Five AI Agent Patterns That Work in a One-Person Business
Past the demos and the hype, a handful of agent patterns reliably earn their keep for solo operators. The five we deploy most, and what each one needs to work.
We've built a few dozen agents for solo operators now, and the honest finding is this: most agent ideas don't survive contact with a real business. The ones that do keep collapsing into the same five shapes. Not fifty. Five.
None of them are the autonomous everything-assistant from the demos. Every one of them runs with a human in the loop by default, because in a one-person business, you are the brand — an agent that sends something wrong sends it as you. The posture that works is: agents prepare, you approve. Here are the five, with what each needs and how each one fails.
1. The morning drafter
Works overnight, queues drafts for the morning. Follow-ups from yesterday's calls, replies to routine inquiries, the check-in email to the client you haven't touched in two weeks. You wake up to eight drafts, approve five, edit two, kill one. Twelve minutes instead of ninety.
What it needs: access to your sent folder and calendar, plus 20–30 samples of emails you actually wrote, so it drafts in your register and not LinkedIn-speak. And a hard rule: it never sends. It only drafts.
Failure mode: approval fatigue. Around week three you start bulk-approving without reading. The fix is volume discipline — a drafter that queues 8 good drafts beats one that queues 25 mediocre ones.
2. The intake researcher
Fires when a new lead arrives. Before you've even seen the inquiry, it has pulled the company's site, their LinkedIn, recent news, and a guess at size and budget, and attached a half-page brief to the lead record. Every first call starts with you already oriented — 20 minutes of pre-call research you no longer do at 11pm the night before.
What it needs: a trigger (form webhook or inbox rule), somewhere structured to write to, and a fixed brief format. Same five sections every time, so your eye knows where to look.
Failure mode: confident nonsense — the wrong "Acme Consulting," a revenue figure from 2021. Label everything unverified as unverified, and treat the brief as a head start, never as fact.
3. The extractor
Turns unstructured input into structured records. Call recording in, CRM update out: decision-maker, budget range, timeline, objections, next step. Rambling three-paragraph email in; task with a deadline out. This is the pattern that quietly fixes the "my CRM is always stale" problem — the notes stop depending on your discipline at 5:47pm on a Friday. It's the engine behind the CRM that updates itself.
What it needs: a transcription source, a strict output schema, and a review step where you glance at what it extracted before it writes anywhere permanent.
Failure mode: schema drift. Let it freestyle new fields and in a month your records are inconsistent again — the exact disease it was hired to cure. Lock the schema down.
4. The watchdog
Monitors one thing and speaks only when action is needed. An invoice crossing 30 days. Weighted pipeline dropping below your floor. A client who normally replies in hours going quiet for six days. The rest of the time: silence.
What it needs: read access to one system, a numeric threshold, and restraint. The whole value is the signal-to-noise ratio.
Failure mode: becoming a newsletter. A watchdog that reports "all clear" every morning trains you to ignore it, and then you miss the one alert that mattered. If it has nothing to say, it says nothing. This is boring on purpose, and that's the point.
5. The chaser
Polite persistence, industrialized. Unpaid invoices, unanswered proposals, unscheduled kickoff calls — anything that needs a nudge on day 3, day 7, and day 14. Humans are terrible at this because each individual nudge feels both awkward and skippable. An agent doesn't feel awkward, and for one client this pattern alone pulled average payment time from 38 days to 16.
What it needs: the state of each open item (paid? replied? booked?), a short sequence of escalating-but-warm templates in your voice, and clear escalation rules: after nudge three, or the moment a reply contains anything unexpected — a dispute, a complaint, a "can we talk?" — it stops and hands the thread to you.
Failure mode: chasing on bad data. One nudge sent for an invoice the client already paid costs more goodwill than ten timely nudges earn. The chaser is only as good as the records it reads — which is why it usually gets deployed after the extractor, not before.
The pattern behind the patterns
Look at the list again. Drafts, briefs, records, alerts, nudges. Every output is either reviewed by you or trivially reversible. Nothing signs contracts, quotes prices, or makes promises. That's not timidity — it's what makes these agents deployable this month instead of "someday, once we trust it."
Start with whichever one maps to your ugliest recurring hour. For most solo operators that's the morning drafter or the chaser. Run it two weeks, keep score of minutes saved and errors caught, and only then add the second. One reliable agent beats five flaky ones, every time.
Want this running in your business?
Youductive audits how you operate, writes your custom automation playbook, then builds the agents and apps that execute it. Fixed pricing, two-week delivery.
Get your playbook