Lesson 5 of 6 · Agents

An agent in your business, safely

Four lessons of theory come down to one decision: which job, with which permissions, checked how. Here is a way to pick the job, the guardrails that should be on before day one, and a ladder to climb instead of a switch to flip.

Everything so far was about understanding agents. This lesson is about using one, in a real Southern Utah business, in a way you would be comfortable explaining to your accountant, your insurance agent, and the customer on the other end of the email. The rule of thumb from lesson 1 carries the whole thing: treat it like a first-week employee with a computer login.

Picking a job an agent can actually do

Good first jobs share four traits. They are frequent (daily or weekly, so the setup pays back). They are written down (the inputs are emails, files, forms, not phone calls and hunches). They have a clear finish line (you can look at the output and say done or not done in ten seconds). And they are cheap to get wrong at first, because at first it will sometimes be wrong.

Three jobs that fit that shape for most small businesses:

Inbox triage. Read every new email, tag it (quote request, existing customer, vendor, junk, needs the owner), and write a morning summary. Read tools only. Output is one message to you. If it mislabels something, you notice in the summary and nothing happened.
Quote drafting. For each quote request, pull the matching prices from your sheet and fill your quote template into a draft. A Washington gutter company with a per-foot price list and a standard template is a perfect fit. The draft is a file or a Drafts entry; a human sends it.
Bookkeeping prep. Every week, read the receipts folder and the bank export, categorize each line, flag anything it cannot match, and write a sheet your bookkeeper starts from. It never touches the books. It prepares the pile.

Jobs that do not fit yet: anything with a phone in it, anything where the finish line is "the customer is happy," and anything that moves money. Not because agents cannot be pointed at them, but because you cannot check the work fast, and checking the work is the job you are keeping.

A test that settles most arguments: describe the job to a friend in one breath, then describe how you would know it was done right in one more. If the second sentence is "I would just know," it is not a first agent job. If it is "the sheet has a row for every receipt and the totals match the bank," it is.

Guardrails that go on before day one

These are not advanced settings. They are the default posture, and each one maps to something you learned in lesson 2 or 4.

Read-only first. Every agent starts with no write tools at all. It reads, it reports to you, that is the whole job for the first stretch. The permission sheet from lesson 2 is the contract.
Human approval on anything outside-facing. When write tools do come on, anything a customer, vendor, or the public could see gets an approval prompt showing the full action, not a summary.
Logs it cannot edit. Every action the agent takes is written somewhere the agent has no write access to. Most harnesses keep a transcript; make sure you know where it is and that you look at it on a schedule.
No payments. Full stop. No card on file, no payment tool, no refund tool, no ad-spend tool. Not this year. If a product only works with a card attached, it is not a first agent.
A fence. The agent works in a copy of your data, a separate mailbox, or a folder made for it. The best-case outcome of a mistake in a fenced setup is that you delete the fence and start over.

That list will feel like overkill on a job as small as inbox tagging. It is meant to. Guardrails are cheap to install on day one and expensive to add after something has gone out with your name on it.

The rollout ladder

Rather than deciding once how much to trust an agent, climb. Each rung turns on a little more, and each rung has a signal you collect before moving up. If the signal is not there, you stay, and you fix the instructions rather than loosening the permissions. Tap through.

Tap a rung: what unlocks, and what you watch before climbing

Unlocks: The agent reads: inbox, calendar, review feed, the quote sheet. It writes one thing, a daily summary to you.
Guardrail: No write access anywhere. No tools that send, move, or pay.
Signal before the next rung: Two weeks of summaries you agree with. If it keeps misreading which emails are quote requests, stay here and fix the instructions.

A pattern, not a product. Whatever tool you use, the rungs are about which permissions you turn on and what evidence you collect first.

Two notes on the ladder. First, rung four is the top for most businesses, and it stays narrow: a short explicit list of boring, reversible jobs. Money, deletions, and anything to the whole customer list live behind a human permanently, and that is not a temporary state of the technology. It is how accountability works. Second, the ladder goes down as well as up. A bad week, a new tool plugged in, a new person on the team: drop a rung, collect the signal again, climb back.

The hard part, simply

The hard idea: the agent is a probability, and your process has to be built for the times it is wrong. Not the average case. The five percent case.

Analogy: a new scheduler at a St. George insurance office who is right 95 percent of the time. Sounds great. Twenty calls a day means one wrong appointment a day, every day. Whether that is fine depends entirely on what "wrong" costs and whether someone catches it. If the mistake is a slot booked at 2 instead of 3 and the front desk confirms every appointment, it is a non-event. If the mistake is a policy cancellation and nobody reviews, it is a lawsuit. The accuracy did not change between those two sentences. The process did.

Worked example: the quote-drafting agent for the gutter company. Imagine it produces 40 drafts a week and gets 38 right. Rung two setup: all 40 sit in Drafts, you read each before sending, the two wrong ones (it used the steel price for aluminum) take you twenty seconds to fix. Cost of the errors: zero. Time saved: most of a day. Now imagine someone skipped to full access because the first week looked perfect. Same 38 right. The two wrong ones went out. One customer accepted a quote $400 under your cost. You eat it or you make an awkward call. Same agent, same accuracy, same week. The only variable was whether the process assumed it would be wrong sometimes. Build for the two, not the thirty-eight.

Where this fits in your week

An agent at rung two or three does not remove you from the job. It changes your part of it from doing to reviewing. That is a real change and a real win, but it is a different kind of work, and it is worth naming what it looks like so it does not quietly get skipped.

A morning routine at rung two: read the summary (two minutes), open Drafts, send or fix each one (ten to twenty minutes), glance at the log for anything you did not expect (one minute). Weekly: read the full log once, and update the instructions with anything it keeps getting wrong. "Aluminum is the default unless the customer says steel" belongs in the instructions the first time it matters, not the fifth.

One more piece of hygiene: tell people. A line in your email signature or on the quote ("drafted with AI assistance, reviewed by our team") costs nothing and gives you the honest position if anything ever slips through. Customers in Washington County have been reasonable about this; they are far less reasonable about being surprised.

Try this yourself

Before you sign up for anything, write the plan. Paste this into the AI app you already use and fill in the blanks. The answer is your rung-one setup, and it doubles as the note you send your bookkeeper or your business partner so nobody is surprised.

I run a [type of business] in [city]. I want to start with an AI agent at "watch only": it reads, it reports to me, it changes nothing. The first job is [inbox triage / quote drafting / bookkeeping prep].

Write me a one-page rollout plan with four rungs. For each rung: which tools are on (read, write, or off, tool by tool), which actions need my approval, what the daily and weekly review looks like, and one concrete signal I must see before moving to the next rung. Rung four must still have no payment tools and no mass sends. End with a two-sentence note I can send to my bookkeeper explaining what the agent will and will not touch.

If the plan looks like more oversight than the job is worth, that is useful information too: pick a smaller job. When you are ready to think about where all this is heading and which promises to believe, that is the last lesson.

Next lesson6. Where this is going

Last updated August 24, 2026