.png)
AI for Roofing and HVAC Companies: What to Automate First, and What to Never Hand an Agent
Somewhere in the last six months, an AI vendor showed a roofing or HVAC company a voice agent answering a call at 9 pm on a Sunday. Maybe it was yours.
It was impressive, and the underlying problem is real. Invoca's 2026 home services benchmarks found that 52% of inbound calls are answered by a person. Filter out misdials and quick hang-ups by looking only at calls lasting more than 30 seconds, and the answer rate reaches 73%.
Read it the useful way. Roughly one in four people genuinely trying to reach a home services business does not get through. In a trade where a no-heat call in January goes to whoever answers first, that is not a rounding error.
Answering the phone is the obvious first move, and it is usually the right one.
The hard part is the next one. An agent that answers your phone is also an agent that can say things on your behalf. Where that line sits is your call, and it rarely comes up before you have signed something.
Two systems and one execution layer
Before the lists, get the architecture straight, because most contractors have it wrong in a way that costs them later.
Salesforce holds business state. The customer, the job, the price, the stage, the crew assignment.
Slack holds operational conversation. What people flagged, discussed, and decided.
AI is not your system of record. It is the execution layer sitting on top of the systems that hold your data and your conversations. Agentforce reads from connected data and takes actions within the permissions, actions and guardrails you configure.
That distinction tells you where the risk sits. An agent is only as good as what it reads, and it is only as safe as what you let it do. Both settings can hurt you, and they fail in different ways. Bad permissions produce visible damage. Bad data produces invisible damage.
One terminology note. Since April 2026, Salesforce calls Topics Subagents and the Topic Selector the Agent Router. The functionality did not change, but documentation still using the old terms predates the rename.
What to automate first
Ranked by return, not by how impressive the demo looks.
1. Answering the phone when nobody can.
After hours, during storm surges, when the office line is already busy. For many home-services businesses, this is one of the clearest early automation opportunities. Your crews are on roofs and in crawlspaces, and they cannot answer.
2. Post-job follow-up.
Review requests, satisfaction checks, seasonal maintenance reminders. These tasks are easy to defer because they compete with whatever is urgent that day. An agent does it every time, within minutes of job completion rather than three days later.
3. Status updates to the homeowner.
Read-only, outbound, based on the record. "Your permit cleared, your install is still scheduled for Thursday." Many status questions are repetitive and can be handled from information already stored in the system.
4. Internal summarization.
What happened on this job, what is blocking it, what changed since Friday. This is where the Slack side earns its place. If the conversation about a job lives on the job record, an agent can summarize it. If it lives in a group text your systems cannot see, it cannot.
5. Record hygiene.
Logging call outcomes, updating fields after a completed job, flagging records with missing or inconsistent information. Unglamorous, and it compounds.
The point isn't to have an agent “clean” your entire CRM. It is to keep routine updates from becoming another task your team has to remember. An agent can handle low-risk record updates and flag anything that needs human review, based on the rules and permissions you configure. Salesforce's Agentforce can use connected business data and perform configured actions subject to those controls.
The work is simple. The consistency is the value.
What to never hand an agent
These are refusals, not "be careful" items. Configure them as hard boundaries in the agent, not as instructions you hope it follows.
Price and scope commitments.
An agent must never quote a roof replacement or commit to a scope. Roofing price depends on decking condition, layer count, pitch, and penetrations, and none of that is knowable until somebody is physically on the roof. HVAC sizing depends on a load calculation. An agent quoting from a picklist creates an expectation you will either eat or argue about, and the homeowner heard a number from your company.
Safety and code judgments.
This is the one that should stop you. A homeowner calls in January with no heat and asks whether they can relight the pilot themselves, or whether it is safe to run the system. An agent must escalate, every time, without exception. Not "provide general guidance." Escalate. The failure mode here is not a lost job.
Insurance claim and supplement language.
Anything that could become part of a claim or coverage decision should require human review. A confidently wrong statement about coverage becomes a record your team has to defend later.
Warranty determinations.
Whether a callback falls under your workmanship warranty or the manufacturer's is a margin decision and sometimes a legal one. It is also the exact question a frustrated customer asks at 8 pm, which is precisely when the agent is the only one available. Escalate.
Financing and credit outcomes.
Eligibility decisions can enter regulated territory. An agent can explain the financing options available. It should not independently determine or communicate whether a customer qualifies.
Binding crew capacity.
An agent committing a crew to Thursday without checking materials, permits and measure verification is how you generate a reschedule, and a reschedule is expensive in a way that never shows up as a line item.
The pricing and scheduling refusals will feel overcautious in a demo, because both are things a good vendor will show you working. They stop feeling overcautious the first time one goes wrong.
The prerequisite nobody mentions
Agentforce grounds its answers in your data through the Einstein Trust Layer, using dynamic grounding and secure data retrieval so the agent works from your records rather than generic model knowledge. That is the right architecture, and it has a consequence.
The agent inherits your org.
Grounding retrieves what your configuration points it at. It does not adjudicate what it finds. So the quality of the answer depends on whether your records agree with each other, and nothing in the retrieval step tells you when they do not.
If the same homeowner exists on several records with different job histories, retrieval surfaces some of that and not all of it, according to how you configured it. If two automations write the same status field in different orders, whatever the record currently holds is what gets retrieved. If the field a question depends on is populated on some jobs and blank on others, the blanks come back as blanks.
None of that produces an error. It produces an answer.
A messy org with a person in front of it produces a coordinator who says "hang on, let me check something." A messy org with an agent in front of it produces a fluent answer, quickly, with nothing to indicate what it was missing.
If you run Improveit 360 or Salesforce Field Service, start there. Those objects hold the job data an agent will reach for first, and they usually carry the most accumulated mess.
Where this argument is wrong
The obvious objection to a list this long is that it hollows out the agent. If it escalates safety, pricing, warranty, and scheduling, what is it doing at 9 pm?
Escalating is not the same as going silent. An agent that captures the situation, asks the right diagnostic questions, checks whether this is an emergency, books what is safely bookable, and tells the caller a named person will confirm in the morning has already beaten voicemail by a wide margin. The homeowner spoke to someone.
Their details are in your system. Your team opens Monday with a triaged list instead of six hang-ups.
The refusals do not remove the value. They remove the version of the value that costs you a lawsuit.
The other limit is real too. Not every contractor needs an agent strategy, and the threshold is not about size. It is about whether your bottleneck is information handling. If your problem is that you cannot hire installers fast enough, or your close rate is soft, or material costs are eating margin, an agent will not touch any of that. You will have automated the part of the business that was working.
The test is simple. Look at where your team loses hours this month. If most of that time goes to answering, chasing, updating, and summarizing, an agent has real upside. If it goes to selling, installing, or fixing, start somewhere else.
Not sure your org is ready for an agent? Implementology's Agent Quickstart begins with a strategy workshop to identify your highest-value use cases, then covers environment setup, permission configuration and an initial agent build, typically in three to four weeks. We start with the data the agent needs to read and the boundaries it should never cross. Talk to our Agentforce team.
FAQ
What should roofing and HVAC companies automate with AI first?
Answering the phone when nobody can. Roughly one in four people genuinely trying to reach a home services business does not get one, and in emergency trades the first contractor to answer usually wins the job. Post-job follow-up and homeowner status updates come next.
Can an AI agent quote a roof replacement?
It should not. Roofing price depends on decking condition, layer count and pitch, none of which is knowable before someone inspects the roof. An agent quoting from a template creates a commitment your company then has to honor or dispute.
Is Agentforce worth it for a contracting business?
It depends on whether your job data already lives in Salesforce. Agentforce grounds its answers in connected data using your existing permissions, so it is strongest when Salesforce is already your system of record. If you run a vertical CRM and only use Salesforce for sales, the agent has less to work from.
What does an AI agent actually cost a contractor?
Licensing is the smaller number and it changes often, so check current pricing with Salesforce or your partner. The cost that surprises people is the preparation. Cleaning the objects an agent reads, defining escalation boundaries, and running a pilot through a full season all take real time. Budget for that work rather than assuming the licence covers it.
Do we need to clean our CRM before deploying AI agents?
You need to clean the objects the agent will read, not the whole org. Duplicate customer records and contradictory status fields produce confidently wrong answers, because grounding reflects your data rather than correcting it.
If the agent escalates everything important, what is it actually doing?
Triage, capture and booking. It answers, identifies whether the situation is an emergency, gathers the details your team would otherwise chase, books what is safely bookable, and hands over a complete record. That is a large improvement on voicemail without putting a commitment in the agent's hands.
What is the difference between an AI chatbot and an AI agent?
A chatbot holds a conversation. An agent holds a conversation and then completes the task, checking availability, booking the slot, and writing the outcome back to your system. For a contractor, that difference is the whole point, because a missed booking is lost revenue rather than a lost chat.
Comments