Meet the agents — six a small business might build

Agents at Work — CC BY 4.0

Talk about “AI agents” in the abstract and it stays abstract. So here are six an ordinary small business could genuinely put to work — the ones I keep coming back to through this course, because each teaches a different way to get it wrong. Once you can see the six failure modes, you can spot the one hiding in whatever you build. Working through them is Anchor 1 in practice: you learn the tool by studying where it breaks.

Read them as a menu, not a curriculum. You don’t need all six. You need to recognise which risk your agent carries.

The Recruiter — the risk is other people’s data

Reads and ranks job applications. It looks like an ideal agent job: a repeatable read over a stack of files. It’s also the one where the comfortable safeguards fail — the evidence is that these tools carry real bias, stripping the names doesn’t remove it, and a human “just signing off” mostly rubber-stamps the machine’s pick. We go deepest here, because it teaches the most valuable skill an agent-builder has: knowing which work an agent shouldn’t touch, however capable it gets. It’s also the one with the sharpest legal exposure — someone else’s rights are on the line.

The Bookkeeper — the risk is your numbers

Reconciles the accounts, chases overdue invoices, drafts the GST return. The data is mostly yours, so the privacy stakes are lower — but a wrong figure carries your name to a customer or the tax department. The risk here is plain accuracy, and the discipline is checking in proportion to what a mistake would cost.

The Legal Aid — the risk is the advice boundary

Reads a contract, flags clauses, answers “is this allowed?” Genuinely useful for a first pass — and genuinely dangerous when it invents a clause or a citation, or quietly slides from “general information” into what reads like legal advice you’d act on. The discipline is knowing the tool’s limits and where the line to a real lawyer sits.

The Competitive Analyst — the risk is the web

Watches competitors, tracks prices, scans tenders. The trap is treating “found on the internet” as “true,” acting on a confabulated competitor claim, or scraping in ways that breach a site’s terms. The discipline is provenance — where did this come from, and can I stand on it?

The Market Analyst — the risk is the research

Reads the trends and hands you a synthesis. The failure is subtle: a confident, fluent summary that has quietly laundered a guess into a “finding.” The discipline is refusing to let polish stand in for evidence — a cited fact and a plausible-sounding claim are not the same thing, however smooth the prose.

The Front Desk — the risk is committing the business

Triages the inbox, sends quotes, books jobs, answers customers. The danger isn’t a wrong summary — it’s an agent that makes an offer, agrees a price, or commits you to a date. A tool can’t contract on your behalf, but it can put words in your mouth that a customer will hold you to. The discipline is a hard gate on anything binding.

The one question that runs through all six

Different work, different failure — but one question underneath every panel: what’s the one way this could go wrong that I’d have to answer for? For the Recruiter it’s an unfair rejection; for the Bookkeeper a wrong number; for the Front Desk a promise you didn’t make. Name that first, and the rest of the design follows.

Start with one, start small

You learn this by doing it once, small, on work you already understand — not by reading six lessons and then building the hardest thing. Pick the agent nearest your bottleneck, give it the narrowest useful slice of the job, and grow its role as you learn how it behaves. The discipline transfers: get one right and you can see the shape of the next. The rest of Tier 1 is about choosing that first slice well — which work is fair game, and the question most guides skip: whose data, and whose call.