The monthly fee on the quote is one of four cost lines. Here is the full structure — including the two that show up on no invoice — and a worksheet to price it from your own inputs.
A five-person HVAC company was quoted one monthly number for an AI receptionist and budgeted for exactly that number. The first month's reality was four numbers: the quoted platform fee, a per-conversation overage after a promo mailer doubled inbound volume, eleven hours of office-manager time correcting answers built from a price list that had changed two weeks earlier, and one write-off after the agent quoted an after-hours surcharge that no longer existed. The fee they had planned for was the smallest of the four.
Nothing about that purchase was necessarily wrong. It was priced wrong, which is a different problem — and a fixable one, before you sign anything.
The fee you see and the four costs you do not
An agent that talks to customers has four cost lines. Two appear on invoices. Two appear on your calendar and in your mistakes column, which is why they get forgotten.
- Build and setup. The one-time work: deciding what the agent handles, writing the facts file, configuring behavior, and testing against real questions before launch. If you build it yourself, this is your hours. If you pay someone, get it as a fixed line on the quote — not the word "included."
- Per-conversation model usage. Every conversation consumes model capacity, and most platforms pass some of that through — as a per-conversation rate, a metered allowance, or an overage above a plan cap. This line scales with your volume, not with your plan tier, so a good month of marketing quietly raises it.
- Maintenance time. Someone has to keep the facts file true — hours, prices, promotions, service areas — and read a sample of transcripts every week. Budget this as real hours, because it is real hours: in my experience one to three hours a week for a small business, more in the weeks right after launch.
- The failure cost of one bad promise. Sooner or later the agent will state something wrong with total confidence — a price, a date, a policy. The cost is whatever it takes to make it right: an honored bad quote, a refund, a redo, or a review you cannot delete.
| Cost line | When it hits | Who notices it | What it looks like in practice |
|---|---|---|---|
| Build and setup | once, before launch | nobody, if you self-build | your hours, or a one-time invoice |
| Model usage | every conversation | nobody until the invoice scales | grows with volume, ignores your plan tier |
| Maintenance | weekly, indefinitely | nobody, which is the problem | facts-file edits plus transcript review |
| Failure cost | unpredictably | you, loudly | honored bad quote, refund, public review |
A cost worksheet you fill in yourself
Every figure below is a placeholder. Replace all of them with your own numbers, and check current pricing with any vendor you evaluate rather than relying on figures from any article, this one included. The worksheet is the structure; your inputs are the answer.
AI AGENT COST WORKSHEET — replace every bracket; figures below are placeholders
1. Setup (one-time)
your build hours ............ [x] h x $[rate] = $[a]
or paid setup ............... $[b] — get it in writing
2. Model usage (monthly)
conversations per month ..... [n]
cost per conversation ....... $[c] (ask the vendor how they meter it)
monthly usage ............... [n] x [c] = $[d]
overage buffer (+25%) ....... $[e]
3. Maintenance (monthly)
facts-file upkeep ........... [x] h x $[rate] = $[f]
transcript review ........... [x] h x $[rate] = $[g]
4. Failure allowance (monthly)
expected mistakes ........... [n] x $[average cost to fix] = $[h]
First year = [a or b] + 12 x ([d] + [e] + [f] + [g] + [h])
Two notes on the inputs. The overage buffer exists because volume is the input you control least — run the worksheet twice, once at your average month and once at double volume, and see how the total moves. The failure allowance is not pessimism; it is an honest estimate of how often a confident wrong answer reaches a customer and what one instance costs you.
Why the cheapest build is often the most expensive to run
The cheap build is recognizable: a generic system prompt, no facts-file discipline, and an agent instructed to be helpful above all else. It launches fast and costs little. Then it improvises.
An agent with loose instructions answers everything, including the questions it should escalate. Every improvised answer is a transcript you must review, a correction you must write, and occasionally a promise you must honor. The expensive-to-run pattern is not high usage — it is high improvisation. A build that costs more upfront, with a tight facts file, a narrow scope, and clear escalation rules, pushes cost out of the recurring lines (review, correction, failure) and into the one-time line. That is almost always the right trade, because the one-time line happens once.
[!TIP]
Ask any vendor or builder one question before you compare prices: "what does the agent do when it doesn't know?" The correct answer is a specific escalation behavior — take a message, book a callback, hand off to a human. "It uses the model's general knowledge" is the sentence that turns the cheapest quote into the most expensive month.
Questions to ask any vendor quoting a monthly fee
Get the answers in writing, ideally written into the quote itself. If a question cannot be answered in one sentence, that is information too.
- How do you define a "conversation" for billing? Turns, sessions, or unique contacts? This one definition can move the invoice more than the plan tier does.
- What happens above the included cap? A per-conversation rate, throttled service, or a hard stop?
- Who updates the facts file, and what is the turnaround? Same-day matters when your hours or prices change on a Tuesday.
- Can I export raw transcripts? In a usable format, without requesting permission per export. You need them for review, and you will need them if you leave.
- What is the exit? Notice period, data export, and what happens to the configured behavior if you go.
Four mistakes that inflate the bill
- Comparing builds on setup price only. The setup line is paid once; the usage, maintenance and failure lines are paid monthly. A quote that looks 40% cheaper on setup can be far dearer across year one.
- Letting the agent answer everything. Scope the agent to the questions that repeat and escalate the rest. Breadth is where improvisation — and therefore failure cost — lives.
- Skipping the transcript review habit. Failures that nobody reads become patterns, and patterns become public reviews. An hour a week of reading is cheaper than one 1-star review describing a promise the agent never should have made.
- Budgeting for month one. Month one is launch volume with your full attention on it. Price year one: twelve months of usage and maintenance, plus one volume spike you did not schedule.
What to do today
- [ ] Fill in the worksheet with your real numbers — once for an average month, once for a double-volume month.
- [ ] Get the definition of a "conversation" and the overage behavior in writing from every vendor you are talking to.
- [ ] Name the person who owns the facts file and block their weekly review hour on the calendar.
- [ ] Write your escalation rule in one sentence: exactly what the agent does when it does not know.
The fee on the quote is a real number, but it is one of four. Price all four before you sign, and the purchase becomes arithmetic instead of hope.
This guide lives at AgentFast, where you'll also find a free audit tool that checks your own agent's instructions against these failure modes. Posted here with a canonical URL pointing at the original.
