Dispatch, Pricing, Diagnostics and the Job Record: Four Workflows to Hand to AI First in Home Services

Which home services workflows should be automated with AI first? Dispatch, pricing, diagnostics and the job record, if the rule is written and checkable.

Short answer: The first workflows to hand to AI in a home services company are not the ones in the demos. They are repeated decisions in the middle of the business, where the rule can be written down, the result can be checked within days, and a mistake can be caught before the customer sees it. Three of them are dispatch, pricing and diagnostics. The fourth is not a decision but the memory: the job record that says what was reported, found, quoted, done and charged. In all four, AI should prepare the first draft, and a named person should approve anything that reaches a customer. The order matters too: my working hypothesis is that it is safest to start with the record and finish with pricing. That is a hypothesis, not a finding: I found no independent study that ranks home service workflows by return.

The obvious first move with AI in a service company is the front door: answer the phone, reply to the review, post to social. If your biggest leak is missed calls, start there. But the larger losses may be in the middle, which you can test in your own numbers: who goes to which job, what it is priced at, and whether the first visit fixes the problem. This essay is about that middle. I did not search for measured results from front-door AI such as answering tools; they may exist.

Three vendors announced agents in November 2026. What they claim, and what they don't

This section reports what the vendors say about their own products. I have not tested any of them.

  • Simpro opened early access to JustDo on November 9, an AI operations agent inside its field service platforms. The company says it turns an email, photo or PDF into a draft job or quote, shows a preview, and waits for approval before anything sensitive happens. It cites an internal benchmark of about 18.5 hours a week of administration per trade business.

  • Motive launched Agents for field teams, including HVAC and plumbing, on November 9. The company says they can take over routine work such as check-in calls and driver tracking, and flag a critical fault code and open a maintenance work order. Its examples lean toward landscaping crews and fleet data.

  • Handle opened a waitlist on November 10 for an agent for local service businesses. The company's press release (Business Wire, Nov 10) says that when asked for a quote it uses what it knows about the business's services and prices and prepares a draft for the owner to review and send.

Two observations, in my reading. For quotes, Simpro and Handle's press release both describe a person approving before sensitive actions or quotes go out, but Handle's own homepage demo also shows its agent sending a quote and chasing an invoice on its own, so ask any vendor which messages need approval and which do not. Motive's agents work on fleet data and bring in managers when something needs attention. And Simpro's 18.5 hours is a baseline, not a result; its method is not published in what I read. None of the announcements I read reports measured outcomes from the agents themselves. Incumbents like these may also bundle all four workflows into one platform, which would make my ordering moot.

The test: written, checkable, reversible

I use a working test, a hypothesis for sorting workflows and not a validated method. A workflow is a good first candidate when three things are true.

  1. Written. You can write the rule down so a new hire could follow it. If the rule lives in one person's head, AI cannot follow it. I made the same point in If you can go remote, you can automate: documentation comes before automation.

  2. Checkable. You can compare the result with something real within days: an arrival window, a price book, a completed job.

  3. Reversible. A mistake can be caught and undone before the customer sees it. This is the logic of the rule in Trust Levels for AI: the further an action is from the person accountable for it, the more the AI should draft and the less it should send. That rule is my opinion, not a standard.

Against that test, as a judgment and not a measurement: dispatch is usually written, easily checked and reversible until the truck rolls. Pricing is often unwritten and not reversible once a quote is sent. Diagnostics is partly written and slow to check, but reversible because a technician is standing at the unit. The job record passes all three.

Dispatch: strong evidence elsewhere, little for home services

Dispatch is matching a job to a technician, a time window and a route, then changing the plan when the day goes wrong. In my view the rule is mostly writable: skills, location, window, parts on the truck.

The best-documented example I found comes from a parcel network, not a home service company. In a 2017 Interfaces paper, a report for the Franz Edelman competition, UPS's own engineers described ORION, a system that was then providing optimized routes to its 55,000 U.S. drivers. The modernization effort behind it began in 2003, the system cost more than $295 million to build and deploy, and the authors expected, in the abstract's words, savings of $300–$400 million a year. The abstract also says UPS needed "extensive change in management practices" so that users and executives would accept the system.

The caveats are heavy. The authors work at the company, the figures are company-reported and the savings expected, not independently measured, and ORION is optimization software, not a language model. It says little about what a 12-truck HVAC shop would save. It suggests the algorithm was not the only hard part: acceptance by the people who follow the plan mattered too.

My hypothesis: what AI adds to dispatch is reading messy inputs (a customer's free-text request, a technician's note) and proposing a plan for a dispatcher to approve. Each change the dispatcher makes, logged with a reason, becomes a rule nobody had written down.

Pricing: a rule you can write, and an agent can only follow

In a conversation with another owner this summer, my advice was to walk your competitors' customer path yourself, as a buyer, and to write the price policy before hiring anyone to sell. Do it where the strong competitors actually are, the places that pay for leads, not the cheap marketplaces; customers are an unreliable guide to what others charge. And do not measure yourself against the cheapest operators: the question to put to customers is who will still be around next year to stand behind the work.

You can hire someone to sell for you, but to hire one you need a price policy and confidence that the job is worth that price. Without it, the salesperson comes back with "we're too expensive, give a discount," you agree because you hear it from the salesperson, he collects his commission, and you are in the red. Walking the competitor's path is the founder's job.

That is my opinion, not research. My hypothesis for AI: an agent that drafts quotes can apply a written price policy more consistently than a person under pressure to close, but it cannot tell you what the policy should be. Without a policy, you hand an agent the salesperson's ambiguity at higher speed.

In my view pricing has the sharpest reversibility problem: a quote that has been sent has been said. In Moffatt v. Air Canada (2024 BCCRT 149), a British Columbia small-claims tribunal held an airline responsible for wrong information its website chatbot gave about a bereavement fare. The tribunal wrote that the airline "is responsible for all the information on its website," whether it comes "from a static page or a chatbot," and ordered it to pay $812.02 in total. That is one decision, in one Canadian province, on a small claim against an airline. It does not decide U.S. law and is not legal advice. Ask your own lawyer what your agent can bind you to.

The practical reading: an AI that drafts quotes against a written price book, with an owner or estimator approving before they go out, is a good candidate. An AI that sets prices alone is not.

Diagnostics: the advisor, not the decider

Diagnostics is the step between "the unit isn't cooling" and "this part, this fix." The rule is partly writable: manufacturer documentation, symptom patterns, the unit's history. The result is checked slowly, because the true cause is often known only after the repair. But it is reversible in the way that matters: the technician is at the unit and decides.

The nearest research is not about the trades. Brynjolfsson, Li and Raymond studied a generative AI assistant given to 5,172 support agents at one software company, most outside the U.S. (Quarterly Journal of Economics, 2025). Agents resolved about 15% more issues per hour on average, and about 30% more among the less skilled and less experienced. For the most skilled, the authors report small gains in speed and small declines in quality. That is one firm, in chat support, not field service.

A second study, from the GPT-4 era, shows the opposite risk. In a preregistered experiment with 758 Boston Consulting Group consultants doing a business strategy task chosen to sit outside the AI's capabilities, correct answers fell from 84.5% without AI to 70.6% and 60% in the two AI groups, which the authors summarize as about 19 percentage points lower on average (Dell'Acqua and co-authors, Harvard Business School working paper; the percentages are as reported on the HBS AI Institute page). Frontier models have moved since, and these were consultants on a managerial task, not technicians.

My hypothesis: read together, they suggest a diagnostic assistant may help newer technicians most and may mislead on unusual faults. That is an inference from other jobs, not a result for technicians. It argues for an easy path to a senior technician and for tracking first-time fix and callbacks. I did not find a published residential HVAC callback-rate benchmark with a disclosed sample, period and definition, so you will need your own baseline. I also found no independent controlled study of AI diagnostic tools for residential HVAC or appliance technicians; I did not rely on the vendor-published case studies that search turns up.

The fourth workflow: the job record

Three of the workflows are decisions. The fourth is the memory that lets you improve them.

By the job record I mean one structured entry per job: what the customer reported, what the technician found, what was quoted, what was done, what was charged, the photos, and whether the customer accepted. It is the least glamorous of the four, but you cannot learn which dispatch assumptions fail unless arrival times are recorded, test a price policy unless quotes, discounts and outcomes sit together, or judge a diagnostic assistant without knowing what the fix was. Without the record, the other three cannot be checked.

It also passes the test best. The rule is writable (which fields, which photos), checkable the same day (complete or not), and a bad entry is easy to fix. Trade press reports that JustDo checks completed job cards before they go to the customer, which fits this pattern; that is a vendor claim.

One practice I recommend to other owners: record a walkthrough of a task, have a note-taking tool summarize it, and have an AI model turn the summary into a written procedure, so you spend the time once. That is an operating habit, not a studied method. My investment thesis says operating data (jobs, parts, time, outcomes) is what compounds in this industry. That is my opinion about where value accrues, and the record is where it starts.

The strongest case against

The best argument against this ranking is that it has the risk backwards.

The middle is where mistakes cost the most. Answering a call or replying to a review is lower stakes. A wrong dispatch wastes a truck day, a wrong price quietly costs margin, and a wrong diagnosis can cost a callback or, with gas and electrical work, something worse. Ranked by risk, the edges go first.

The rules in the middle are often tacit. The best dispatcher, estimator and technician carry rules they could not write down on demand. If the rule is tacit, the AI misses it.

The leak may be at the front door. If calls go unanswered, no scheduling improvement matters, and I have no independent data on where a typical $1–5M company loses the most.

What does this reveal? Part of the ranking survives. The risk argument is right about pricing, which is why it goes last here and why the approval step stays; it is right that "first" cannot mean "fully automated." The tacit-rule argument points at the real work: writing the rules down. If you can't, none of the four is ready. Where your money leaks is a measurement, not an opinion.

When this fails

  • There is no written rule or the data is thin. Pricing without a price book, dispatch without agreed priorities, job notes left blank.

  • Nobody really reviews the drafts. A rubber stamp is automation with a delay. Track how often people change the draft.

  • The work is safety-critical. Where a diagnosis touches gas or electrical work, a licensed person decides and the AI stays advisory.

  • The volume is small. With three trucks, one person can dispatch from memory. I don't know where the line is below which setup costs more than it saves.

  • The economics shift. Simpro's own page says JustDo draws on a credit allotment. Measure what the work costs, including checking the drafts.

What this changes for operators, investors and founders

For owners and founders. Write the rule before you buy the tool. A price policy, a dispatch priority and a definition of a complete job are decisions only you can make; software can apply them but not invent them. As a hypothesis, a sensible order is the job record, then diagnostics as an advisor to the technician, then dispatch as a proposed plan, then pricing against a price book you trust.

For investors and acquirers. A company that says an AI runs dispatch or pricing should be able to show the written rule, the log of overrides, margin per job over time, and what it costs to run and check. Using AI does not make a home service business a software company, and it should not be valued like one. Written rules reduce dependence on the few people who carry them in their heads, which is not the same as a higher multiple.

How to test this in 30 days

Run this beside your current process.

  • Week 1: write three rules. A price book with discount limits, dispatch priorities, and the required fields of a complete job. If you can't, that is your finding.

  • Week 2: the record. AI drafts the job record from notes and photos. A person fixes it. Count the corrections.

  • Week 3: diagnostics and dispatch as drafts. AI suggests likely causes and tomorrow's schedule. The technician and dispatcher decide. Log each disagreement and why.

  • Week 4: pricing as a draft. AI prepares quotes from the price book for an estimator to approve. Compare drafts with what was sent.

Then look at four numbers: how often people changed the draft, time spent on records, first-time fix and callbacks where the assistant was used, and margin on quotes. If changes are frequent, fix the rule or the data before widening the test.

Frequently asked questions

Which home services workflows should be automated with AI first?

The ones where the rule can be written down, the result can be checked within days, and a mistake can be caught before the customer sees it. My hypothesis is that dispatch, pricing, diagnostics and the job record fit that test better than most customer-facing tasks, provided a person approves anything that reaches a customer. I found no independent study that ranks workflows by return.

Should I start with an AI answering service or chatbot instead?

If missed calls are your biggest measured leak, yes. Those tasks are lower risk. The argument here is for owners whose losses are in scheduling, pricing and repeat visits. Measure where you lose jobs first.

Can AI set my prices?

It can apply a written price policy and draft quotes against it. It cannot decide what the policy should be, and a quote that has gone out may bind you. A Canadian tribunal held an airline responsible for what its chatbot said about fares; that does not govern U.S. law, so ask your own lawyer. Have a person approve quotes until you know how often the drafts are wrong.

Where this leaves you

The evidence is thin, so here is how I weigh it. On solid ground: the vendors announced these agents in early November 2026, Simpro and a Handle press release describe a person approving quotes, and a tribunal held an airline to what its chatbot said. The research I cite is real but comes from other settings: one parcel carrier, one software firm, one consulting experiment.

What is my own view: a price policy has to exist before you hire a seller, autonomy should shrink with distance from the accountable person, and operating data is what compounds. What I do not know is where a typical company loses the most money, whether these tools beat a good dispatcher, estimator or technician in practice, and what it costs to check the drafts. Hence my suggestion: run the 30-day plan beside your current process, and let the override rate and margin per job, not the demo, decide how much more to hand over.

The demos show AI talking to customers. The money is in the decisions made before the customer hears anything, and in writing down the rules so something can check them.