AI Agents for Small Business: Past the Chatbot

The distinction is simple enough. A chatbot responds to a question. An agent is given access to your tools and carries out a sequence of steps: read the enquiry, look up the customer, draft the quote, update the CRM, schedule the follow-up. It is the difference between something that talks and something that does.

That difference is also exactly why the interesting question is not what an agent can do. It is what you should let it do without checking with you.

Tasks worth handing over

• Moving data between systems. The copy-paste between your form, your CRM and your accounting tool. Nobody enjoys it and everybody eventually mistypes something.
• First-pass enquiry handling. Enrich the lead, score it, draft a reply for a person to send.
• Follow-up that would otherwise be forgotten. The second and third chase that nobody gets to on a Friday afternoon.
• Drafting from templates. Quotes, proposals, standard documents, prepared from your own data and waiting for approval.
• Routine reporting. Pulling the same numbers into the same summary every Monday.

Tasks that stay with a person

Anything that spends money. Anything that goes to a customer unreviewed. Anything with legal or contractual weight. Anything where being wrong is expensive to undo.

The useful design pattern here is that the agent prepares and the human approves. You keep the time saving, which was mostly in the preparation anyway, and you keep the judgement where judgement belongs. Businesses that skip this step and let an agent send freely tend to find out why it was a bad idea in front of a customer.

Start with one thing

The projects that work start narrow. Pick the single most repetitive task your team does, automate that end to end, and run it alongside the manual process for a fortnight while you compare the output. Expand only once it has earned it.

Projects that start with "let's automate our operations" tend to produce a large amount of configuration and very little relief, because nobody had defined what the process actually was before handing it over.

Things to insist on

• A log of every action. What ran, on what input, with what result. You should be able to audit it rather than trust it.
• An obvious stop. One switch that halts it, that somebody who is not technical can reach.
• Your systems, your accounts. The agent should work inside the tools you already pay for and own.
• A defined scope. An agent with broad access and vague instructions is a liability, not a member of staff.

Where the value really comes from

Not from headcount. For a small business, the gain is that things stop falling through. The enquiry that arrived at 9pm on Saturday gets a reply before Monday. The follow-up nobody remembered happens. The report is ready without somebody spending their morning on it. None of those are dramatic on their own, and together they are usually worth more than the automation everyone talks about.

A fair warning

An agent built on a messy process automates the mess, faster and at greater volume. If the underlying workflow is unclear, fix that first. It is the less exciting half of the project and it decides whether the rest works.

We build these as scoped, reviewable systems rather than open-ended automations: see AI agents, or custom web solutions where it needs to connect to something you already run.