Skip to content
be-human.ai
← Back to the journal

AI by industry

Superintelligent Agents for HVAC Companies: Maintenance Plan Renewals

How HVAC companies can build maintenance plan renewal and reminder workflows today, with clear memory, narrow permissions and approval limits that scale.

Be Human typographic cover with the title: Superintelligent Agents for HVAC Companies: Maintenance Plan Renewals
Conceptual illustration created with AI · Be Human

An HVAC company can build maintenance plan renewal workflows today by separating three things: a clean record of each customer and unit, a written rule for when a renewal enters the window, and a short list of actions an agent may take without a person. Start with one renewal workflow, keep the agent's tools narrow, and measure whether plans renew without adding callbacks. Those foundations let the workflow absorb more capable agents later without a rebuild.

Key takeaways

  • Start with renewals because the work is recurring, dated and easy to audit.
  • Keep the memory of each customer and unit separate from the rules that decide when a renewal is due.
  • Give the agent narrow, named tools and a written list of actions that need a technician or office manager.
  • Measure renewal outcomes and callback volume from the first week, not after the agent seems finished.
  • Design permissions and decision records so the workflow can pass to a more capable agent without losing its operating knowledge.

Why renewals are the right first workflow for an HVAC company

Maintenance plans create a steady stream of predictable work. Each plan has a start date, a visit schedule, terms the customer already accepted, and a list of equipment that must be serviced. When a plan approaches renewal, the office has a clear question to answer: will this customer renew, and what must be true for the renewal to be easy? That question is narrow enough to automate carefully and important enough to matter to revenue.

Renewal work also fails quietly. A reminder goes to an old email address, a condenser replaced last season still appears on the plan, or a customer who had two callbacks in the spring receives a cheerful offer without anyone noticing the history. Those failures cost trust, not only a sale. Because the failure modes are visible, a renewal workflow gives the team something concrete to inspect.

A general dispatch agent that touches everything is a tempting first project, and it is usually the wrong one. Dispatch involves technicians, parts, weather and emergencies. Renewals involve a contained set of records and a clear end state: renewed, declined or paused. The dispatch and callback side of the business has its own problems, which are covered in the plumbing dispatch and callbacks guide, and it is better handled after renewals run cleanly.

What the agent needs to know about each customer and unit

An agent that writes renewal messages without the right context will produce polite messages that are wrong. Before any capability is added, the business needs an accurate memory of each account. For a renewal workflow, that memory usually includes:

  • The plan terms: start date, renewal date, visit count and what the plan actually covers.
  • Each covered unit: location on site, make and model where known, install date and last service date.
  • Service history from the last two or three visits, including open issues the technician flagged.
  • Recent callbacks, complaints or declined quotes, so the message does not ignore a sore point.
  • Contact preferences: who approves the renewal, which channel they use and when they do not want messages.
  • The office's pricing rules, referenced from the source document rather than copied into every prompt.

Keep the account record as the source of truth and let the agent retrieve only what the current message needs. Anthropic's guidance on context engineering makes the general point that an agent works better with the smallest set of high-signal context for the task at hand. For a renewal, that means the current plan and its units, not the entire customer history. The same separation between a knowledge source and the model that uses it appears in retrieval-based designs generally. Every factual claim in a draft, such as which unit is covered, should point back to a record a coordinator can open.

The recall logic used for chiropractic visit recalls follows a similar pattern: the reminder is only as good as the visit history behind it.

A step-by-step plan for the first renewal workflow

Use this sequence to build the first version. Do not add the next capability until the current step produces a record a person can check.

  • Step one: export active plans, covered units and last visit notes into one account view, and mark every record the office would not trust today.
  • Step two: write the renewal rule in plain language. A plan enters the window a fixed period before its renewal date, and it leaves the window when the customer renews, declines or a person pauses it.
  • Step three: define the message types the agent may draft, and list the facts each message must cite from the account view.
  • Step four: give the agent read access to the account view and write access only to a draft queue. Nothing is sent in this phase.
  • Step five: have a coordinator review every draft for two weeks, recording whether each one was approved, edited or rejected, and why.
  • Step six: allow sending for one message type only, such as a reminder that restates existing terms with no change, and keep everything else in the draft queue.
  • Step seven: review the edit log and decline reasons each month, then decide what to widen, one change at a time.

Which actions an agent may take alone and which need a person

Write the boundary down before any tool is granted. The list below is a starting point for an HVAC office, to be adjusted by the people who run it.

  • The agent may read plan terms, unit records and service history.
  • The agent may draft a reminder that restates existing plan terms without change.
  • The agent may schedule a reminder at a time the office has approved.
  • The agent may record a customer reply such as "not renewing this year" as a note on the account.
  • The agent must not offer a discount, change plan pricing or promise coverage the plan does not list.
  • The agent must not renew, cancel or pause a plan without a named person approving it in the system.
  • The agent must not add or remove covered units until a technician has confirmed the equipment on site.
  • The agent must hand any message that mentions a complaint, a warranty question or a safety concern to a person.

These rules belong in the system, not only in the instructions. OWASP's authorization guidance is a useful reference for this: access decisions should be enforced by the application and checked on each request, rather than assumed from how the agent was prompted. Customer messages also deserve suspicion, because a reply can contain text written to steer the agent. The OWASP guidance on prompt injection prevention describes why instructions found inside retrieved content should be treated as data, not as commands.

Illustrative scenario: a customer replies to a renewal reminder asking whether the plan covers an air handler that another company installed last year. The agent records the question on the account and routes it to the office manager, because coverage for equipment the company did not install requires a person to check the plan terms and the unit record.

How to measure whether renewals are closing

A draft queue is not a working workflow. Measure at the level of the account first: did the plan renew, did it lapse, and did the customer call with a complaint within a short period after the reminder. Those outcomes live in the office system and do not depend on anyone's judgment of the model.

Then measure the agent's own behavior: how often coordinators edited drafts, the reasons for edits, which tool calls failed, and how often the agent handed off. OpenTelemetry's observability primer describes traces, metrics and logs as complementary ways to understand a running system. For this workflow, a trace should show which account was read, which records were used, what draft was produced and which person approved it. Keep that trail even when nobody looks at it daily, because the first complaint about a renewal often arrives weeks later.

Build a small evaluation set from past accounts, including the awkward ones: a replaced unit, a customer who moved, a plan with two spring callbacks. Rerun that set whenever the prompt, the rules or the underlying model changes. OpenAI's evals guide and Anthropic's writing on evaluating agents describe the same practice, and the key discipline is keeping the set small and maintained by a person who knows the business.

Building for more capable agents later

The useful question is not whether a future agent will run renewals alone. The useful question is whether your records and permissions will let you make that decision calmly when the time comes. Three design choices matter most.

First, keep tools narrow and named. Each tool should do one thing with clear inputs, such as reading unit history or creating a draft reminder. The Model Context Protocol is one open way to describe how an agent connects to external tools and data, and it rewards the same discipline: small, documented capabilities that can be reviewed separately.

Second, give the agent its own identity with the lowest permissions it needs, and revoke or narrow them whenever the workflow changes. A more capable model does not need broader access to do the same job well.

Third, record decisions in a form that survives a change of tools. A renewal decision should be a stored record with a reason, not a line buried in a chat history. The NIST AI Risk Management Framework organizes this kind of work into governing, mapping, measuring and managing risk, which is a practical frame for deciding what the office should review and how often.

Be Human designs and operates this kind of workflow as a managed service. An engagement starts with one useful role, the memory and tools that role needs, the approval boundaries and an observable success criterion, such as renewals closing without added callbacks. After launch, managed operations can include monitoring, recurring evaluations on past accounts, integration maintenance when the field service system changes, permission reviews and controlled improvements. The work is remote and available in English and Portuguese.

If you want to find the first task in your own operation that is worth handing to an agent, the free Be Human AI diagnostic is a practical place to begin.

Frequently asked questions

Do we need a new field service platform before building a renewal agent? You usually do not. The workflow needs a reliable account view of plans, units and visit history, which many offices already keep in their current system. If that data is scattered across spreadsheets and inboxes, cleaning it up is the first step, and it is worth doing regardless of the agent.

How many customers should be in the first pilot? Start with a set small enough that a coordinator can review every draft during the early weeks. The exact number matters less than whether the team can read each draft, check its facts and record why it was changed. Widen the group only after the edit log shows the same kinds of changes are no longer needed.

Can the agent offer a renewal discount to close a lapsing plan? Not in a well-designed workflow. Pricing changes, discounts and promised coverage should stay with a named person, because they carry financial and contractual consequences. The agent can surface that a customer is hesitant and route the case to the office manager with the relevant history attached.

What happens when the agent is replaced by a more capable one? If the workflow was built with narrow tools, a documented rulebook, stored decision records and an evaluation set of past accounts, the replacement can be tested against the same cases before it touches live customers. The business keeps its operating knowledge, and the new agent inherits a workflow rather than starting from nothing.

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.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →