Probative LabsHowever the AI is billed, check what you got.
Per resolution, per conversation, per session, in assists or credits, or per seat: each model hides waste in a different place. Every one is checked against what happened next, and brought to one number you can compare: what each resolution that held cost.
| How it is billed | Who bills this way | What is checked |
|---|---|---|
| Per resolution | Intercom Fin, Zendesk AI agents | Resolutions billed twice, with no matching conversation, not resolved or outside the period, and those that did not hold |
| Per conversation | Salesforce Agentforce, Jira Service Management's virtual service agent | Billing errors, outcomes the contract does not bill, and how many paid-for conversations ended in a hand-off, an abandonment or a repeat ticket |
| Per session | Freshservice Freddy AI Agent | More sessions billed than your records hold, and a second charge inside one session |
| Usage units | ServiceNow Now Assist (assists), Copilot Studio and Agentforce (credits) | Where the units went by feature, units used in test instances, units spent on conversations that failed, and when the pool runs out |
| Per seat or employee | Moveworks, Microsoft 365 Copilot | Licensed against active users, the cost of seats nobody used, and what each held resolution cost |
Billing models from public pricing pages and guides, October 2026. Your contract may differ.
The same desk, billed five ways
One simulated month (2026-05) of a 40-team service desk, checked under each model. Prices are illustrative, not any vendor's.
| Billed | Price (illustrative) | Paid | Per resolution that held | What the check found | |
|---|---|---|---|---|---|
| Per resolution | $1.00 per claimed resolution | $650 | $1.19 | 110 billed resolutions disputable, and 14% of claimed resolutions did not hold | Report |
| Per conversation | $2.00 per conversation | $1,742 | $3.18 | 37% of the conversations paid for were handed to a person, abandoned, or came back as a ticket | Report |
| Per session | $0.90 per 24-hour session | $786 | $1.43 | 19 more sessions billed than the desk's records hold | Report |
| Usage units | $0.50 per assist, from a 45,000-assist pool | $3,262 | $2.57 | 14% of assists ran outside production, and the pool runs out on 2026-08-05 | Report |
| Per seat or employee | $2 per employee a month, 6,000 employees | $12,230 | $22.30 | 809 of 6,000 employees used it (13%) | Report |
Per resolution that held: what the month cost, divided by the AI's claimed resolutions that held beyond the normal return rate after your agents' own resolutions. It is the number to compare across vendors and billing models.
Now Assist bills in assists. Its reporting counts a walk-away as resolved.
ServiceNow's own guides describe self-solve analytics that mark a conversation resolved when at least one AI asset ran, and a resolution rate that counts no response as positive. A chat the person left without finishing closes as "Closed Complete, No User Response".
The check reads the Interactions table directly and shows, for each way ServiceNow closed a conversation, how often the person needed help again. For assists, it shows where they went by skill, how many ran in development and test instances, how many went to conversations that failed, and when the pool runs out.
Came back within 7 days, by how ServiceNow closed the chat
Simulated desk. After your agents' own resolutions: 13.9%.
| State reason | Counted resolved | Came back |
|---|---|---|
| Topic Complete | 1,335 | 20% |
| Auto Closed | 887 | 16% |
| No User Response | 797 | 56% |
| User Ended | 455 | 29% |
| Left for search | 122 | 38% |
Questions
Our contract has no per-resolution price. Is this still useful?
Yes. Whatever the unit, the question at renewal is what the AI delivered for what it cost. Per-conversation, per-session, usage and seat pricing all come down to the same number here, what each resolution that held cost, with a list of what to raise with the vendor.
What do you need from us?
Your ticket export and the AI's conversations, plus the bill or the usage export. For seats, the fee and the number of seats. Everything runs on exports, read-only.
Can we run it ourselves?
Yes: probative reconcile --model, probative usage and probative seats are in the free engine. See what Cloud adds.
See what your AI bill bought.
Design partnerships are free: you get the check and a readout; I ask for feedback, an anonymized case study and a reference call if it helped.