# Zendesk AI: How to Measure Deflection and Resolutions

*How-to — 2026-10-20 — by Mahmoud Zalt*

Where to find Zendesk's AI agent numbers, what deflection and automated resolutions actually mean, and which metric tells you the truth.

**Short answer.** **Two different numbers get called the same thing, and only one of them is honest.** Deflection counts conversations that did not reach a human. An automated resolution is Zendesk's own billing unit and is checked by a language model reading the conversation after a 72 hour window with no follow-up. You find both in the AI agents workspace dashboard and in Admin Center under Account, then Usage.

**About this guide.** This guide is based on publicly available information and Zendesk's own documentation, checked on 17 August 2026. Sistava is not affiliated with Zendesk. Interfaces and pricing change often, so confirm the current steps in their documentation before you rely on them.

Every support team that switches on an AI agent gets asked the same question within a month: is it working. The trouble is that the easiest number to produce is also the one that flatters you most, and the number that actually tells you something takes three days to settle.

This guide covers what Zendesk means by each term, where the numbers live, and how to tell whether the figure on your dashboard represents customers who got helped or customers who gave up. Those two look identical in a deflection count.

## What is the difference between deflection and an automated resolution?

Deflection is the broad one. A conversation counts as deflected when it does not end with an escalation to a human. That includes conversations where the customer got a perfect answer, and conversations where the customer closed the window in frustration, because from the counter's point of view both look like nobody escalated.

An automated resolution is narrower and it is the unit Zendesk uses for billing AI agent usage. Their documentation describes a verification step where a large language model reads the text of the conversation after a 72 hour window with no customer follow-up, and confirms that the request was satisfactorily resolved. A conversation that fails that check does not become an automated resolution.

That 72 hour window is the part worth internalising. It means today's number is not final, and a dashboard read on Monday afternoon will not match the same period read on Thursday. If you are reporting weekly, report on a period that has already settled.

## What are Zendesk's automated resolution tiers?

In May 2026 Zendesk split the single automated resolution measure into three tiers, so that a conversation where the AI agent helped a bit is no longer counted the same way as one it finished on its own. Their documentation dates the introduction to 18 May 2026.

| Tier | What it means | Allowance |
|---|---|---|
| Assisted escalation | The AI agent contributed to the interaction before a human agent completed the resolution | Their page states this does not count against your resolution allowance |
| Contained resolution | The AI agent handled the interaction to completion without the customer requesting further assistance | Their page states this does not count against your resolution allowance |
| Verified resolution | The AI agent successfully resolved the interaction, confirmed by the language model check | Their page does not state an exemption for this tier, unlike the two above |

Zendesk describes the reasoning as making sure you do not overpay for straightforward tasks while still reflecting the value of harder ones. For measurement purposes the useful part is that the tiers give you three honest categories rather than one blunt one, and that alone is worth reading your dashboard differently for.

## Where do you actually find these numbers?

In more than one place, and each answers a different question. The dashboard in the AI agents workspace answers how is it performing. The usage page in Admin Center answers how much of my allowance have I used. Do not use one to answer the other's question.

1. **Open the AI agents workspace dashboard** — It is the default landing page in the AI agents workspace, reachable from the Dashboard icon in the sidebar. Its Insights section shows automated resolutions as a percentage of conversations, escalations as a percentage, and total conversations.
2. **Check usage in Admin Center** — Go to Admin Center, then Account, then Usage, then Automated resolutions. This shows allowance usage for the current billing period, the total consumed, and a daily usage chart with an option to view cumulatively.
3. **Use the Forecast tab for planning** — Zendesk's documentation describes a Forecast tab that projects usage from a moving average of the past 7 to 30 days and estimates when your allowance may run out.
4. **Look at individual tickets** — A ticket carries a Resolution type field, which shows Automated when it qualified. This is how you check a specific case rather than an average.
5. **Read the conversation logs** — Conversation logs in the AI agents workspace mark automated resolutions with an icon, so you can read the actual exchange behind a counted resolution.

That last step is the one to make a habit. Zendesk's documentation also notes that alerts fire in Admin Center and by email to your billing admin at 80 percent and 100 percent of your allowance, so you are unlikely to be surprised by the usage side. Nothing alerts you about answer quality.

**The other way to look at this.** Deciding what counts as resolved is a question every support team has to answer, whatever tool it uses. At Sistava we built around the same principle, which is that an AI Employee reports what it actually finished rather than what it merely handled, and asks you when it is unsure instead of closing the conversation. You still read the transcripts, because no metric replaces that.

The habit that catches problems earliest is reading a fixed number of real conversations every week, chosen at random rather than picked from the ones that look interesting. Twenty is enough to notice a pattern and small enough that someone will actually do it.

## How do you tell whether a high number is real?

Pair it with a second number that would move in the opposite direction if customers were quietly giving up. The best pairing is repeat contacts: how often the same customer comes back within a few days on the same topic. A conversation that counted as contained and then reappears as a new ticket is not a success, it is a delay.

Satisfaction scores specific to AI conversations are the second check, kept separate from your overall satisfaction score so a large volume of human handled tickets does not hide a problem. Independent write ups on Zendesk metrics make the same recommendation, and it holds for any tool in this category.

The third check is the split between escalation and containment on your hardest topics. If containment is high on billing questions and low on password resets, something is inverted, because the harder category should be escalating more, not less.

## Which metrics are worth reporting to the business?

Four, and deflection is not one of them. Report the numbers that would change a decision, and leave out the ones that only ever go up.

- **Verified resolutions: **the tier where Zendesk's check confirmed the request was actually resolved.
- **Escalation rate by topic: **not the overall figure, which averages away the thing you need to see.
- **Repeat contacts within seven days: **the honest counterweight to any containment number.
- **Satisfaction on AI handled conversations: **kept separate from your overall score so it cannot be diluted.

One more note on the moving target underneath all of this. Zendesk announced on 30 March 2026 that the split between Essential and Advanced AI agent plans is being removed, with the rollout running from 11 May to 12 June 2026, and their announcement gives 31 August 2026 as the point development stops on legacy functionality and 10 December 2026 as the full shutdown. Reporting pages differ between the old and new shapes, so check whether the help page you are reading is marked as legacy before you trust its screenshots.

**Three ways to fool yourself with these numbers.** Reading a period before the 72 hour verification window has passed gives you a figure that will change. Reporting deflection instead of verified resolutions counts the customers who gave up as wins. And judging quality from the dashboard alone hides the specific bad answer that is quietly costing you a category of customer, which only reading transcripts will surface.

None of this is unique to Zendesk. Every product in this space, ours included, has to choose what counts as done, and the honest ones make you wait for evidence before they claim it. When you are comparing tools, ask how each one decides, not what its headline number is.

## FAQ

### What counts as an automated resolution in Zendesk?

It is the unit Zendesk uses to calculate and bill AI agent usage. Their documentation describes a verification step in which a large language model reads the text of the conversation after a 72 hour window with no customer follow-up and confirms the request was satisfactorily resolved. Conversations that fail that check do not count.

### Why does my Zendesk automated resolution number keep changing?

Because of the 72 hour window. A conversation is assessed after three days with no customer follow-up, so a figure read today for a recent period is provisional and will settle later. If you report weekly, report on a period that has already passed that window rather than on the last few days.

### What are the three Zendesk resolution tiers?

Assisted escalation, where the AI agent contributed before a human completed the resolution. Contained resolution, where the AI agent handled the interaction to completion without the customer requesting further help. Verified resolution, where the AI agent successfully resolved it. Zendesk's documentation dates the introduction of the tiers to 18 May 2026.

### Where do you see Zendesk AI agent performance?

The dashboard in the AI agents workspace is the default landing page there, and its Insights section shows automated resolutions as a percentage, escalations as a percentage, and total conversations. For allowance consumption, go to Admin Center, then Account, then Usage, then Automated resolutions.

### Is deflection rate a good measure of AI support performance?

On its own, no. Deflection counts conversations that did not reach a human, which includes customers who got the answer and customers who abandoned the conversation. Pair it with repeat contacts within a few days and with satisfaction on AI handled conversations, and the picture becomes trustworthy.

### Does Zendesk warn you before you run out of automated resolutions?

Yes. Their documentation says you are alerted in Admin Center and your billing admin receives an email when you approach the limit, with thresholds at 80 percent and 100 percent of your allowance. There is also a Forecast tab that projects usage from a moving average of the past 7 to 30 days.

## Sources

- [Zendesk help, ](https://support.zendesk.com/hc/en-us/articles/9570369117338-About-automated-resolution-tiers)About automated resolution tiers
- [Zendesk help, ](https://support.zendesk.com/hc/en-us/articles/8922391373978-Monitoring-automated-resolution-usage)Monitoring automated resolution usage
- [Zendesk help, ](https://support.zendesk.com/hc/en-us/articles/10684348229274-Using-the-dashboard-to-monitor-and-manage-AI-agents)Using the dashboard to monitor and manage AI agents
- [Zendesk help, ](https://support.zendesk.com/hc/en-us/articles/10487730059034-Announcing-expanded-access-to-AI-agent-capabilities-for-all-Zendesk-customers)Announcing expanded access to AI agent capabilities for all Zendesk customers
- [eesel AI, ](https://www.eesel.ai/blog/zendesk-ai-agent-metrics-to-track-automated-resolutions)Zendesk AI agent metrics to track automated resolutions

**Tags:** zendesk, automated-resolutions, support-metrics, deflection-rate, customer-support-automation