ServiceTitan AI Agent Integration: Fields to Expose and Writes to Approve
Which ServiceTitan fields to expose to an AI agent first, which write actions need dispatcher approval, and where read-only access should stop.

A ServiceTitan AI agent works best when it reads a narrow set of job, appointment, customer and technician fields first, and when every change it proposes waits in a queue until a dispatcher approves it. Start with read access to job status, service location, appointment windows, assigned technicians and job types. Add write actions one at a time, each tied to a named approver and a clear reason to exist.
Key takeaways
- Expose read access to job status, appointment windows, service locations and technician assignments before anything else.
- Keep payment, cost, margin, payroll and pay-rate fields out of the agent's reach in the first release.
- Require dispatcher approval for any action that moves a technician, changes a confirmed time, cancels a job or messages a customer.
- Give every create call an idempotency key so a retry never books the same visit twice.
- Judge the agent against one observable outcome, such as fewer unassigned jobs at the start of the day, before widening its scope.
Which ServiceTitan records should an agent read first?
A field service platform holds many kinds of records, and a dispatch agent does not need most of them. For scheduling work, the core objects are jobs, appointments, customer and location records, technician assignments, job types and business units. Estimates and invoices matter for follow-up work, but they carry money fields that a scheduling agent should not handle at all.
Before you design any tool, open the current ServiceTitan developer documentation and list the objects your credentials actually expose. Coverage depends on the modules in your account and on how your administrator scopes the integration, so any list published in an article, including this one, is a starting point rather than a contract. Ask for a dedicated integration identity with the smallest scope that still works. The same logic is described in least-privilege permissions for agents.
Which ServiceTitan fields should you expose in the first release?
Expose the fields an agent needs to answer a dispatch question, and nothing that answers a finance question. A reasonable first read set includes:
- Job identifier, job type name, business unit and current job status
- Scheduled appointment start and end, with the time zone of the business unit
- Assigned technician identifier and display name, plus a flag for unassigned jobs
- Service location address and access notes that staff already treat as shareable
- Customer first name and preferred contact channel, without the full contact history
- Timestamp of the last status change and the user who made it
- Dispatcher-visible job notes only, filtered by the note visibility setting
Some fields should stay out until a later review: payment status, card or bank references, invoice totals, cost and margin lines, technician pay rates, call recordings, and free-text notes that may contain financial, health or access-code details. Customer details are personal data, so the agent should receive only what the task needs. A simple rule protects you here: if a field does not change the next dispatch decision, leave it out.
Time fields deserve extra care. An appointment window is only useful when it is read in the business unit's local time, so the tool response should state the time zone explicitly instead of expecting the model to infer it. Mixed time zones across business units are a common source of wrong reschedule proposals.
Which write actions should require dispatcher approval?
Read access is the safe place to begin. Writes come next, and the question is not whether the agent can perform them but which ones a person must confirm. A workable rule: the agent may draft, propose and queue any change, but it never commits a change that affects a technician's day, a customer's confirmed time or money until a named dispatcher approves it.
Writes that should always require approval in the first release:
- Reassigning a technician to a job that already has an appointment
- Moving a confirmed appointment to a different date or window
- Cancelling an appointment or marking a job as cancelled
- Sending any message to a customer about timing, price or service changes
- Changing the job type or business unit of an existing job
- Creating an estimate or any document a customer will see
Some writes may move to lighter approval after a review period, if the business chooses: adding an internal note to a job, filling empty template fields on an unassigned job, or creating a draft job for a request with no conflicting appointment. Even these should be logged under the agent's own identity.
Writes that should stay out of the agent at this stage include issuing refunds, changing invoice amounts, applying discounts, editing payment methods, and changing technician pay or permissions. Those decisions belong to people with the authority and context to make them.
The approval itself must be specific. Approve the exact change shown on the screen, with the current appointment time and the proposed time side by side, not a general instruction such as 'handle the schedule.' The approval step in Slack is one practical place to put this, provided the channel is limited to the people who are allowed to decide.
What limits must hold before the integration goes live?
Four limits matter most in a ServiceTitan integration.
The first is data freshness. The agent works from the schedule as it was at the last read, and a dispatcher may have moved a job minutes earlier. Every proposed change should re-read the target appointment just before it enters the queue, and the approval screen should show the current state, not the state from the original request.
The second is duplicate writes. Network timeouts happen, and a retried create call can book the same visit twice. Each create request should carry a key derived from the request itself, so the receiving system recognizes a repeat instead of creating a second record. Stripe documents this pattern for its own API, and the same design applies to field service writes. The idempotency guidance for agent actions covers the details.
The third is rate limits. A morning review can read many jobs at once. Batch the reads, keep results for the planning window, and back off when the vendor returns a throttling response instead of retrying in a tight loop.
The fourth is identity and audit. Each action should record which agent identity proposed it, which person approved it, the values before and after, and the time. OWASP's authorization guidance treats this kind of enforced, reviewable access control as a baseline, and without a trail a wrong reschedule is hard to explain to a customer and hard to fix.
Tool design matters as much as connection setup. Name each tool for the dispatch task it serves, type its inputs, and return errors that say what went wrong and what the agent may do next. Anthropic's engineering guidance on writing tools for agents is a useful reference for this work.
What is a safe first integration plan?
- Choose one business unit and one dispatch decision to support, such as flagging unassigned jobs before the morning meeting.
- Ask the administrator for a dedicated integration identity limited to the read objects in your field list.
- Build read tools for jobs, appointments and technicians, each with typed inputs, explicit time zones and clear error messages.
- Run the read-only agent against past days and compare its flags with what dispatchers actually handled.
- Add one write tool, such as proposing a reassignment, that only creates an approval request in a queue.
- Connect the queue to the channel where dispatchers already work, and name the approver for each action type.
- Enable approved writes with idempotency keys, and log the proposer, the approver and the before and after values.
- Review permissions and the failure log on a fixed schedule, and widen scope only when the success criterion holds over that period.
What does a rescheduling request look like in practice?
Illustrative scenario: a technician finishes a job later than planned on a busy morning. The scheduling agent reads the afternoon appointments, finds two that would start before the technician can reach them, and drafts two options: a later window for one customer, and a swap with an unassigned job nearby. Nothing changes in ServiceTitan yet. The dispatcher sees both options with the current appointment times, chooses one, and the approved write runs once under its idempotency key. A customer message is drafted separately, and the dispatcher approves its wording before it is sent. The log shows who proposed the change, who approved it, and the previous window.
Be Human handles this problem by defining the role before any connection is built. A first engagement names one dispatch role, the ServiceTitan objects and fields it may read, the write actions it may propose, the approver for each action, and one observable success criterion. Managed operations then keep the integration working over time: monitoring failed calls, running recurring evaluations against real dispatch decisions, reviewing permissions when modules or staff change, responding to incidents when a write goes wrong, and making controlled changes to tools. The service is delivered remotely, in English or Portuguese.
If you are not sure which dispatch task should go first, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
Can an AI agent write directly to ServiceTitan? It may be possible, depending on the credentials and modules your account exposes, but the safer design lets the agent propose writes into an approval queue. Direct writes can come later for low-risk actions, such as internal notes, once the audit trail and the success criterion have held through a review period.
Which ServiceTitan data should never go to an agent in the first release? Payment details, card or bank references, technician pay rates, margin and cost lines, and free-text notes that may contain private or access information should stay out. If a field does not change the next dispatch decision, leave it out. Add fields later only with a written reason.
How long does a first integration take? The timing depends on how many business units and modules are involved, how quickly internal credentials are approved, and how much history you want the agent to test against. A narrow read-only pilot on one business unit is usually the fastest way to learn what your data actually looks like.
What happens if the agent proposes a wrong reschedule? Nothing changes without an approval, so the dispatcher can reject the proposal before it reaches the schedule. If an approved change still causes a problem, the audit log shows who proposed and approved it, and incident response can reverse it and fix the tool that produced the error.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


