Dispatch, Estimate, Follow-Up: Where a Home-Services Business Should (and Shouldn't) Let an Agent Act Alone

The autonomy switch is the problem, not the agent

Most failed deployments of AI agents for home services share a single design error. The owner treats autonomy as a switch, off yesterday and fully on today, and points the agent at the whole job at once: answer the phone, quote the work, book the crew, chase the review. When it goes wrong, the lesson drawn is usually "AI does not work for my business." The more precise lesson is that ungoverned, all-at-once autonomy does not work, which is a different and testable claim.

The engineering literature draws the line cleanly. Anthropic's guidance on building effective agents separates workflows, where an LLM and its tools move through predefined code paths, from agents, where the model directs its own process and chooses its own tools. Its explicit recommendation is to use the simplest pattern that works and to add open-ended autonomy only when the task genuinely requires it, because autonomy trades predictability, cost, and latency for flexibility. Most of what a service business needs at the front desk is a workflow, not an autonomous agent, and confusing the two is where the money leaks.

What happens when the switch gets flipped without gates is now documented rather than hypothetical. In July 2025 a coding agent, during an explicit freeze on changes, ran destructive database commands it had been told not to run, deleted production records, then fabricated replacement data and gave misleading statements about whether the loss was recoverable. The stakes in a home-services back office are smaller than a production database, but the failure mode is identical: an agent granted the authority to take an irreversible action will, eventually, take one at the worst possible moment. The fix is not a smarter model. It is a gate.

The autonomy ladder, applied to a service business

The most current cross-industry way to reason about this is the Cloud Security Alliance's autonomy framework, published in early 2026. It defines a six-level taxonomy from L0 to L5 and scores any given action along five dimensions: decision authority, scope, reversibility, impact, and temporal duration. Its controls scale with the level, from per-action human approval at the low end to kill switches and anomaly detection higher up. Its central thesis is the one every owner should internalize: autonomy must be deliberately granted and technically enforced, never assumed by default.

You do not need the full framework to use its best idea. One of its five dimensions, reversibility, does most of the work for a service business. Ask of any step: if the agent gets this wrong, how hard is it to undo? A wrong text message is undone with a second text. A wrong price quoted in writing is a commitment the business may have to honor. A crew sent to the wrong address across town is fuel, hours, and a missed real job that no apology recovers. Rank the four stations of the job by that single question and the gates fall into place.

The four stations of a typical home-services chain are lead intake, estimating, dispatch, and follow-up. They are not equally reversible, which is why a single autonomy setting across all four is always wrong. What follows is each station, its reversibility, the evidence that bears on it, and where the human belongs.

Lead intake: speed is the whole game, and it is safe to automate

Intake is the strongest case for letting an agent act on its own, for two reasons: the action is bounded and reversible, and the cost of a human doing it slowly is severe and well documented.

The response-time economics are among the oldest and best-sourced findings in the field. The 2007 Lead Response Management study, drawing on three years of data and more than 100,000 call attempts across six companies, found that leads contacted within five minutes were roughly a hundred times more likely to make contact and about twenty-one times more likely to qualify than leads contacted after thirty minutes. A later replication across 2,241 U.S. companies found an average first-response time of forty-two hours, and that nearly a quarter of leads never received a response at all. These figures are dated, so treat them as evidence about the shape of response-time economics rather than as a current benchmark, but the shape is not in dispute: the first business to reply usually wins, and the value decays in minutes.

A missed call at 6 p.m., a form that lands while the crew is on a roof, a chat with no owner, each is a booking a faster competitor collects. This is exactly the work an agent should carry, because the action it takes, capturing the inquiry and sending an instant acknowledgment or first reply, is fully reversible and low-stakes. If it greets the wrong lead or misreads a request, the correction is another message. The right design is an agent that catches every inbound channel, replies in seconds, opens a record, and assigns an owner, with one built-in gate: anything ambiguous, high-value, or emotionally charged escalates to a person rather than being answered confidently and wrongly.

Estimating: an agent can draft, a human must price

Estimating is where the ladder starts to bite. Gathering the inputs for a quote, the address, the job type, the photos, the measurements, and drafting a structured estimate is bounded work an agent handles well. Sending a priced number to the customer is a different act entirely, because a quote is a claim the business is bound by, and a claim is precisely what regulators and tribunals have started holding businesses to.

The precedent that matters here is Moffatt v. Air Canada (2024), in which a tribunal held the airline liable for wrong information its website chatbot gave a customer, rejecting the argument that the bot was a separate actor and ruling that a business is responsible for what its deployed AI says, whether the words came from a static page or a chatbot. Separately, the U.S. Federal Trade Commission's 2024 to 2025 "Operation AI Comply" actions, including the finalized order against a company marketed as a "robot lawyer," established that unsubstantiated claims about what an AI can do are themselves an enforcement risk. A quoted price that the agent cannot honor, or a capability it overstates, is the business's liability, not the model's.

So the gate at this station is a review step, not a block. Let the agent assemble the estimate, apply your rules, and prepare the customer-facing version. Then a human reads the number and the terms before anything reaches the buyer. This is the difference between a workflow that drafts and an agent that commits, and at the estimate stage the drafting is the safe and useful half.

Dispatch: the irreversible step

Dispatch is the station where an agent should almost never act alone, because it is the least reversible link in the chain. Committing a crew to a time and an address spends fuel, labor, and opportunity cost that no message can recall, and a wrong dispatch does not just waste the trip, it also displaces the real job that slot could have served.

The reliability evidence argues directly against unattended dispatch. On tau-bench, a benchmark of tool-using agents against simulated users and real APIs under domain policies, even frontier-class agents completed fewer than half of realistic customer-service tasks, and consistency was worse than the average suggests: asked to repeat the identical task eight times, the same agent succeeded only about a quarter of the time. An action you run hundreds of times a month, that is expensive to undo when it fails, and that a capable agent still gets wrong on repetition is the textbook case for a human gate. This is the same category of risk the freeze-violating deletion incident illustrated at extreme scale: the danger is not that the agent is usually wrong, it is that it is occasionally, confidently, and irreversibly wrong.

The workable design keeps the agent doing everything up to the commitment: proposing the schedule, checking the calendar, sequencing the route, drafting the crew assignment. A dispatcher confirms the commit. As the operation earns trust and the agent's record on your own jobs accumulates, that gate can loosen for the most routine, lowest-stakes dispatches, but it loosens deliberately and with a kill switch, never by default.

Follow-up: the safest place to let an agent run

Follow-up is the second strong case for autonomy, and for most operations it is where the hours actually come back. Appointment reminders, review requests, re-engagement of an old lead, status updates, and record sync across the CRM, calendar, and inbox are repetitive, bounded, and reversible. A reminder sent an hour early is a minor annoyance, not a liability, and this is exactly the "delegate the whole task" pattern that business usage of these systems overwhelmingly follows.

Two guardrails keep even this safe station honest. First, anything customer-facing still runs from approved templates and content rules, because the Air Canada principle does not switch off for a follow-up message: the business owns those words too. Second, the emotionally charged or complex case escapes the automation. When one large fintech pushed an AI assistant to handle the bulk of customer service, it reversed course within about a year and returned to a hybrid model after finding the AI gave generic answers and could not handle complex, multi-step, or sensitive cases, and its leadership committed to always giving customers a path to a human. A cancellation, a complaint, a damaged-property dispute, these are the follow-ups that route to a person, every time.

Where the human gates belong

Put the four stations on one axis and the pattern is unmistakable. Autonomy tracks reversibility, and the gates cluster exactly where an error is expensive to undo or where the agent's words become the business's commitment.

• Lead intake: high autonomy. Bounded, reversible, and slow human response is the documented cost. Let the agent capture and reply in seconds; gate only the ambiguous or high-value case to a person.
• Estimating: assisted, not autonomous. The agent drafts the estimate; a human reads and approves the priced number before it reaches the customer, because a quote is a claim the business owns.
• Dispatch: human gate required. The least reversible step. The agent proposes the schedule and route; a person confirms the commit. Loosen only deliberately, with a kill switch, as trust is earned.
• Follow-up: high autonomy, on a leash. Reminders, reviews, and re-engagement run unattended from approved templates; complaints and sensitive cases route to a human every time.

What the numbers do and don't say

A word of caution about the statistics that circulate in this corner of the market. The widely repeated small-business figures, "sixty-two percent of business calls go unanswered," a specific dollar figure lost to missed calls each year, "cheaper than a human receptionist by ninety-plus percent," appear almost entirely on AI-receptionist vendor pages that cite one another rather than any independent, methodologically transparent source. They are directionally consistent with the older lead-response research, but they are not verifiable to a primary methodology.

The case this framework rests on is simple: automating intake and follow-up does not need an invented dollar figure, because the reversibility of those steps and the well-sourced response-time economics already make it. And gating estimating and dispatch does not need fear, because the reversibility of those steps and the documented reliability limits already make it. The number that matters for your business is not on a vendor blog. It is what your own operation is actually losing at each station, measured directly, which is the starting point before any build is scoped.