The word “agent” has been stretched to cover everything from a chat box on a website to an imaginary employee who runs the company while you sleep. Neither one is useful to an owner trying to decide where the money goes. The working definition is narrower than the first and far less dramatic than the second.
An agent has three things: access to the systems where the work actually lives, one job it is responsible for, and permission to take specific actions in those systems. Access, a job, and a door it is allowed to open. Remove any one of the three and you have something else — a report, a macro, or a demo.
agent vs. automation
An automation fires when X happens, does Y, every time. A new lead lands in the CRM, so a text goes out. An invoice ages past thirty days, so a reminder queues. The logic is written down in advance, it runs the same way on the thousandth job as on the first, and when the situation does not match the script it either does the wrong thing or does nothing at all.
An agent reads the situation across systems and decides which Y — then does it, or asks. Given that same aging invoice, it can check whether the job is actually closed, whether a supplement is still open with the carrier, whether a check came in that nobody applied, and whether the last two emails to that address bounced. Then it drafts the right message, or flags that the invoice should not be chased at all.
Both have a place, and the order matters. Most companies need automations first: they are cheaper, they are predictable, and a good share of the admin drag in a home-service business is genuinely mechanical. Bring in agents where judgment is involved — where the right next step depends on facts scattered across three systems, and a person currently spends fifteen minutes assembling those facts before deciding anything.
what agents do today in a company like yours
None of these are hypotheticals. Each one is a pattern already running somewhere, including inside our own restoration company.
- intake — a recorded walkthrough or intake call becomes a draft scope of work, with rooms, affected materials, and a first pass at line items, waiting for an estimator to edit instead of a blank form.
- closeout audits — every job that closes is checked against the books for work performed but never billed: change orders that never made it to an invoice, equipment on site longer than the billing says.
- reporting — the weekly numbers pulled from the CRM, the books, and the ad accounts, assembled, written up in plain language, and posted to Slack before the meeting starts.
- follow-up — open estimates and aging receivables chased on a cadence, each message drafted from what is actually in the job record, on a schedule a human approves.
- negotiation prep — before an adjuster call, the file assembled: what was estimated, what was billed, what was paid, which photos support the disputed line items, and what the last few conversations said.
The common thread is assembly. Almost none of that work is hard. All of it is tedious, all of it requires pulling from more than one place, and all of it gets skipped when the week gets busy — which is precisely the week it is worth the most. The closeout and reporting patterns are what our scorecard runs on live numbers, because a number nobody assembles is a number nobody trusts.
the human stays on the trigger
The rule we use, in our own company and in what we build for members, fits in one sentence: agents read everything, draft anything, and execute only what is reversible.
Reversible work runs unattended. Reading a call transcript, drafting a scope, reconciling two systems, updating an internal record, assembling a report — if any of that comes out wrong, someone edits it and nothing left the building.
Irreversible work stops for a person. Anything that goes to a customer, to a carrier, or to the bank waits in a queue with a name on it. That is not a limitation we plan to quietly remove once the technology improves. It is the line that makes the rest of it safe to run fast: an agent that occasionally drafts the wrong thing is still a good trade, because the wrong draft gets caught at a desk instead of in a customer's inbox.
where to start
Pick one workflow that eats admin hours and has a clear paper trail. The paper trail matters more than owners expect. If the work already leaves a record — a call recording, a closed job, an invoice with dates on it — then the agent has something to read and you have something to check its output against. Work that lives only in one person's head is the last thing to hand over, not the first.
Say your AR sits at 47 days and someone spends a day a week chasing it. That is a reasonable first candidate: it repeats, the rules can be written down, the record exists in the books, and the output is a draft a human approves before it sends. Intake is the other common starting point, for the same three reasons.
Then measure the hours it gave back before you build the second one. We build this work as automations and ai agents, and the live intake module in axial is this exact pattern in production — a call becomes a scope of work that an estimator corrects instead of writes.