What are AI agents? A plain-English guide for small business owners

A chatbot answers; an agent acts. When you type a question into a chat window, you get a reply and the exchange is over — the software has no objective beyond your current message. An AI agent is different. You hand it a goal ("find local businesses with outdated websites and draft an intro email for each"), and it plans the steps, carries them out with tools, and keeps going until the job is done or it hits a moment that needs you.

Three ingredients make software "agentic." First, a goal it holds onto across many steps, not a prompt it responds to once. Second, tools: the ability to search a business directory, open and read a website, keep a working list of leads, or write an email draft — rather than only producing text in a chat. Third, judgment between steps: deciding a lead is not worth pursuing, that a draft needs a second pass, or that the next action is one a human should sign off on.

It is worth saying plainly: none of this is science fiction, and it is not general intelligence. A good agent today is narrow software that is genuinely competent at one defined job — reading eighty small-business websites in an afternoon and scoring each one, say — and honest about everything outside that job.

Abstract definitions only get you so far, so here is one real workflow from Brohns, step by step. A Finder agent searches OpenStreetMap for businesses matching a niche and area — plumbers within 25 km of your city, for example. A Qualifier then opens each business's website and gives it an explainable 0–100 "outdated score," judging per lead whether it is genuinely worth pursuing and writing down why. An Outreacher drafts a first message for each qualified lead, opening with a real, specific observation about that company's site instead of a generic pitch.

Then comes the step that defines a trustworthy agent: it stops. Every draft lands in an approvals queue where you read it, edit it, approve it, or dismiss it — and only an approved message actually goes out, through your own email sender. Find, judge, draft, wait. That last verb is not a limitation bolted on afterward; it is part of the design.

The other defining trait is that the work is inspectable. In Brohns, each agent's actions and its actual reasoning appear on a live timeline — what it searched, why it scored a site 71, why it skipped a lead. You are not looking at a progress bar; you are reading the agent's thinking, which is how you learn whether to trust it.

Agents shine at work that is high-volume, judgment-light-per-item, and text-heavy. Concretely, they are good at:

Where they fail is just as predictable. Agents are weak on context you never wrote down (they do not know your best client came from a referral, unless told), on vague goals ("grow my business" gives an agent nothing to plan against), and on arithmetic — a language model left alone will occasionally state a wrong number with full confidence. They also do not naturally know when to stop, which is why unlimited autonomy is a bad default.

Good products design around these failure modes rather than denying them. In Brohns, for instance, an analyst agent answering questions about your imported data has its numbers computed by deterministic code — the model shapes the question, code does the math. Outreach drafts pass a second, deliberately strict review that checks for invented specifics, hype, and filler before you even see them. And every outward action waits behind an approval gate. The honest framing: agents are excellent junior staff, and junior staff need review.

A single do-everything assistant is like hiring one person to run prospecting, qualification, copywriting, and quality control at once: possible, but mediocre at each. An agent team splits the goal into single, sharp responsibilities. In Brohns, you describe the goal in plain language and Bro — the orchestrator at the center — proposes a tailored team, usually two to seven agents, each with one job: Finder, Qualifier, Outreacher, Builder.

The structure is not cosmetic; the handoffs are quality gates. The Outreacher does not write a word until the Qualifier has marked leads worth pursuing, and the Builder does not build a demo page until a lead has actually replied. Each stage filters the last, so weak inputs get caught early instead of turning into weak emails.

There is a practical management benefit too: when something is off, you know exactly whose desk it sits on. If outreach feels flat, you tune or teach the Outreacher — the Finder is fine. If you are weighing a single assistant against a team for your own situation, the comparison guide on agent teams versus single assistants walks through when each makes sense.

The question every owner should ask about agents is not "how smart is it?" but "what can it do without me?" Brohns's answer is approval-first: no email is sent, nothing is published, and nothing is spent without your explicit sign-off — and this is enforced on the server, where the recipient and content of an approved message come from the database, not merely hidden behind a button in the interface. The queue makes review fast: approve in one tap, edit first, or dismiss.

Around that gate sit standing guardrails you configure once:

Control is not all-or-nothing forever, though. Brohns uses an autonomy ladder: every team starts in approve-everything mode, and as an agent's drafts keep passing your review, you can grant it routine autonomy within limits, per ecosystem — and step it back down whenever you want. There is a kill-switch for stopping everything at once. And because sending runs through your own Resend key or Gmail account, the agent is always writing under your name, from your infrastructure, never from someone else's domain pretending to be you.

Do not begin with "automate my business." Begin with one narrow, checkable goal: more qualified leads in one niche and one region, payment reminders for your own overdue invoices, or first drafts of replies to routine support questions. The more concretely you state the audience, the area, and what success looks like, the better the team an orchestrator can design — a goal an agent can execute reads like a brief, not a wish.

Plan to review everything in the first stretch. That is not wasted time; it is how the system gets good. In Brohns, when you edit a draft before approving it, the agent distills your edit into a lasting lesson — cut the flattery, lead with the specific finding, keep it under 120 words — so the tenth draft needs less of you than the first. You can also teach agents directly, the way you would brief a new hire.

Starting costs nothing to try: Brohns comes with a 14-day free trial including 500 credits, no credit card required, and lead finding works out of the box on the free OpenStreetMap path — you only connect your own email sender when you are ready to actually send. Describe one goal, watch the team run one cycle, and judge it on what lands in your approvals queue.