Price AI work before support eats the margin
Where the hours really go on a white-label AI account, what generates the support tickets, and a pricing shape that survives the second year of delivery.
Plenty of agencies sold an AI line over the winter: answering, chat, review replies, some flavour of assistant attached to a client’s phone number. Those builds went live in February and March, so two months of delivery are now on the books and the second invoice has gone out.
This is the point where the economics show up. An AI service is cheap to sell, quick to set up and expensive to keep running, and the cost lands entirely in support. We have watched agencies price a voice agent off the build and then discover the build was the smallest part of the year.
Here is where the hours actually go on these accounts, and the pricing shape we have settled on for the white-label work we do.
The build is not the job
Setting up an agent is a day. Scoping it properly is two or three, and it is the only part that decides whether the account is profitable later.
Most of the support load on an AI account is generated by decisions nobody made during onboarding. If the scope does not say what happens when a caller asks for a price, then the first caller who asks for a price becomes a ticket, a phone call with the client, a script change, a redeploy and a test. That is an hour of senior time for a question that should have been answered on a form before go-live.
So we gate go-live. No account goes live until we have, in writing from the client rather than from their website:
- The twelve questions the front desk really gets, with the real answers.
- What the agent may never say. Prices, coverage, anything clinical, anything about money owed, anything legal.
- The handoff rules: who, which number, what hours, and what happens when that number does not answer.
- The calendar rules: which appointment types it may book, how long each takes, buffers, and how far ahead.
- Who owns the phone number and who can change the forward.
- Who at the client reads the transcripts and when.
That last one is the one agencies skip and the one that predicts the renewal.
What actually generates tickets
Sorted by how much of the month they eat, from what we see across partner accounts:
- Drift. Hours change, a service is dropped, a staff member leaves, a plan is no longer accepted. Nobody tells you. The agent keeps saying the old thing until a customer complains, and then it is an emergency.
- Script requests. The client listens to one call, dislikes one sentence, and sends a paragraph on a Saturday. Individually trivial, collectively a part-time job.
- Forwarding and numbers. Ports, the old tracking number on a van, the after-hours forward somebody changed during a storm. This is the most common cause of “it stopped working” and it is almost never the agent.
- Expectation gaps. The client believed it would handle something you never agreed it would handle. Cheapest to fix before go-live, most expensive to fix in month four.
- Usage spikes. A busy week produces minutes. If you sold unlimited, that week is on you.
- Staff turnover at the client. The person you trained left. The new person does not know the escalation path, the kill switch, or that transcripts exist.
None of those are technical problems and none of them get cheaper as you add accounts. They scale linearly with clients, which is the whole issue with pricing this like software.
The pricing shape
Four lines, and they do different jobs.
A setup fee priced on the scoping, not the clicking. If the build is six hours and getting the answers out of the client is fourteen, price the fourteen. Charge it whether or not the account goes live, because the work happens either way.
A monthly supervision fee with named deliverables. Transcript review, guardrail changes, a monthly summary, a call. Write the hours into the agreement. This is the line clients try to cut and it is the line that keeps the account working, so make it visible rather than baked into a round number.
Usage billed as usage. Voice minutes, messages, whatever the platform meters. An allowance, a stated overage rate, and an alert before the overage rather than an invoice after it. Never sell unlimited anything that a vendor bills you by the minute for.
A change allowance. One script revision per month included, additional ones billed or banked. This single line removes most of the Saturday paragraphs, because it makes the cost of a change visible to the person requesting it.
For partners reselling, the mark-up belongs on setup and supervision, not on the metered line. Marking up minutes makes you look expensive on the one number the client can check.
Delivering it without a support team
The operational side matters as much as the price. What we run per account:
- A runbook in one place: numbers, forwards, calendar rules, guardrails, escalation path, kill switch, who approves changes. If a teammate cannot take the account over from the runbook, it is not documented.
- One named owner per account. Shared ownership of an AI account means nobody reads the transcripts.
- A quarterly drift check as a calendar item, not a good intention. Hours, services, staff names, plans accepted, calendar types. Fifteen minutes that prevents the emergency.
- The kill switch in the client’s hands, documented, tested once in front of them. An account that can turn it off at nine at night will not call you at nine at night.
- A standing change window. Requests collected, applied together once a week, tested once. Not deployed one sentence at a time.
- Nothing sold as a front desk replacement. Coverage for the hours nobody is there. Sell it as that and the expectation gap never opens.
What we are telling clients
- Go back through every agreement signed since the new year and find the ones promising unlimited usage or unlimited changes. Fix those at the next invoice, not at renewal.
- Add the go-live gate to onboarding this week. No answers to the twelve questions, no go-live date.
- Time-track one live AI account for a month, split into build, scoping, supervision and unplanned support. Price the next three off that split.
- Put a named supervision deliverable on the invoice so the work is visible before the client asks what they are paying for.
- Write the runbook for your oldest AI account first, because that is the one only one person understands.
- Book the quarterly drift check for every account now, through to the end of the year.
An AI line sells well and it renews badly if the support is unpriced. The margin is in the scoping and the supervision, exactly as it is with AI answering done in house, and the platform work underneath is the same CRM build you were already doing.
Written April 21, 2026, and kept as written. Platforms, features and policies mentioned here are described as they stood at the time.
Local measurement without third-party cookies
Call tracking, first-party events and honest ranges: how we measure a local business when the tracking is partial, and how to report it without guessing.
ReadTexting that arrives: registration and consent
Why business texts get filtered before anyone sees them, how we register and capture consent properly, and the five follow-up sequences worth running.
ReadReviews under enforcement: fix your flow
The fake reviews rule has been in force since October 2024. Incentives, staff reviews, AI-written replies, and the review flow we build to stay clear.
ReadThis is what we do for clients: For Agencies.
The platforms have moved on since this was written. We will show you where your business stands right now and what to do first.