TL;DR
Not the model price — the real number. Subscriptions, usage, integration, and maintenance, with a worked example for a 300-lead-per-month service business.
→ See how this applies to your business (free 30-min call)Every quote you'll get for "AI" is quoting one of three different things, and the gap between them is an order of magnitude. That's why the answers people give each other are so wildly inconsistent — they're not disagreeing, they're describing different layers.
The three layers are subscriptions, usage, and build plus maintenance. Price them separately or you will be surprised, and the surprise is always in the same direction.
Layer One: Subscriptions
The predictable, per-month, per-seat costs. These are the ones people quote because they're easy to find.
For a ten-person business giving everyone an assistant and running one automation platform, this layer is realistically $500 to $1,200 per month. It's the layer people budget for, and it's usually not the biggest one.
Layer Two: Usage
This is where the surprises live, because it scales with activity rather than headcount.
Text model calls are priced per million tokens, split between input and output, and the spread between model tiers is roughly twenty-fold. A small fast model might cost cents per million input tokens; a frontier model costs dollars. In practical terms:
The multiplier people forget is retries and reflection loops. An agent that reconsiders its work three times costs four times as much as one that doesn't. Uncapped loops are the single most common source of a shocking bill.
Voice is priced per minute and stacks several components: speech recognition, the model, speech synthesis, and telephony. All-in, a conversational voice agent typically lands somewhere between roughly $0.07 and $0.25 per minute depending on the model tier and voice quality. A three-minute qualifying call is therefore in the range of twenty to seventy-five cents.
Messaging is per segment: SMS is typically under a cent per segment in the US, plus carrier fees and one-time A2P registration. Email is fractions of a cent per send. These feel trivial until you're sending fifty thousand of them.
The rule that governs this layer: cost scales with conversation volume, not with company size. A five-person business handling four hundred inbound leads a month spends more on usage than a fifty-person business handling forty.
Subscriptions are what you budget. Usage is what surprises you. Maintenance is what kills the project in year two.
Layer Three: Build and Maintenance
The largest and least-quoted layer.
Build. Connecting AI to your actual systems — CRM, calendar, phone, billing — and making it survive real inputs. Realistic ranges:
If you're building internally, translate that to engineering time: a competent developer at a fully loaded cost of $150 to $250 per hour, and multiply your honest estimate by two.
Maintenance. The line item nobody includes and everybody pays. Models get deprecated. Vendors change APIs. Your business process changes and the automation doesn't. Prompts drift out of alignment with reality. Budget fifteen to twenty-five percent of build cost annually, or an ongoing management fee if someone else operates it.
Systems without an owner degrade faster than any other software category, because the ground underneath them moves. This is not a hypothetical risk; it's the normal case.
A Worked Example
A home services business: 300 inbound leads per month, average job value $4,200, currently closing about 12 percent because response time averages four hours and nights and weekends go unanswered.
What they run:
Monthly total: roughly $2,035. Build cost, amortized: $18,000 over the first year adds about $1,500 per month in year one.
What it has to produce. At $4,200 per job, the system needs to add about one job per month to cover the ongoing cost, and roughly two in year one including build. What actually happens when response time goes from four hours to ninety seconds and every lead gets qualified is a close-rate improvement of several points — on 300 leads, that's a materially larger number than two jobs.
That's the calculation that matters. Not "what does AI cost," but "what does it cost relative to the revenue currently leaking out of the four-hour response gap."
Where Costs Blow Up
Five failure modes, in rough order of frequency:
Uncapped agent loops. Always set a maximum step count and a per-request budget. Always.
Sending full context on every call. Passing an entire conversation history to every request multiplies input tokens. Summarize and truncate deliberately.
Using a frontier model for classification. A large fraction of agent calls are simple routing decisions that a small, cheap model handles identically. Tier your model usage by task.
Per-seat sprawl. Six AI subscriptions with overlapping capability, three of them barely used. Audit quarterly.
No usage alerting. Set a spend alert on every provider before you launch anything, not after the first surprising invoice.
How to Price a Pilot Instead of a Program
The mistake that makes AI look expensive is committing to a full build before you know the return. Structure it as a pilot and the risk profile changes completely.
A good pilot has four properties:
One process, narrowly scoped. Inbound lead response for one service line, or one lead source. Not the whole funnel.
A defined success metric agreed in advance. Time to first contact, appointments booked, or qualification rate — a single number, measured the same way before and after.
A fixed cost and a fixed end date. Typically $3,000 to $8,000 over 30 to 60 days for a scoped pilot, including usage. If nobody will quote fixed, they don't know what they're building.
A baseline measured first. You cannot evaluate the pilot without knowing what the current process produces. Spend the first week measuring, not building.
The economics of a pilot are usually obvious rather than ambiguous. On 300 leads a month at a $4,200 average job, a two-point close-rate improvement is roughly $25,000 in additional revenue. A pilot that costs $6,000 either finds that or it doesn't, and you'll know in eight weeks rather than after a year-long contract.
The corollary: be suspicious of anyone who won't pilot. A vendor confident in the outcome will happily prove it on a small scope first, because the pilot is their best sales tool.
What You Should Actually Budget
For a small-to-mid local service business getting serious about this:
If a quote comes in dramatically below that, find out which layer they've excluded. It's usually maintenance, and it's usually the one that matters.
The Comparison That Frames It
The honest benchmark isn't zero. It's the alternative: a full-time person answering phones and chasing leads costs a fully loaded $45,000 to $65,000 per year, works forty hours a week, and doesn't answer at 9pm on a Saturday when a good share of high-intent inbound arrives.
The AI system costs less, covers all 168 hours, and — this is the part that actually matters — responds in ninety seconds every single time.
If you want a specific number for your business rather than a range, [book a free strategy call](/book) and we'll price the build and the running cost against your actual lead volume.
Free Weekly Briefing
One AI Marketing Tactic.
Every Tuesday. Free.
What's actually working across our client accounts right now — ROAS moves, follow-up sequences, creative angles. The stuff that isn't in any blog post yet.
No spam. Unsubscribe anytime. 1,200+ business owners already in.