Superintelligent Agents for Plumbers: Dispatch and Estimate Follow-Up

An AI agent for a plumbing company is a software worker that answers calls and web inquiries, sorts each request by urgency, books or proposes a visit within rules you set, and follows up on estimates that have gone quiet. In its useful form today, it handles intake, triage, confirmations and reminders, and passes anything involving money, safety or exceptions to a person. A more capable agent would keep richer context about each home and each past job, which would sharpen triage and follow-up, but the approval boundaries would still be yours to write.

Key takeaways

• Start with intake and triage. An agent that captures the address, the problem, access notes and urgency is worth more than one that sounds impressive in conversation.
• Write dispatch rules as a short decision guide: what the agent may book alone, what a dispatcher must confirm, and what goes straight to an emergency path.
• Estimate follow-up is mostly a memory problem. The agent needs the current quote version, the visit notes, the photos and the contact history in one place.
• Keep pricing, discounts and warranty exceptions with people until your records and approvals have been tested.
• Judge the agent by observable operational measures, reviewed on a schedule, and change the rules when those measures drift.

What a dispatch agent should handle today

A useful first version of a plumbing dispatch agent does four jobs well. It answers first contact at any hour, collects the facts a technician needs, sorts the request into a clear category, and confirms what happens next. Everything else extends those four jobs, and each extension should be added only after the first one runs reliably.

The minimum intake record is short. It should contain:

• The service address and any access notes, such as gate codes, pets or a building entrance that changes after hours
• The problem, described in the customer's own words
• A direct answer on whether water, gas or sewage is involved
• A preferred time window and a phone number the technician can use
• The language the customer wants to use, English or Portuguese

The agent should record what the customer said, not what it assumed. If a field is missing, it asks once and writes down the gap instead of guessing.

Urgency needs the most care. Your written protocol should name the situations that go straight to an on-call person, including any report of a gas smell, which should trigger your emergency instructions immediately and never a routine booking. The agent applies the protocol. Your licensed staff writes it and owns it. For the callback side of the same workflow, see Superintelligent Agents for Plumbing Companies: Dispatch and Callbacks.

Where call triage goes wrong

Most triage failures are not dramatic. They come from a form field labeled preferred date that the agent treats as a commitment, from two customers with the same street name in different zones, or from a word like urgent that means different things to different people.

Illustrative scenario: a customer writes through the website that water is dripping from a ceiling beside a light fixture and asks for the first available morning. A weak setup books the earliest open slot because the request names a time. A configured agent recognizes water near an electrical fixture as an escalation category in the protocol, gives the safety steps your team has approved, and sends the case to the on-call dispatcher with the address, the description and the timestamp. The schedule follows from the protocol, not from the customer's preference.

The lesson is that the agent's first question is not how to be helpful. It is which rule applies. Helpfulness comes after classification is correct.

A decision guide for booking and escalation

Write this guide before you configure anything. A short list is easier to audit than a long set of instructions, and it gives dispatchers and the agent the same standard.

• Book alone: a routine repair or inspection at a known address, with a standard scope and an open window in an assigned zone.
• Book, then notify the dispatcher: a repeat customer with an existing job record and no open complaint.
• Escalate at once: any report of a gas smell, sewage backup, active flooding, or water near electrical equipment.
• Ask a dispatcher first: a new address outside the defined service zone, or a job that may need a specialty the first available technician does not hold.
• Hand to a person: discounts, warranty exceptions, price changes, disputes about past work, or any change to a job already in progress.

Every escalation should carry a short summary so the person receiving it does not have to call the customer again to start over. Anthropic's engineering guidance on building effective agents makes a similar point: start with simple, well-bounded workflows, and add autonomy only where it clearly earns its place.

Estimate follow-up is a memory problem

Many plumbing estimates are sound and still stall. The technician writes the quote, the customer says they will think about it, and the file waits in an inbox. Follow-up usually fails less from lack of effort than from lack of context. Nobody can quickly see which version was sent, which options were priced, or what the customer asked after reading it.

An agent can carry that context if it can read the current quote version, the photos and scope notes from the visit, and the history of contacts. It also needs a clear sequence for when it may speak and when it must stop. Write the sequence before you switch anything on:

• Confirm that the quote attached to the job is the version the customer actually received.
• Send the first follow-up in the customer's preferred channel, restating the scope in plain words.
• Ask one question that moves the decision: approve, request changes, or pause.
• If there is no reply, follow up once more on the cadence you set, then hand the file to a person.
• Record the outcome with a reason code, such as price, timing or a second opinion, so the next quote reflects what you learned.

The agent should never improvise a discount to win the job. If the customer raises price, the case goes to whoever owns pricing. Painting contractors face similar timing questions, and the approach is covered in Superintelligence Prep for House Painters: Estimates and Color Records.

What a more capable agent would change

A far more capable agent would not simply sound better on the phone. It would hold more of the relevant history at the moment of decision and reason across that history with fewer errors. For dispatch, that could mean reading a home's service record before the call: a water heater noted as aging, a repeated complaint about the same drain, or a previous visit where the access instructions changed. It could match the job to a technician whose notes show experience with that kind of work. It could also notice that several calls from one neighborhood describe the same symptom, which may deserve a dispatcher's attention before the next call arrives.

For estimate follow-up, the change is mostly in preparation. A capable agent could draft a revised estimate from visit notes and photos for a person to check, identify which question a customer is most likely stuck on, and adjust its message for a homeowner who has already said the price feels high. None of that should be sent without review until your records show the drafts are reliable.

More capability also raises the stakes. An agent that can read more and act in more places has a larger surface for error and for misuse. A customer message can contain text meant to override your rules, which is why the OWASP prompt injection guidance recommends treating untrusted input as data, not as instructions. Access should be narrowed to what each task requires, following the principle of least privilege described in OWASP's authorization guidance. The right response to a more capable agent is tighter, task-level permissions that are reviewed on a schedule, not a broader standing role.

Choosing the first role and what to measure

Be Human's managed approach starts with one useful role, usually intake and triage or estimate follow-up. The engagement defines the memory the agent may read, the tools it may call, the approvals a person must give, and a success criterion you can observe. Operations then cover monitoring of conversations and bookings, recurring evaluations against real past calls, permission reviews, incident response when something goes wrong, and controlled changes when your rules change. The service is remote and runs in English or Portuguese.

Measure the role with signals your dispatch lead can check in a few minutes:

• Time from first contact to a human-confirmed booking or escalation
• Share of intake records with a complete address and access notes
• Accuracy of emergency routing on a test set of past calls, reviewed by your senior dispatcher
• Estimates with a logged follow-up inside the window you agreed on
• Bookings a dispatcher had to correct, with the reason recorded

Review these measures weekly during the first month. If one is not moving, change the rule behind it rather than rewording the instructions. If you want to see which of these tasks is worth handing to an agent first, take the free Be Human AI diagnostic.

Frequently asked questions

Can an AI agent answer plumbing calls after hours? Yes, for the intake part. It can answer, collect the address and problem, classify urgency against your written protocol, and confirm what happens next. Anything that needs a licensed person to act, including emergencies, should route to your on-call contact with a complete summary.

Should an AI agent set prices or approve discounts on estimates? Not in a first engagement. Pricing rules, discounts and warranty exceptions belong to the people who own those decisions. The agent can explain the scope of a quote that has already been approved and record the customer's objection for review.

What happens when the agent is not sure whether a call is an emergency? It should treat uncertainty as a reason to escalate. Your protocol should say that any unclear water, gas or sewage report goes to a person immediately, and the agent should tell the customer that plainly. The case is recorded, so you can see how often uncertainty occurs and adjust the rules.

How is customer data protected when an agent reads job records? Access should be limited to the fields each task needs, granted per task, and reviewed on a schedule. Customer messages should be treated as untrusted input, and any action that changes money, schedules or records should require a logged approval where your rules call for one. Data handling should also follow the privacy rules that apply wherever your customers are.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.