Where AI agents actually fit in a small business

Most of the AI content you’ll read this year is written for enterprises — companies with data teams, platform budgets, and a year to burn on a pilot. If you run a business of five to fifty people, almost none of it applies to you. Here’s the version that does.

The short version: small businesses don’t have an AI capability problem. They have a selection problem. Today’s models are already good enough for a long list of real work — the deployments that fail usually failed at the moment someone picked the wrong task, not at the moment the technology fell short.

The test that matters

An AI agent is worth deploying when a task passes three checks:

1. It happens often. Weekly at minimum, ideally daily. Automating a quarterly task buys you nothing.
2. It’s describable. If you can write down how a competent new hire would do it, an agent can probably do it. If “it depends” appears three times in your explanation, it can’t — yet.
3. A wrong answer is cheap to catch. Drafting a reply someone reviews before sending: great. Quoting a price that goes straight to the customer: not yet.

That’s it. Notice what’s not on the list: “it’s impressive in a demo.”

Where agents earn their keep today

In practice, the work that passes the test clusters in a few places:

• Inbox and channel triage. Reading incoming email, WhatsApp, or support messages, categorising them, drafting responses, and escalating the ones that need a human. This is usually the single highest-value deployment for a small business, because it runs all day, every day.
• Internal lookups. “What did we agree with this supplier?”, “What’s the status of order X?” — an agent with access to your documents and systems answers in seconds what currently interrupts somebody for ten minutes.
• Recurring document work. First drafts of proposals, summaries of long threads, meeting notes into action items, data copied between systems that were never integrated.
• Monitoring. Watching for the thing that occasionally goes wrong — a stale order, an unanswered lead, a price change — and raising a hand when it does.

Where they don’t (yet)

Equally important is the list to walk away from:

• Anything where the error cost is high and silent. Financial commitments, legal language, compliance.
• Work that’s actually relationship-building in disguise. If the email is the product — sales at the top of your market, sensitive HR conversations — automating it makes it worse.
• Processes you haven’t defined. An agent automates a process. If the process lives in one person’s head and changes weekly, fix that first; it’s cheaper than any software.

Start embarrassingly small

The failure mode we see most isn’t ambition — it’s scope. A business tries to “roll out AI”, stalls for six months, and concludes it doesn’t work. The pattern that succeeds is the opposite: pick one task that passes the three checks, deploy one agent against it, measure for two weeks, then decide. The first deployment teaches you more about your own operations than any strategy document — including ours.

That’s also why our own engagements are structured the way they are: a free call to find the candidate tasks, an assessment that puts numbers on them, and a plan your team can execute one small win at a time. Selection first, deployment second — in that order, the technology mostly takes care of itself.