The Hidden Cost of Zendesk's 72-Hour “Automated Resolution”

A customer opens a chat, gets a reply from the AI, and closes the tab without saying another word. Three days later, that conversation quietly becomes a billed resolution on your invoice. The customer never confirmed anything was solved. The meter counted it anyway.

This is not a bug. It is how Zendesk defines an automated resolution, and once you understand the mechanic, a lot of surprising line items on the bill start to make sense. Here is exactly how the count works.

What Zendesk actually counts, step by step

An automated resolution is the unit Zendesk bills for AI agent usage. You pay only when the AI resolves a request without a human agent stepping in. So far, reasonable. The detail lives in how “resolved” gets decided.

For email and web form conversations, a resolution is counted after 72 hours of inactivity, and only if all of the following are true:

• The AI agent gave a generative reply to the customer's question.
• The customer gave positive feedback or gave no feedback at all.
• No human agent responded to the ticket.
• A separate large language model reviewed the transcript and confirmed the reply was relevant.

That last step is the one people miss. Zendesk runs a secondary LLM check over the conversation text to decide whether the customer's request was genuinely satisfied. Conversations that fail this check are not billed. It is a real quality gate, and it is more than most competitors do.

Messaging works a little differently. For advanced messaging agents the conversation is evaluated 2 hours after the first message by default, configurable up to a 72-hour maximum, so the flat 72-hour figure is really an email and legacy-messaging rule rather than a universal one.

The quirk that inflates the count

Read the four conditions again and notice the second one: positive feedback or no feedback.

Silence counts as success. A customer who abandons the chat out of frustration looks identical, to the meter, to a customer who got exactly what they needed and walked away happy. As long as the AI's last message reads as relevant to the verifier, both get counted as resolutions.

That is the gap between a billed resolution and a real one. Buyers have reported this pushes their counted resolution rate above the satisfaction they actually measure, with estimates of the overcount landing somewhere in the 15 to 30% range depending on channel and traffic. Treat that band as buyer-reported rather than official, because Zendesk does not publish an overcount figure and the true number varies a lot by account.

The point is not that Zendesk is being dishonest. The definition is documented in plain sight. The point is that “resolution rate” and “resolved to the customer's satisfaction” are two different numbers, and only one of them shows up on the invoice.

What the May 2026 update changed

Zendesk tightened this in May 2026, and the change is worth crediting because it cuts the other way.

The old model had a single automated-resolution bucket. The new one splits automation into three tiers:

• Assisted escalation: the AI helped, but a human finished the job. Free. Does not draw down your allowance.
• Contained resolution: the AI handled the interaction to completion, but it was not independently verified. Also free now.
• Verified resolution: the AI resolved it and the secondary LLM confirmed the outcome inside the 72-hour window. This is the only tier that is billed.

The effect is that unverified closes no longer cost you money. That genuinely narrows the overcount problem compared with the older behavior, where anything the AI touched and the customer did not object to could count. The verification bar is now the gate, not just customer silence.

It does not fully close the gap, though. A customer who gives up but whose final AI reply still scans as relevant can pass verification and be billed. The mechanic rewards conversations that look resolved on the transcript, which is not always the same as conversations that were resolved in the customer's head.

Why the counting method hits your budget twice

The 72-hour window has a second-order effect that catches finance teams off guard.

Because resolutions are only confirmed three days after the fact, your billed usage for a given day is not knowable until 72 hours later. During a traffic spike, a sale, an outage, a viral complaint, the meter keeps counting, and since January 2026 overages auto-bill by default unless an admin has switched on the pause setting. The safety valve exists, but it is off out of the box, so a busy month can produce an invoice nobody signed off on. The full three-layer cost picture, base seats plus the Copilot add-on plus per-resolution fees, is laid out in the Zendesk pricing breakdown.

There is a reporting effect too. Because the counted resolution rate can run ahead of real satisfaction, a leadership dashboard showing a healthy automation percentage might be quietly overstating how many customers actually left happy. If you report AR% upward, it helps to footnote what it does and does not measure.

How to reconcile the meter against reality

You do not have to take the counted number at face value. A few habits keep the resolution report honest.

Sample the transcripts. Pull a random set of billed resolutions each month and read the last few messages. You are looking for the pattern where the AI answered, the customer went quiet, and nothing actually got done. If that shows up often, your real rate is lower than your billed rate, and you now know by roughly how much.

Cross-check against a satisfaction signal you control, like a post-chat CSAT prompt or a reopened-ticket count. If billed resolutions climbed but CSAT did not, the meter is running ahead of the experience. Zendesk's own AI agent reporting now separates Contained from Verified outcomes, which helps, but the reconciliation is still yours to do. The goal is not to catch anyone out. It is to make sure the number in the board deck means what the room thinks it means.

How a cleaner resolution definition looks

The fix is not to distrust every automated resolution. It is to tie the count to something the customer confirmed, and to price it so you are not penalized for automating well.

Robylon takes a narrower line on what counts. Its email support agent reports a 60 to 80% autonomous resolution rate that is validated against your own historical tickets during onboarding, so the number is measured on your data before you commit rather than inferred from silence after the fact. And because pricing is credits-based rather than per verified resolution, a busy month does not trigger an uncapped overage, and automating more does not mechanically raise your bill. For the head-to-head on cost and resolution quality, the Zendesk AI versus Robylon comparison puts the two side by side.

None of this means Zendesk's definition is wrong. It means you should read your resolution report knowing what a resolution is: an AI reply, three days of silence, and a second AI agreeing it looked relevant. Useful. Just not the same thing as a happy customer.

Want a resolution number you can trust because it was measured on your tickets, not inferred from silence? Robylon resolves 60 to 80% of customer emails autonomously with AI agents that take action across Zendesk, Shopify, Stripe, and 60+ other integrations. See how Robylon resolves email