside has to spend a few hours explaining how the work is done today. That internal time is the hidden cost of any automation project, and it is worth counting from the start.

An agent is not a chatbot

A chatbot answers. An agent executes: it reads, queries the system, decides within the limits you set, writes the result where it belongs and notifies whoever needs to know. The practical difference is that an agent needs permissions over your systems, which forces you to decide in advance what it can do alone and what has to pass through a person.

When not to automate

  • When the process is about to change in the coming months: automating something that is about to die is throwing money away.
  • When the volume is low and one person absorbs the work without effort.
  • When nobody internally can explain the rules: sort them out first, automate second.

How to tell whether it worked

Before starting, write down how much time the process takes today and how many errors are caught per month. Without that baseline, any later conversation about return is an opinion. With it, two months in you know whether to continue or stop.

Automate ifIt is frequent, describable and mistakes are cheap to fix
Wait ifThe process is about to change or the volume is low
Tidy up first ifNobody can explain the rules currently followed

Want to know what your first process would be?

We look at it with you in one session. If we do not see a clear candidate, we will say so.