Note · AI

The three-question filter

Before you buy anything, take the task you have in mind and answer three questions.

1. Does it repeat at least weekly, in roughly the same shape?
2. Can you write the rules on one page, exceptions included?
3. If it fails silently for a week, do you find out?

Three yes answers makes the task a candidate. Two makes it a project. The third question is the one most owners skip, and it is the one that decides whether you keep using the thing in month 3. An agent that breaks quietly is worse than no agent, because you stop checking the output and start trusting it.

The first three jobs worth automating

1. Lead response. A form submission or missed call answered in minutes instead of the next morning. The pattern that works is drafting, not deciding: the agent writes the reply, books the slot it sees open, and flags anything that smells like a fit problem for a human. Reported response-time improvements are dramatic, from hours to under 5 minutes, and the sales effect comes from the speed, not the wording.

2. Receivables and invoice chasing. Every unpaid invoice on a polite schedule: day 7, day 14, day 30, then a human. This is the most boring automation on the list and usually the fastest to pay for itself. Somebody has to reconcile payments against the ledger, and that somebody can be a scheduled job with a shared inbox. Treat hours-saved figures from vendors as their numbers, not yours.

3. Intake and scheduling follow-up. The welcome email, the form, the calendar link, the reminder the day before, in order, every time. No judgment required, and the failure mode, a client who never got the reminder, is exactly the kind of thing small businesses eat the cost of personally.

Each of those three has the same shape: a trigger, a written rule, an output a human can scan in under a minute.

What to leave alone

Leave the agent out of anything where it moves money, gives regulated advice, or writes something a customer could quote back at you in a dispute. Leave it out of pricing decisions and out of hiring, firing, and clinical or legal judgment.

That rule is narrower than it sounds. A receivables agent can draft the escalation note without having access to the bank account. An intake agent can book the appointment without touching the payment detail. Scope, not capability, is what keeps this safe.

What it actually costs

Three cost buckets, and only one of them is software.

• Tooling. Automation platforms and agent seats commonly land between $20 and $200 a month for a small business, depending on how many tasks run and how often they fire.
• Setup. Doing it yourself costs 2 to 4 weeks of part-time work on one workflow. Paying an agency or a freelancer typically runs a $1,500 to $5,000 setup plus a $300 to $900 monthly retainer, and that range moves with how much of the integration is custom.
• Maintenance. Forms change, calendars break, an API key expires. Budget an hour a month per live workflow, or the retainer that covers it.

Those are reported market ranges with real spread, not quotes, and not guarantees. Ask for a fixed-scope pilot price before you sign a retainer.

Run a 14-day pilot before you commit

The pilot is the whole game, and it is cheap if you run it in this order.

1. Days 1 and 2, the baseline. Log how the task runs today: minutes per run, runs per week, errors per 20 runs. Write it down. Without a baseline you cannot tell improvement from novelty.
2. Days 3 and 4, the rule sheet. One page: trigger, steps, the fields that matter, and what the agent does when it is unsure. Escalate is a valid step.
3. Days 5 to 11, live but supervised. Run real cases with a human approving every outbound message. Count the ones you had to fix and why.
4. Days 12 to 14, the verdict. Compare against the baseline. If the error rate is higher than the human error rate was, stop and keep the process, not the tool. A written process that a person follows is still a win.

The four ways this fails

1. No written process. An agent cannot automate a workflow that exists only in your head and three inboxes. It will reproduce the confusion faster.
2. No owner. Someone has to read the exception log weekly. Name the person, even if the answer is you for 30 minutes on Fridays.
3. No exception path. Every task has a case the rules miss. If the agent has nowhere to send that case, it guesses.
4. Buying a platform to find a task. This is the expensive order. Pick the task first, then the cheapest thing that finishes it, which is often a form, a filter, and a scheduled reminder. Some workflows end there, and that is a good outcome.

The 30-day sequence

• Days 1 to 3: pick one workflow you have done by hand, at least weekly.
• Days 4 to 7: write the rule sheet, and run the dumb version first.
• Days 8 to 14: add the agent only where judgment is needed, with a human approving anything a customer receives.
• Days 15 to 21: measure against the baseline you wrote before you started.
• Days 22 to 30: decide, in writing, whether to keep it, fix it, or kill it. Keeping a workflow that saved 4 hours a week for $80 a month is the whole point.

The bottom line

The businesses getting value from agents this year are not running exotic stacks. They picked a boring, repeating job, wrote down how it should work, and made the agent prove it against a number.

Not financial, legal, or tax advice: verify vendor claims, pricing, and your own workflows with qualified professionals before acting.

← All notes