Every field service software vendor now has an AI story. Some of it is useful. Some of it is a demo that works on stage and nowhere else. And some of it would make you nervous if you read the fine print: software that sends messages to your customers or changes invoices without anyone looking.
The useful question isn't "should we use AI agents in the office?" It's "which tasks should an agent do by itself, which should it only suggest, and which should it stay away from?" This post gives you a framework for answering that, examples by function, a way to measure whether an agent is earning its keep, and a list of red flags in vendor claims.
A two-question risk framework
Before turning on any agent, ask two questions about each thing it might do.
1. Can it be undone, cleanly? Adding an internal note, flagging an invoice for review or drafting an email can be reversed with no one the wiser. Sending a text to a customer, charging a card or paying a supplier can't be taken back.
2. Who sees it if it's wrong? An internal mistake costs some time. A mistake the customer sees costs trust. A mistake involving money costs money, and sometimes a phone call you'd rather not make.
Put those together and you get three levels:
- Low risk: run automatically. Reversible, internal, low-dollar. Tagging, categorizing, matching, drafting, flagging. If it's wrong, someone fixes it in a minute.
- Medium risk: suggest, a person approves. Anything a customer will see, or anything that changes a dollar amount on a record. The agent does the work of preparing it; a person clicks approve.
- High risk: keep a person in charge. Irreversible money movement, legal or credit decisions, anything unusual. The agent can surface information, but the decision and the action stay with a person.
Most back-office work falls into the first two. That's good news: it means an agent can take a large share of the effort out of a task while a person still makes every call that matters.
Why the approvals queue matters
"AI proposes, a person decides" only works if proposals land somewhere sensible. That place is an approvals queue: one list of everything agents want to do, each with the reasoning and the records it would touch, and approve, edit or reject buttons.
A good queue gives you three things.
Batching. Your office manager reviews twenty collection reminders in five minutes at 9 a.m., instead of being interrupted twenty times.
Context. Each proposal shows why: "Invoice 4512 is 21 days past due, the customer paid the last three invoices by text link, the last reminder was 7 days ago." If the reasoning is wrong, you'll see it.
A dial, not a switch. As you trust a specific agent with a specific task, you can move it from suggest to automatic for low-risk items only, and move it back if something goes wrong.
Examples by function
Here is how the framework plays out for four common back-office jobs. Treat these as starting points; your own risk tolerance may differ.
Invoice review
Before invoices go out, an agent checks them against the job: missing line items the tech mentioned in notes, prices that don't match the pricebook, a membership discount that wasn't applied, a missing signature or photo.
- Automatic: flagging invoices for review, adding an internal note.
- Suggest: changing a line item or a price.
- Person only: writing off a balance or issuing a credit.
Collections
An agent watches open balances and prepares reminders by the channel each customer responds to.
- Automatic: sorting the A/R list by priority, drafting reminders.
- Suggest: sending reminders to customers, applying a late fee under the rules you set.
- Person only: credit holds on a commercial account, sending anything to an outside collector.
Membership visits
An agent finds members whose tune-ups are due and proposes appointments that fit the schedule.
- Automatic: building the "due this month" list.
- Suggest: offering appointment times to the member, placing the visit on the board.
- Person only: changing what the membership includes or its price.
Supplier bills and accounts payable
An agent reads supplier invoices, matches lines to purchase orders or to the job where the tech bought parts at the counter, and flags price changes against your pricebook cost.
- Automatic: extracting lines, matching to POs, flagging mismatches.
- Suggest: posting the bill, updating a pricebook cost.
- Person only: paying the bill.
Notice the pattern. Agents do the reading, matching and drafting, which is where the hours go. People keep the decisions that are customer-facing or move money.
How to tell if an agent is helping
"It feels faster" isn't a measurement. Track these from the first week.
- Acceptance rate. The share of proposals approved without changes. Low acceptance means the agent is making work, not saving it.
- Edit rate. The share approved after edits. High edit rates tell you what the agent gets wrong; often it's a setting, not the AI.
- Undo rate. For automatic actions, how often someone reverses one. This should be close to zero before you widen what runs automatically.
- Time to decision. How long proposals sit in the queue. If they pile up, the queue is too noisy or nobody owns it.
- The business outcome. The reason you turned it on: days to get paid for collections, pricing errors caught for invoice review, visits completed on time for memberships, cost changes caught for AP.
Run a simple before-and-after. Write down the outcome metric for the month before you turn an agent on, then compare the first full month after. If the outcome hasn't moved and the acceptance rate is low, turn the agent off. That's a legitimate result.
Red flags in vendor AI claims
When you evaluate any AI feature, from us or anyone else, these are worth pushing on.
- "Fully autonomous." Ask what it does without a person, and what happens when it's wrong. If there's no undo and no queue, that's a risk you're taking on.
- No record of what it touched. Ask to see the audit log for an agent action. If it's not there, you can't investigate a complaint.
- The model writes your numbers. Language models are good with words and unreliable with arithmetic. Revenue, margins and balances should come from a calculation engine; the AI should only explain them.
- It's on a waitlist, or only in the top tier. A feature you can't use today isn't a feature yet. Ask what is generally available on your plan, now.
- Your own tools are locked out. Check whether your software's terms let you connect your own AI tools to your own data. ServiceTitan's Terms of Use, for example, bar letting third parties, including AI agents, access the service on a customer's behalf.
- Vague reasoning. "Our AI decided" isn't an explanation. Every proposal should say why, in a sentence a person can check.
How TradesWork approaches it
Our agents for invoice review, collections, membership visits, supplier bills and automations each have three settings: Off, Suggest, or Auto-apply low-risk. Every proposal lands in one approvals queue, low-risk automatic actions can be undone for 24 hours, and everything is audited. When you ask the business a question in plain English, the numbers come from the metric engine, never from the model. And because there's an MCP server, REST API and webhooks, you can bring your own AI tools under the same permissions.
Whatever you use, start every agent on Suggest. Move individual tasks to automatic only after the acceptance and undo rates have earned it.