What «70% resolved autonomously» means

From the engine room – numbers, decisions and lessons from running support.

Few numbers come up faster in a conversation about AI Agents than the resolution rate: «70% resolved autonomously.» The number looks precise. But it’s only as good as its definition, and that almost never comes with it.

The definition, and whose it is

A case counts as resolved autonomously when the customer makes no further contact for 72 hours after the bot interaction and no human had to step in. That definition isn’t ours, it’s Zendesk’s. The Automated Resolution is their official unit of measure, the one AI Agents are also billed on, including an AI evaluation that checks whether the agent’s answer was relevant at all.

The definition is deliberately unspectacular. Nobody can look inside a customer’s head to know whether an issue is settled. The most reliable observable signal is that nothing follows: not in chat, not on the phone, not by email. And because the same number lands on the invoice, it has to hold up to harder scrutiny than any marketing rate. When you pay per resolved case, you look closely at what counts as resolved.

The justified no counts too

The rate also includes cases where the customer’s request can’t be granted for legal reasons. That belongs in there, and rightly so: «resolved» doesn’t mean «the customer got what they wanted», it means «the issue has been handled conclusively». A cleanly explained no, one the law makes mandatory, is a resolution. The customer knows where they stand without having to wait for a human. In a regulated environment that’s not an edge case, it’s everyday reality.

Where the definition reaches its limits

Silence is a proxy, not proof. Someone who simply gives up after the bot interaction looks the same in the stats as someone whose problem got solved. Zendesk’s relevance check doesn’t change that, because relevant isn’t the same as satisfied. So the rate never stands alone for us: next to it belong the reasons for the repeat contacts, the escalations, the satisfaction scores, and a look at the conversations themselves. The interactions are reviewed weekly; the feedback loop between the team and the bot is a fixed process.

And there are situations where silence wouldn’t be a success but an alarm. What you need then is in the note on fallbacks.

The third question: percent of what?

Besides the time window, the denominator decides what the number means. 70% of all requests? Of the bot-eligible ones? Or of the ones who chose the bot voluntarily? On our Voice rollout we let callers choose: bot or human, right at the start. Around 70% pick the human up front. This Opt-in logic depresses the rate structurally, and it’s still right: it puts the experience first and delivers real acceptance data. Compare rates without knowing the denominator and you’re comparing nothing.

How to spot a reliable rate

Three questions are enough:

1. When does it count as resolved? Which time window, which signal, and across every channel or just one?
2. Does the justified no count? And is it explained to the customer cleanly, instead of brushing them off?
3. Percent of what? All requests, bot-eligible, or Opt-in?

For context against our own numbers: our chat sits at 65%, against a stretch goal of 70. In the first year, 40–60% is realistic, depending on the contact reasons; after that the rate grows in operation, with every optimization cycle. The whole story with all the numbers, including the ones we didn’t reach, is in the Munich note.

A rate without a definition is marketing. A rate with a definition is an operational metric, and that’s the only kind worth talking about.