Note · AI

What an agent is — and is not

An AI agent is software that does multi-step work: read the email, decide what the client actually wants, draft the reply, check the calendar, propose three meeting times. A chatbot answers questions. A cron job repeats a fixed task. The agent sits between them — it chains steps together and makes judgment calls. That judgment is the whole reason to use one, and the whole reason to write the process down first. An agent only acts well when the steps and the rules are clear, because it is not guessing the rules — it is following the ones you wrote.

The three workflows worth handing off first

• Lead response. Speed is the measurable win. Average response times of four hours drop to minutes when a form submission auto-replies with a real answer to the lead's actual question and flags the ones that need you. The agent reads, replies, and only pages you when fit or budget is obvious.
• Invoice follow-up. Polite, escalating reminders on a schedule you set. This is where a small practice or agency reclaims real hours — accounting firms automating invoice processing report dropping roughly 40 hours a month of manual work to about 6, with a person reviewing only the exceptions.
• Client onboarding. Welcome email, intake form, calendar link, first-week checklist — fired in order the moment someone signs. No missed steps, no "sorry, I forgot to send that."

What stays human

Anything where the relationship is the product: complex complaints, upsell conversations, renewals, and the first call with a big new client. Automate the speed and scale — first responses, confirmations, routine FAQ — and keep the moments that decide whether someone stays a customer. The small-business advantage over a corporation is the personal connection; automating every touchpoint erases exactly that.

The order that does not waste money

1. Write the workflow down first. Every step, every decision, who owns it. If you cannot write it, you cannot automate it.
2. Try the dumb version first. An email template, a spreadsheet, a cron job. Plenty of workflows do not need AI, and the dumb version is free.
3. Add the agent only where judgment is needed. Reading a message, deciding fit, drafting a reply. Not for "everything."
4. Measure two weeks of baseline before you start — current response time, hours spent, error rate — then measure again after. Otherwise you cannot tell whether the agent helped.

The traps

• Data first. The analyst estimate that most enterprise AI projects get abandoned because the data is not ready applies double to small business. If your leads live in four inboxes and a notebook, fix that before you buy anything.
• The agent that generates work. A bot that "organizes" everything and emails you hourly summaries is a new job, not a removed one. If the output does not change a decision or an action, it is noise.
• Cost creep, not the sticker price. Entry-level agents run about $9–$25 per agent per month, which is fine. The budget blowup is paying for agents on workflows that were never written down — an agent cannot fix a process that does not exist.

The checklist before you buy anything

• One workflow written end-to-end, with the owner named.
• Two weeks of baseline numbers: time spent, response time, error rate.
• A dumb automation tried first wherever it can work.
• AI added only for the judgment steps, with a human approving anything that goes to a customer.
• A "what does this change?" test on every output the agent produces.

The small-business AI-agent moment is real. It just arrives in the same order as every other tool: process, then automation, then intelligence — not the other way around.

← All notes