
AI for HVAC Repair Companies: Calls, Visits and Warranty Records
How an AI agent can organize HVAC service calls, technician visits, parts quotes and warranty records, with human approval before any commitment is made.
Want to apply this in your company?
Book a call with Juan →An AI agent can organize the workflow of an HVAC repair company: receive the service call, collect equipment details, suggest a visit window, prepare the parts order and log the warranty. It should not confirm a visit, approve a quote or promise a deadline before a technician or manager approves it. This article walks through that workflow step by step, the limits of each stage, and what to test before the agent talks to real customers.
Key takeaways
- The agent organizes the call, but the visit commitment and the quote amount only exist after human approval.
- Equipment data (brand, model, approximate age, symptom and serial number) is the foundation for almost everything else in the workflow.
- Warranty depends on manufacturer rules and the installation record, so the agent logs and checks but does not decide on its own.
- Every decision needs a record: who approved it, when, with what information, and what reply was sent to the customer.
- Before serving real customers, the workflow should be tested with simulated calls and reviewed by a technician.
Where the call starts and what the agent needs to know
In most shops, service requests arrive by WhatsApp, phone, a website form or a referral from an installer. The first problem is not technical. It is the lack of a standard. One customer writes 'the air is not cooling', another sends a photo of the display with an error code, and another says only 'it's dripping'. An agent that answers all of them the same way makes the experience worse, and an agent that answers without asking anything creates unnecessary visits.
For that reason, before any automation, map the minimum information the team needs to decide the next step. Usually this means the service address, the type of unit (split, window, cassette, multi-split), the brand and model, the symptom described in the customer's own words, whether there is an error code on the display, and the approximate installation date. The invoice or installation receipt becomes necessary when the case involves warranty.
This mapping also defines what the agent must not attempt. Gas, electronic board or electrical risk diagnosis requires a qualified technician. The agent can collect the symptom and classify urgency according to rules the company has written, but it should not give technical guidance that replaces a visit. When a customer reports a burning smell, water leaking with a risk of electrical damage, or a gas smell, the correct path is immediate handoff to a person on the team.
Step by step: from the call to the visit
A simple workflow, which is usually the starting point, can follow this order. Each step has a clear rule about what the agent does and when it hands the task to a person.
- Greet the customer and confirm whether it is a maintenance, installation or warranty request, without trying to solve everything in the first message.
- Collect the minimum equipment and location data, using closed questions whenever possible to reduce typing errors.
- Classify the case as urgent, scheduled or pending information, using the rules set by the company, not a free-form judgment from the model.
- Open the service record in the company system, with all fields completed and the conversation attached.
- Check available schedule slots and present window options, without confirming anything to the customer.
- Send the visit proposal to the technician or scheduling manager for approval, with a summary of the case.
- Only after approval, send the customer the confirmed window, including the company name, date, time period and anything they need to prepare.
The last step needs the most care. If the customer replies 'that works' to a window that was not approved, the agent should not negotiate on its own. It records the reply and returns the decision to the responsible person. This handoff point must be written into the prompt and also into the technical workflow, not left to the intent of whoever configured it.
If the shop uses WhatsApp, consider the customer service window rule. Messages outside the window opened by the customer follow different restrictions, and the agent needs to know when it can reply freely and when it must use an approved template. The topic is covered in more detail in WhatsApp 24 Hour Customer Service Window: How an AI Agent Should Act.
Parts quotes: what the agent can prepare and what stays with the technician
The parts quote is where a shop loses the most time and where errors most often appear. The technician has to identify the right component (compressor, capacitor, board, sensor, fan motor, filter, refrigerant), check supplier availability and calculate labor. An agent can help with the organization: build a list of likely parts from the service record, check the catalog or internal price table, prepare a draft quote and separate what still needs confirmation.
What should not be delegated is the final part selection and the approval of the amount. A part can have different compatibility across generations of the same model, and a wrong reference code creates rework and an extra visit. For that reason, the agent's draft should always show the source of each item (catalog, equipment history, customer information) and state clearly what has not yet been verified.
A practical rule: the agent can fill in, compare and suggest, but it does not send a quote with a final amount until the responsible person reviews it. This separation also prevents a language model from inventing a price or a parts delivery time. Prices and lead times must come from a controlled source, never from the model's memory.
Warranty records without promising what has not been verified
Warranty is one of the topics where customers expect the fastest answer and where the company can least afford to improvise. Coverage depends on the product, the period set by the manufacturer, the installation conditions and who performed the service. It is not something the agent should infer from a conversation. It should check the registered rule and, when there is no clear rule, hand the case to a person.
The warranty record has value even when the final answer takes time. It should store the installation date provided, the receipt sent (collecting no more than necessary), the serial number, the description of the defect, the relevant photos and the decision made. This history helps the shop spot repeats: the same model with the same symptom across several installations may point to an installation problem, a batch issue or site conditions, and that matters to both the technician and sales.
On personal data, the care point is purpose. Address, phone number and environment photos serve to handle the call, not for other uses. Brazil's data protection authority (ANPD) guides the processing of personal data for a defined purpose and limited necessity, and that principle should shape what the agent may ask for and store. If the workflow involves automated decisions that affect the customer, such as denying a warranty, the company needs human review available. The article on LGPD automated decision-making explains this in more detail.
Illustrative scenario: a call on a hot day
Illustrative scenario: a customer writes in the morning that the living room split is not cooling and sends a photo of the display showing an error code. The agent identifies the model from the label, checks whether the unit has a registered installation, asks whether there is unusual noise or water dripping, and classifies the case as urgent because the customer mentions a child sleeping in the room. The shop's rule sends urgent cases to the on-call queue.
The agent opens the record, attaches the photo and the conversation, and offers the customer two arrival options the team has confirmed as available. It does not promise an exact time. While the customer waits, the agent gives only the guidance the shop has authorized, for example, not restarting the unit repeatedly and keeping doors closed. The technician receives the summary, approves the afternoon window, and only then does the agent send the confirmation with the technician's name and the time range.
The next day, the agent logs in the history that the compressor was replaced with the part the technician specified, with the date and the service note. If the customer later asks whether the part is covered, the agent checks the record, states the registered warranty period and refers any question about coverage to the responsible person. This example is illustrative and does not describe a real shop.
How to start with one workflow
The safest way to deploy this kind of agent follows the method we use in our projects: map the current process and its baseline, limit data and actions, pilot one real workflow with records and a success metric agreed in advance, and then decide whether to expand, adjust or stop. In practice, this means starting with one type of call, for example scheduled maintenance, and leaving installation and warranty for a second stage. The metric can be simple, such as the share of calls that have all minimum data filled in from the first message, compared with today, without invented targets.
Before connecting the agent to real customers, test it with simulated calls that include ambiguous cases, off-topic messages and attempts to get the agent to approve amounts. This kind of testing reveals the limits of the workflow. For tools and permissions, the agent should have access only to the actions it needs, such as checking the schedule and opening a record, without the ability to change prices or cancel a warranty. The OWASP project describes prompt injection risk, where a customer message tries to override the model's instructions, and that is a concrete reason to separate what the agent can do from what it only suggests.
If the shop also services other devices, such as phones or household appliances, the logic of intake and status tracking is similar. An example is in AI for Electronics Repair Shops: Intake, Status and Warranty, and the combination of scheduling and parts appears in AI for Auto Repair Shops and Dealerships: Quotes, Parts, Scheduling.
About the term superintelligence in this context
Some vendors describe any customer service agent as superintelligence. For a repair company, the question that matters is different: does the system follow clear rules, record what it did, and send the right decision to the right person? An agent configured well for service calls is useful because it is predictable, not because it seems all-knowing. Read Is ChatGPT Superintelligence? What It Is and What It Is Not before accepting this kind of promise in a commercial proposal.
The NIST AI Risk Management Framework is a useful reference for organizing these risks into governance, mapping, measurement and management, without relying on marketing terms.
Frequently asked questions
Can an AI agent book a technician visit on its own? It is not recommended. The agent can check the schedule, suggest windows and prepare the request, but the confirmation to the customer should happen only after the technician or manager approves the window. This keeps the shop from making commitments it cannot keep.
Can the agent give the customer a part price and delivery time? Only if that data comes from a source the company controls, such as an internal table or a supplier system, and the final quote is reviewed by a person. Prices and lead times should never be estimated from the model's memory, because that creates promises the shop cannot honor.
How should the agent handle warranty requests? It logs the case with the required data, checks the registered coverage rule, and hands it to the responsible person when there is no clear rule or when the customer disputes a decision. It should not deny or approve a warranty on its own, and the customer needs to know how to request a human review.
What data should an HVAC agent avoid collecting? It should avoid anything not needed to handle the call, such as personal documents without a purpose, photos showing private life beyond the equipment, and sensitive data without justification. The practical rule is to collect the minimum needed for the service and record the purpose of each data point.