AI agents for contractors: what data they need to work well

What makes an AI agent different from a chatbot#

An AI agent differs from a chatbot because it takes actions in your systems, such as booking a job, assigning a technician, drafting an estimate or sending a payment reminder, rather than only answering questions. That makes an agent's quality depend on two things a chatbot never needed: the history it reads and the permissions it holds.

For a contractor, those systems are usually the field service platform, the phone and call tracking system, accounting and the customer messaging tool. An agent connected to them inherits every shortcut, gap and inconsistency that lives in the records, and it repeats them at a speed no dispatcher could.

Agent by agent: minimum records and permissions#

Each contractor agent needs a different minimum set of records and permissions, and the table lists them for the agents most companies consider first. Treat the human checkpoint column as a starting policy and loosen it only after an agent has proved itself on your own jobs.

Why your history matters more than the model#

Your job history matters more than the model because an agent learns your business rules from the records, not from a manual nobody wrote down. If the same repair is booked under three different job types, a booking agent cannot learn which slot length and which technician it needs.

Durations, skills and outcomes are the usual weak spots. Job durations recorded only as scheduled windows, not actual arrival and completion times, teach a dispatch agent the wrong lengths. Skill tags set when a technician was hired and never updated send the wrong person to the wrong equipment. Callbacks logged as new jobs hide which repairs failed, so a follow-up agent cannot spot a pattern.

The same weak spots limit reporting today, so fixing them pays off whether or not an agent is ever switched on.

Which agent to try first#

The best first agent for a contractor is the one whose rules are simple to check and whose mistakes are cheap to catch. Starting there builds trust in the records and the review process before an agent touches the dispatch board or pricing.

Match the choice to the problem that costs you most today, then confirm the readiness items for that agent before switching it on.

Data readiness checks before you switch an agent on#

A data readiness check before switching an agent on catches most of the avoidable mistakes. Work through the list with the service manager and the office lead, and fix the items that affect the agent you plan to try first.

• 1Job types: one job type per kind of work, with retired types closed rather than left selectable.
• 2Technician skills: current certifications, equipment experience and service areas for every technician.
• 3Price book: current prices, with old items retired rather than left active.
• 4Customers and locations: duplicates merged, with do-not-contact flags and access instructions in their proper fields.
• 5Actual times: arrival and completion times recorded, not only scheduled windows.
• 6Callbacks: linked to the original job rather than logged as new work.
• 7Payment terms: recorded on commercial accounts, so collections messages match the agreement.

Permissions and guardrails that keep agents safe#

Agent permissions should follow the same rule as a new hire's: only what the job requires, widened as trust is earned. Many field service platforms and integration tools let you create a separate user or connection for the agent, which keeps its actions traceable and easy to switch off.

Keep payment card data out of the agent's reach, and route refunds, discounts and payment plans to people. Confirm that call recording and messaging disclosures cover automated systems where your state requires it, and check with counsel where you are unsure. Review the agent's action log closely in the period right after launch.

Check the vendor's terms as well. Some agent products use customer interactions to improve their models, so confirm what data the vendor keeps, for what purpose and whether you can opt out, before connecting the agent to customer records.

Illustrative: a mechanical contractor pilots two agents#

Illustrative: a fictional mechanical contractor with residential and light commercial service tries a booking agent on after-hours calls and a collections agent on overdue invoices. The COO gives both agents their own user accounts with read-only access for the first phase, so every proposed action lands in a review queue.

The booking agent's draft appointments reveal that the same maintenance visit appears under several job types created by different managers over the years. The team merges them and closes the duplicates before granting write access. The collections agent handles residential balances well but stumbles on commercial accounts with negotiated terms that were never entered in the system, so the controller adds the terms and keeps disputed invoices with a person.

Neither fix involved the agent itself. Both came from records the company had been living with for years.

Your records and the agents being built for the trades#

Developers building agents for the trades need real examples of how bookings, dispatch decisions, estimates and collections actually played out, and contractors are among the few organizations that hold them. A company's linked history can be licensed under defined terms while the company keeps ownership.

At SourceX, such a license follows the SourceX five-step transaction: Supply, Rights, Preparation, Approval and Delivery. Preparation strips customer names, addresses, phone numbers and access details before release, and the company approves the scope and the release itself. The first step needs only a description of the records, not the records.