Last month we found Acme Co was billed $14k by a competitor for AI resolutions. 38% were tickets where their human team had to step in and fix. We refunded $5,320.
That refund was not a gesture. It was the system doing what we built it to do. Acme had imported a quarter of billing history, connected Slack, Zendesk, and their help center, and asked us to replay the last invoice against Beamdesk rules. The competitor invoice said 18,421 AI resolutions. Our audit log said 11,425 clean AI handoffs, 4,283 human-repaired outcomes, 1,911 negative follow-ups, and 802 tickets with CSAT below the workspace threshold. The clean number was still valuable. The padded number was not honest.
This is the pricing rule we use: seat fees cover the product, included resolutions cover ordinary volume, and metered AI resolution overage is billable only when the AI actually carried the ticket across the finish line. If a human edited the answer before send, if a human had to follow up after send because the first answer missed, or if the customer rated the outcome below your threshold, the overage line for that ticket is $0.
What We Count
We count a resolution when Beamdesk receives a customer message, routes it through the model roster, retrieves the necessary knowledge or tool output, sends the reply, and the conversation reaches a resolved state without a qualifying human correction. The result has to survive the operational checks that a support lead would use outside an invoice: no agent rewrite, no agent rescue, no reopened ticket caused by the AI answer, and no customer signal saying the resolution was bad.
The invoice row is generated from the same event stream you see in the audit log. We do not keep a private billing counter. A ticket earns the billable flag when the resolution event is written and keeps it only while later events do not disqualify it. That means the charge can move from billable to zero after the fact. If the customer replies six hours later with "this did not answer my question" and a human agent fixes the ticket, the billing status changes. If CSAT lands two days later below your configured threshold, the billing status changes. The invoice follows the truth, not the first optimistic state transition.
ticket BD-84319
ai_reply_sent 2026-04-18 09:42:11
resolved_by_ai true
human_edit_before_send false
human_rescue_after_send false
negative_follow_up false
csat_score 5
billing_outcome metered_resolution
overage_charge $0.20The adjacent ticket may look similar in a dashboard but it invoices differently. If an agent opens the draft, changes the refund amount, and sends it, Beamdesk records a human-edited resolution. The AI may have saved time by gathering context and writing a usable first draft, but it did not resolve the ticket. We do not call that an AI resolution.
What We Do Not Count
We do not count drafts. We do not count classifications. We do not count macro suggestions, sentiment labels, intent tags, duplicate detection, or internal notes. Those are product features, and they are covered by seats and normal plan limits. We also do not count a ticket as billable just because the AI was the last named actor before a status changed to resolved. That shortcut is exactly where invoices get padded.
The common gray area is the "agent assist" case. A model reads the thread, finds the order, proposes a refund, and writes a reply. The human agent changes "refund issued" to "store credit issued" because the customer used a final-sale discount. That is useful AI work. It is not autonomous resolution. Beamdesk keeps the audit trail, credits the assist in productivity reporting, and bills $0 in metered resolution overage.
Another gray area is the polite but wrong answer. The AI sends a confident message, the customer replies "no, I already tried that," and a human agent closes the case. Some vendors still count the first answer as resolved because an automation technically responded. We count the customer correction as a negative follow-up. The event removes the charge and places the conversation in the exception set for QA.
How The Audit Log Works
Every auto-resolution carries a signed outcome record. The record includes the source messages, model route, retrieved knowledge chunks, tool calls, final reply, confidence score, Watchtower QA score, customer follow-up window, CSAT result if available, and all human touches. The important part is not that the log exists. The important part is that billing reads from it.
Human touch is explicit. Opening a ticket is not enough to zero a charge, because managers review closed conversations all day. A qualifying touch is an edit before send, a public follow-up that changes or repairs the AI answer, a manual status reversal caused by the AI answer, or a CSAT result under the workspace threshold. Teams can set the threshold. Most start at 4 out of 5 for transactional support and 80 out of 100 for enterprise workflows. When the threshold changes, new tickets use the new rule and old audit rows remain immutable.
billing rule:
billable = resolved_by_ai
and not human_edit_before_send
and not human_rescue_after_send
and not negative_follow_up
and csat_score >= workspace.csat_billing_thresholdYou can click from invoice to ticket to evidence. If finance asks why a line item dropped from 8,200 outcomes to 6,740 outcomes, support ops can export the disqualified rows with reason codes. No one has to argue over screenshots from an admin panel.
Why Most Vendors Do Not Do This
Outcome pricing is attractive because it sounds aligned: pay when the AI solves work. The hard part is admitting that many "AI resolutions" are mixed outcomes. Support teams know this. Agents clean up half-correct answers, add missing policy caveats, apologize for tool mistakes, and reopen tickets after bad deflections. A vendor that sells per resolution has a direct incentive to blur those assists into resolutions.
There is also a product reason. If the billing system only sees conversation status, it cannot tell the difference between an AI handoff and a human rescue. Fixing that requires a durable event model, editor instrumentation, CSAT ingestion, exception workflows, and invoice logic that can revise a charge after customer feedback arrives. It is much easier to say "resolved by AI" when the bot sent a message and the ticket was closed.
We think that is a tax on honest support teams. The better the human team is at repairing AI misses, the more the vendor can overclaim. The invoice improves while the customer experience quietly depends on unpaid human cleanup.
The Math On A Typical Workspace
Take a 20-seat support team on a Pro-style plan with 200 included resolutions per seat and $0.20 metered overage. They receive 14,000 tickets in a month. Beamdesk attempts automation on 8,900. Of those, 6,700 are clean AI resolutions, 1,100 are human-edited before send, 620 receive a negative follow-up, and 480 fall below the CSAT threshold. The plan includes 4,000 resolutions, so only 2,700 overage resolutions are billable.
14,000 inbound tickets
8,900 AI attempts
6,700 clean AI resolutions
-4,000 included plan resolutions
2,700 billable overage
2,700 x $0.20 = $540A loose outcome vendor could count all 8,900 attempts, subtract the same 4,000 included outcomes, and charge 4,900 overages. At $0.20, that is $980. If the vendor charges $0.30 or $0.40 per claimed resolution, the gap widens quickly. The customer is not buying more automation. They are buying a more generous definition of the word "resolved."
Honest pricing also changes internal behavior. Agents stop worrying that correcting an AI answer will punish the company twice, once in labor and again on the invoice. Support leads get a clean split between automation savings and quality debt. Finance can forecast overage from verified outcomes instead of marketing claims. Product teams can see which knowledge gaps create expensive human rescues.
The Rule Is Simple On Purpose
We do not want a philosophical billing meeting every month. If the AI resolved it cleanly, overage can be billable. If a human had to make the answer true, complete, or acceptable, overage is $0. If the customer says the outcome missed the mark, overage is $0. The audit log keeps us from asking anyone to trust our interpretation.
AI support will only be trusted when the accounting is as inspectable as the answer. That is why Beamdesk treats billing as a product surface, not an invoice appendix. See your own audit log: /settings/billing/audit-log