Google put Gemini 3.7 Flash on the price list this month at $0.75 per million input tokens and $3.75 per million output tokens, held there through the end of 2026. A single customer question and its answer burns a few thousand tokens. That puts the raw model cost of handling one support message at a fraction of a penny. The vendors selling you an AI support agent bill between $0.99 and $2.00 every time their bot closes a ticket. The distance between those two numbers is the entire business model, and most operators sign without ever measuring it.
This is not an argument against buying. Software costs more than tokens for real reasons: retrieval, testing, escalation rules, and a phone number to call when the thing tells a customer something wrong. It is an argument for walking into the demo with your own number already calculated. Build that number this week, before your renewal date or your first sales call. It takes about ninety minutes and it changes the conversation.
Seats Went Away. Outcomes Took Their Place.
The per-seat era is closing. Zendesk restructured in May 2026 and now bills on a three-tier resolution model. An Assisted Escalation, where the AI gathered information but a person closed the ticket, is free. A Contained Resolution, where the customer stopped replying but a verification step did not confirm the fix, is also free. Only a Verified Resolution, confirmed by a separate evaluation model within 72 hours, draws against your allowance and gets billed. Third-party analyses put that rate near $1.20 to $1.50 on committed volume and roughly $2.00 pay-as-you-go. Zendesk does not publish the overage rate itself.
Everyone else picked a different unit. Fin publishes $0.99 per outcome, billed only when the conversation closes without a human. Salesforce Agentforce sells either $2.00 per conversation or Flex Credits at about $0.10 per action, and you pick one per org. Freshdesk charges $0.10 per session. Gorgias runs roughly $0.90 to $1.27 per resolution, and each AI resolution also counts as a helpdesk ticket against your plan tier. Same category, five different meters.
Here is the part that matters for your invoice. The word “resolution” is a contract term, not a fact about the world. One vendor counts a closed conversation. Another counts a verified one. A third counts every session including the failures. Ask for the written definition before you ask for the price. A vendor who will not put the definition in the order form is telling you something.
Run the Math at 400 Messages, Not 25,000
Every vendor pricing table on the internet models a 20 to 50 agent support team handling 5,000 to 25,000 conversations a month. That is not a six-truck plumbing company or a three-location dental group. Your numbers are smaller and your break-even sits in a completely different place, so the vendor’s total-cost table tells you nothing useful.
Work a real example. Say 400 inbound customer messages a month across your web form, your business texts, and transcribed after-hours voicemail. Assume the agent closes 60 percent without a person, which is a fair planning number and below what most vendors advertise. That is 240 resolutions and 160 escalations.
- Per outcome at $0.99: about $238 a month. You pay nothing on the 160 that failed.
- Per conversation at $2.00: $800 a month. You pay full price on all 400, including every one the bot fumbled.
- Per session at $0.10: looks like $40, until you learn one customer issue often spans three sessions. Budget $120 and get the session definition in writing.
- Per verified resolution at $1.50, plus platform: $360 in resolutions, plus a $115 seat and a $50 copilot add-on for the one person who works the queue. Call it $525.
Now the other side of the ledger. If a part-time front desk person costs you $2,400 a month and the agent removes 40 percent of their message load, your break-even is roughly $960. Every option above clears that. If instead the owner answers texts at 9pm for free, your break-even is close to zero and the honest answer might be no. Do this subtraction before the demo, not after the second invoice.
Four Questions That Decide the Bill
Print these and take them into the call. They are the four variables that separate a quoted price from an actual one.
- What exactly counts as a billable event? Get the definition in the order form, not the marketing page. Ask specifically how an escalation is treated, and what happens when a customer replies four days later.
- What is the free allowance and does it reset? Zendesk Suite plans include only 5 to 15 free verified resolutions per agent per month depending on tier. That allowance disappears fast at any real volume.
- Is there a hard spend cap? Zendesk moved to automatic overage billing in January 2026, replacing an opt-in model, and teams on older contracts found out on the invoice. Fin publishes spend caps. Ask for one in writing or assume there is not one.
- Who owns the failure? When the agent quotes a wrong price or promises a same-day appointment you cannot staff, what is the remedy? For a local business, one bad answer to a regular customer costs more than the month of software.
That fourth question is the one operators skip and the one that bites. Google published a design guide this month showing a support agent talked into issuing an excessive refund, and its answer was to put hard limits outside the model: a maximum dollar amount, fixed destinations, an approval gate. A prompt that says “never refund more than the order total” is a suggestion. A control that makes the larger refund technically impossible is a policy. Ask which one your vendor built.
When Assembling Beats Buying
The alternative got much stronger this year. Xero showed agentic workflows at Xerocon Denver that flag unreconciled transactions, duplicates, and missing documents, then walk the user through resolving them, and it is expanding integrations so you can query live Xero data from Microsoft 365, Claude, or ChatGPT. Anthropic shipped Claude for Small Business in May with connectors into QuickBooks, PayPal, HubSpot, Canva, Docusign, and Google Workspace. Bardeen launched an agent you teach by performing the task in front of it, then approving the plan it writes. Meta AI can now read your Facebook, Instagram, and Ads data alongside Google Workspace and build the report.
None of that carries a per-resolution meter. It also carries no service level, no vendor to call at 6pm, and no one but you when the output is wrong. So split the decision on risk rather than price. Buy the metered product when the queue faces customers and a mistake costs you the relationship. Assemble your own when the work is internal and reversible: the weekly exception list out of your books, the follow-up drafts that a person still sends, the data cleanup nobody has time for. Start where the error is visible and easy to undo.
The Signal to Watch Before October
Watch whether the metered vendors start publishing their overage rates. Right now the per-unit price is the marketing number and the overage is the real one, which is why only some of them print it. If disclosure becomes standard by the fourth quarter, the market is maturing and you can negotiate on the posted number. If it does not, assume every quote you see is the floor and price your contract against the ceiling instead.
The second signal is closer to home. Pull your own message log for the last 30 days and count three things: total inbound, how many were answerable from information already on your website, and how many needed you specifically. That third number is the ceiling on what any agent can take off your plate, and it is usually smaller than the demo implies. Run the count Tuesday morning. It costs nothing and it is the only figure in this article that is actually about your business.
Atlas Unchained builds the websites, automations, and stacks that local operators run on, and we price the work, not the tickets. If you want the week’s operator news filtered down to what changes your Tuesday, subscribe to The AU Brief below.
