Two dispatchers, same leads, same scripts. One books 70% of the calls she answers; the other books 40%. Over a year, on a $400 average ticket, that gap is real money walking out the door — and most owners never see it, because nobody has time to listen to 400 calls a week.
We don’t sell call-QA software, so here’s the operator’s version. AI doesn’t make your dispatchers better. It makes the invisible visible: it transcribes and scores every call, so you can coach from what actually happened on the phone instead of from the two calls you happened to overhear. The economics are simple — a few points of booking rate is worth more than most of your ad budget. The legal part is where people get sloppy, so we’ll handle that first.
The big legal caveat: record the calls the right way
Before you point any AI at your recordings, you have to be allowed to record them. This is where home-service owners trip — and it’s avoidable.
The U.S. splits into one-party and all-party (often called two-party) consent states:
- In one-party states, only one person on the call needs to consent — and since that’s you, you can record your own business calls.
- In all-party states, everyone on the call must be notified. About a dozen states require this (the count is genuinely contested — most lists land between 11 and 13). The standard list: California, Florida, Connecticut, Delaware, Illinois, Maryland, Massachusetts, Montana, Nevada, New Hampshire, Pennsylvania, and Washington.
California is strict on purpose — Penal Code §632 covers confidential communications, and §632.7 specifically requires all-party consent for cell and cordless calls regardless of confidentiality. Don’t lean on “I’m a party to the call, so I already have consent” — in California that argument is shaky, because the law turns on whether the caller was notified, not on your status as a party. Florida treats illegal recording as a third-degree felony (a first, non-commercial offense can drop to a misdemeanor). And here’s the rule that makes this simple: on a call that crosses state lines, follow the strictest law that could apply. A Charlotte shop (North Carolina is one-party) taking a call from a customer in Florida is on the hook for Florida’s all-party rule.
The fix is the one you’ve heard a thousand times yourself: play a short announcement at the very start of every call — “This call may be recorded for quality and training purposes.” Played before any substantive conversation, it satisfies the notification requirement, and the recording is legal because the caller heard it and chose to keep talking — implied consent. That one announcement covers you everywhere at once. You don’t need a per-state recording matrix; you need one announcement on every line, inbound and outbound. (This is general guidance, not legal advice — consent law evolves, so confirm your setup with counsel if you operate across many states.)
Tell your own team in writing, too. In all-party states your dispatcher is also a party to the call, so “we record and review calls for quality” belongs in the employment agreement or handbook, signed. It’s a one-line fix that closes a real gap for a process whose whole point is reviewing your employees.
Mind the data, not just the consent. Calls are full of names, addresses, and sometimes card numbers. The moment you ship recordings to an outside AI vendor, you’ve handed customer PII to a third party — so use a vendor that will sign a data processing agreement (DPA), and never pipe payment-card audio into a public, general-purpose AI tool. In California, CPRA treats this as personal information you’re responsible for. None of this is a reason to skip AI QA; it’s a reason to choose the vendor deliberately.
Feeding recordings to an AI doesn’t add a new consent layer for the caller — the announcement already handled that. The AI is just a faster way to review your own calls. The added duty is on the data side, above.
What the AI actually does: transcribe, then score
Once the recordings are legal, the workflow is two steps.
Step one — transcription. The AI turns audio into searchable text. That alone is useful: you can find every call where “warranty,” “second opinion,” or a competitor’s name came up.
Step two — scoring. This is the part that changes behavior. You give the AI a rubric — the things a great call does — and it grades every call against it. A practical home-services rubric:
- Greeting. Did the dispatcher answer with the company name, promptly, in a warm tone? A robotic or slow pickup loses callers in the first five seconds.
- Qualifying questions. Did they capture the problem, the full address, and urgency? Calls that skip qualification book worse and create bad dispatches.
- The booking ask. The single biggest one. Did the dispatcher ask for the appointment — “I can get a tech out Thursday morning, does that work?” — instead of “we’ll call you back”? Most missed bookings die here.
- Objection handling. When the caller hesitated on price or timing, did the dispatcher have an answer, or did they fold?
- Follow-up on missed calls. On voicemails and abandoned calls, did anyone call back — fast? Industry data is brutal here: most home-service callers who don’t reach a live person never call again. (One caution: if your follow-up is an outbound call or text to a missed lead, you’re now in TCPA territory — keep callbacks tied to a customer who just contacted you, and don’t bulk-text cold numbers.)
Each call gets a score and, more importantly, a flag on the exact moment it went wrong. That’s the difference between “Maria needs to improve” and “Maria booked 9 of 12 calls but never made the booking ask on the 3 she lost — coach the ask.”
Keep a human in the loop. AI transcription and scoring are good, not perfect — accents, crosstalk, and sarcasm trip them up, and a wrong transcript produces a wrong score. Treat the AI as the thing that surfaces the calls worth your attention, not as the final word. Spot-check the flagged calls yourself, and never hand someone a write-up or an HR consequence off an unreviewed AI score.
What to measure (and what each number tells you)
Scoring individual calls is the coaching tool. These three metrics are the management tool:
- Answer rate — the share of inbound calls a human actually picks up. Across home services, roughly 14% or more of calls go unanswered, and after-hours and weekends are far worse. If this is low, it’s a staffing or routing problem, not a coaching one.
- Booking rate — the share of answered calls that become scheduled jobs. Typical home-services book rates run 30–50%; 65% is strong; top CSRs clear 75%+. The 85%-and-up numbers you’ll see quoted are real but mostly come from narrow call types — maintenance and membership-renewal calls — not general inbound. This is the number AI QA moves, because it’s a function of how the call is handled.
- Abandoned calls — callers who hang up before anyone answers. Every one is a near-certain lost job and a wasted ad dollar, since you paid to make that phone ring.
Where your booking rate sits tells you how much room coaching has to work with:
The diagnosis falls out of the combination. High answer rate, low booking rate = a coaching problem; your people are picking up but not closing. Low answer rate = a capacity problem; no script fixes a call nobody answered. Most owners assume they have a lead problem when they actually have a phone-handling problem — and you can’t tell which without these numbers. Capturing them cleanly is exactly what proper call tracking is for.
Coaching from the data, not from vibes
Data only matters if it changes behavior. The loop that works:
- Score every call automatically against your rubric.
- Pull the weekly pattern per dispatcher — not 300 calls, but “here are the 6 calls each person lost, and the moment they lost them.”
- Coach one behavior per person per week. “This week, every call gets a booking ask.” Narrow beats comprehensive.
- Re-measure booking rate the next week. If it moved, the coaching worked. If it didn’t, the problem was somewhere else on the rubric.
Real-world: when shops actually run this loop, booking rate moves in double digits — some home-services teams report jumps from the 40s and 50s into the 70s once dispatchers consistently qualify and ask for the appointment. The AI didn’t make that happen. The visibility did.
One thing to watch: a rubric goes stale, and dispatchers will “teach to the test” — saying the booking-ask words without genuinely closing. Recalibrate the rubric every so often, and always anchor the score to the outcome that pays the bills: did the call become a booked job? A high call score that doesn’t convert is a rubric problem, not a win — which is exactly the loop the next two sections close.
The tools landscape (generically)
You don’t need an enterprise contract. The options stack roughly into three tiers:
- Home-services-specific platforms that score booking calls and emergency intake against contractor-shaped rubrics out of the box.
- General AI call-scoring / speech-analytics software built for sales and support teams, configured to your rubric.
- A lightweight pipeline — take the recordings your call-tracking provider already captures, transcribe them, and score against a rubric you write yourself.
The platform matters far less than the discipline. The wrong way is buying a fancy tool and looking at the dashboard once a month. The right way is every call scored, spot-checked, tied to booked jobs, one coaching point per dispatcher per week — which works on the cheapest tier just as well as the priciest. Whichever tier you pick, confirm the vendor will sign a DPA before you send it a single recording.
Tie call handling to booked jobs — or it’s just a vanity dashboard
Here’s the step most QA setups skip, and it’s the whole point. A high call score means nothing if those calls don’t turn into paid jobs. You need the loop to close all the way to revenue: which scored calls became booked, paid jobs, at what ticket, from which marketing channel.
Done right, you stop reporting “we answered 320 calls” and start reporting “we answered 320 calls, booked 198, at a $410 average ticket — and the calls we lost clustered on the missing booking ask, which cost us roughly $18,000 last month.”
That’s a number you can act on. Wiring call outcomes back to revenue is the job of conversion tracking and revenue attribution — without it, call QA is a hobby; with it, it’s the cheapest growth lever you have.
FAQ
Is it legal to record my dispatchers’ calls and run them through AI? It depends on the caller’s state, and you must follow the strictest law that could apply. About a dozen all-party-consent states (commonly California, Florida, Connecticut, Delaware, Illinois, Maryland, Massachusetts, Montana, Nevada, New Hampshire, Pennsylvania, and Washington) require everyone on the call to be notified. The universal fix is a short “This call may be recorded for quality and training purposes” announcement at the very start of every call — the recording is legal because the caller was notified and kept talking, not because you’re a “party.” Feeding the recordings to an AI doesn’t add a new consent step, but it does add a data duty: use a vendor that signs a data processing agreement and keep card numbers out of public AI tools.
What should the AI actually score on each call? The moments that move booking rate: greeting and tone, qualifying questions (problem, address, urgency), a clear booking ask, objection handling, and follow-up on missed or voicemail calls. Each call gets scored against that rubric with the failing moment flagged, so coaching targets one fixable behavior. Spot-check the AI before you coach on its scores — it misreads accents, crosstalk, and sarcasm.
Which numbers tell me if call handling is the problem? Answer rate (share of calls a human picks up), booking rate (share of answered calls that become jobs — 30–50% typical, 65%+ strong, 75%+ for top CSRs, 85%+ mainly on maintenance/repeat calls), and abandoned calls. High answer rate with low booking rate means it’s a coaching problem, not a staffing one.
Do I need an expensive enterprise QA platform? No. You can start by transcribing and scoring the recordings your call-tracking provider already captures, against a rubric you write. The discipline — every call scored, spot-checked, tied to booked jobs, one coaching point per dispatcher per week — matters far more than the tool. Just make sure the vendor will sign a data processing agreement.
The bottom line
Your booking rate swings on who answers the phone, and you can’t manage what you can’t hear. AI lets you hear all of it — legally, if you put a recording announcement on every line and pick a vendor that will sign a DPA — and turns 400 unheard calls a week into one weekly coaching point per dispatcher. Keep a human spot-checking the scores, anchor everything to booked jobs, and the payoff isn’t a tidier dashboard; it’s a few points of booking rate, which beats almost anything you can buy with the same money.
If you suspect you’re losing jobs at the phone but can’t prove where, that’s exactly what we dig into. Audit my call handling and we’ll tie your calls to booked, paid jobs so you can see where the money is leaking.