AI support cost per resolved ticket needs a mature issue cohort, not a count of closed conversations

Divide AI spend by the conversations a vendor marked resolved and you get a number on day one. It is a narrow number, not the full workflow cost per verified issue. It divides by an event that can be reversed next week, it counts conversations rather than issues, and it leaves out the people and the platform that finished the work the AI did not.

This article proposes a reporting policy for AI support cost per resolved ticket. Report it for a fixed cohort of issues, as of a dated cutoff. The numerator is everything the cohort cost by the cutoff, resolved or not. The denominator is the issues in that cohort whose follow-up window has closed with verified resolution evidence. When that denominator is zero, the figure is undefined, and the report should say so. While some issues are still open, the figure is provisional and labeled that way.

Below is a small synthetic ledger at two dates, with the arithmetic. Every figure in it is invented for illustration.

Four different things get called "resolved"#

1. A billed conversation outcome. The vendor's unit. Intercom's Fin outcomes page defines a resolution as the customer confirming the last AI answer (confirmed) or leaving without asking for more help (assumed). Read on 12 November 2026, it lists $0.99 per resolution and per configured Procedure handoff for Fin over chat and email, charges at most once per conversation, and does not bill escalations from Fin's default behavior or workspace rules. If the customer returns to the same conversation, even in a later billing period, the resolution is deducted. The page does not cover a customer who comes back in a new conversation or on another channel.
2. A closed ticket. A helpdesk state. Tickets close for many reasons, including inactivity.
3. An assumed resolution. Silence after an answer: a guess about the customer, not evidence about the issue.
4. A verified resolution. Your own policy, applied to evidence you can check later. Under this article's policy, it is the only kind that enters the denominator.

The units differ too. One conversation can carry two issues; one issue can span two conversations on two channels. The ledger has one row per issue, listing the conversations it draws on.

The policy: evidence, window, cutoff#

For this example:

• Verified evidence: the system of record shows the requested change (refund issued, setting changed, subscription cancelled), or, for an information request, the customer explicitly confirms. Under this policy, silence alone does not qualify. Treating "Ok thanks" as confirmation is itself a choice; some teams will want system evidence for that too.
• Follow-up window: 14 days from the issue's last answer or resolution event. An issue that is still open has no window yet.
• Observation cutoff: the as-of date of the report. Costs recorded by the cutoff count.
• Status: an issue whose window ended before the cutoff is mature. It is verified if it has qualifying evidence and no linked contact arrived inside the window, and unverified otherwise. An issue whose window has not ended is pending; an issue with no resolution event is open.

Fourteen days is a choice for this example, not a recommendation. Set it per issue type from how long your customers take to come back: a billing error may only resurface at the next invoice. A longer window allows more time to observe repeat contacts, but delays reporting and does not remove linkage or verification errors. Print the window next to the number.

The cohort and the cost rules#

The cohort is every issue whose first contact arrived on the AI-assisted chat entry point between 1 and 7 October 2026: eight issues in seven chat conversations. Each conversation's assumed Fin billing category is in the first table.

Cost rules, applied identically to every row:

• AI cost: the vendor outcome fee ($0.99 per billed outcome, from the page above) plus the team's own model calls on that conversation, an intent classifier and a handoff summary priced from its own logs. Those own calls are separate attributable spend, not a surcharge on Fin's answer. A deducted resolution removes the fee.
• Human cost: minutes from the helpdesk time log at a loaded $0.50 a minute. One source per minute of work: if review minutes are in the time log, do not also apply a per-approval cost from another system.
• Platform cost: an assumed $600 monthly AI tooling cost over 1,200 issues first contacted in October, so $0.50 per issue. Allocate per issue, not per resolution: per-resolution allocation makes unresolved issues look free and ties the allocation to the outcome being measured.
• Shared conversations: C-206 carries two issues. Its one vendor fee and its own model calls ($1.10 together) are split equally; human minutes go to the issue the handoff note names.
• Unknown is not zero. If a channel's human minutes are not tracked, mark that cost unknown and do not publish a figure that depends on it.

Snapshot A, as of 12 October: undefined#

The I-102 row assumes Intercom's Human in the Loop, a beta feature on chat and email in which a teammate approves or answers and Fin sends the reply and owns the resolution. The example bills it as a resolution. If your reviewed replies run through a separate workflow with no vendor fee, drop the fee and say so.

Cohort cost so far: $34.56, which is $7.56 AI, 46 minutes ($23.00) human and $4.00 platform. Mature verified issues: zero. Cost per verified resolution as of 12 October is undefined.

The vendor view on the same day: four resolutions (C-201, C-202, C-204, C-205) and three Procedure handoffs, so seven billed outcomes and $6.93 in fees. Two shortcuts can be computed from it, and each needs its definition printed beside it. Cohort AI cost over resolutions only: $7.56 / 4 = $1.89. Cohort AI cost over all billed outcomes: $7.56 / 7 = $1.08. Both leave out human and platform cost and count conversations. Neither is an early estimate of the cohort figure.

Snapshot B, as of 30 October: five verified, one open#

Changes since snapshot A:

• I-101 reopened. The customer came back to C-201 on 14 October. A person spent 12 minutes, changed the setting, and the customer confirmed. Under Intercom's rule the resolution fee is deducted, so AI cost drops to $0.11. New window ends 28 October: mature, verified.
• I-102, I-103, I-105, I-107 passed their windows with no linked contact: mature, verified.
• I-106 passed its window in silence: mature, unverified.
• I-108 is still open, now with 20 human minutes.
• I-104 is mature; whether E-301 is the same issue is unknown.

Numerator, all eight issues, costs recorded by 30 October: AI $6.57 + human 70 minutes ($35.00) + platform $4.00 = $45.57.

Denominator: five mature verified issues (I-101, I-102, I-103, I-105, I-107).

Cost per verified resolution as of 30 October, E-301 not linked: $45.57 / 5 = $9.11. It is provisional: I-108 is open, so its cost will grow and it may still become verified. For a final figure, fix an observation horizon for the cohort in advance and report anything still open at the horizon as open.

The vendor shortcuts moved too. Three resolutions remain after the deduction, so $6.57 / 3 = $2.19 over resolutions, or $6.57 / 6 = $1.095 over billed outcomes.

Snapshot A is not wrong and should not be overwritten. It answered "what is known on 12 October". Keep each report with its as-of date and treat a later one as a restatement. Under this policy, a contact that arrives after an issue's window is a new issue, which bounds restatements.

The uncertain link is a range, not a guess#

E-301 came from the same account three days after I-104, about an invoice, with no invoice number. A person spent 14 minutes on it, the corrected invoice is in the billing system, and nothing followed within 14 days. Link evidence, strongest first: same conversation, then same account plus the same order or invoice ID, then same account plus a similar topic. Only the last is present, and an account match does not prove the same person or the same issue.

Compute both sides and state the assumption:

• Not linked: I-104 stays outside the denominator, and the figure is $9.11. E-301's first contact is email, so it belongs to an email-entry cohort for 8 to 14 October, not to this chat cohort.
• Linked: E-301's 14 minutes ($7.00) join I-104, which becomes verified through a person on email. ($45.57 + $7.00) / 6 = $52.57 / 6 = $8.76. No extra platform allocation, because there is no extra issue.

Report "$8.76 to $9.11 as of 30 October, depending on one unconfirmed cross-channel link". Here linking lowered the figure, because it added a verified issue along with the cost. Neither side reliably flatters the AI, which is one more reason not to pick one quietly. Either way, the vendor still counts C-204 as a resolution.

A stable customer or account mapping across channels helps identify candidate repeats; a pseudonymous end-user id, as in per-user AI token quotas, is one option. If no reliable mapping is available, report that limitation and do not claim complete cross-channel resolution counts.

Split by path without losing what failed#

"$1.56 per autonomous resolution" is true of one row and misleading on its own. The not-verified line is 31% of cohort cost ($14.16 / $45.57). Two of its three issues began with AI answers, but only C-204 is a vendor resolution under the example's classification; C-206's outcome is a Procedure handoff. Keep issue paths and conversation outcomes in separate columns. A handoff is a path, not a resolution: I-108 was handed off and is still open.

Classify by the path the issue took by the cutoff. I-101 started as an autonomous resolution and ended as human work. A successful tool call does not settle the path either: the refund API returning success is not the customer's problem going away, the gap described in a tool call can succeed and still produce the wrong outcome.

Compare with a baseline built the same way#

A human-only baseline is comparable only as a ledger with the same issue definition, verification rule, window and cutoff; the same cost scope, including its own helpdesk tooling at the same loaded rate; and a comparable intent mix, compared intent by intent or reweighted. Seasonality, product changes, staffing and pricing can differ between periods, so a before-and-after difference alone does not establish that AI caused the savings. In our view, a concurrent holdout, routing a random share of eligible issues to people in the same weeks, is the comparison worth reporting. Eight issues, like this example, support no conclusion at all.

Where Worksi fits, and where it does not#

Worksi's ROI view prices every model call that passes through the gateway, records outcomes you send from code or the API, and compares them with baselines from before AI. Its total cost is AI spend plus platform cost prorated to the period plus human review cost, where human review is runs that needed an approval or an operator action multiplied by your cost per review. A sensible use: record an outcome as successful only after your ledger marks the issue mature and verified.

It does not build the cohort ledger, link conversations across channels, verify resolutions, or capture support minutes outside approvals and operator actions. If your support AI is a vendor's hosted agent, a gateway sees only your own calls, like the classifier and summary here, and the vendor invoice stays the fee input. The Worksi SDK is in private preview and distributed on request; any OpenAI or Anthropic SDK works with the gateway without it.