---
title: "How to tell whether your AI is finishing the job"
url: "https://customerrelations.io/insight/how-to-tell-whether-your-ai-is-finishing-the-job/"
author: "Aarohi Kulkarni"
published: "2026-10-01"
updated: "2026-10-01"
---

# How to tell whether your AI is finishing the job

It's 11pm. A customer opens the chat on your website and says they want to cancel their subscription before it renews in the morning. The bot replies in two seconds with a link to a billing article. They read it. The third paragraph says to contact support to cancel, so they close the tab and write an email instead. They explain the whole thing again. At 9am the renewal goes through. The next message from that customer is a refund request.

On your support dashboard, that conversation looks good. A question came in, the bot answered, and nobody on the team had to touch it. A lot of AI support tools count that as deflected. Some call it resolved.

In September a USA Today reporter put out a request for comment on "doom loops" in travel customer service, where a customer keeps getting sent back to where they started. I build these systems, so I answered from that side: the loop usually starts with what the company chose to measure.

Deflection measures one thing: the customer stopped talking to you. It's an easy number to count, which is probably why it's the one on the dashboard. People stop for two reasons. Either the problem got solved, or they gave up. The metric can't tell the difference. The customer who gave up at 11pm looks the same as a happy one until they write in again, or leave at the next renewal.

Last May, someone posted their own numbers on an AI vendor's community forum. Over four weeks, the tool had marked 16 of their tickets as resolved. They checked all 16. Nine had not been solved at all.

Customers judge an experience largely by its worst moment and by how it ended. It's called the peak-end rule, and it works against deflection. The customer who tried to cancel at 11pm will remember explaining themselves twice, on two channels, and paying for a month they tried to stop. They learn that your website is the slow way to get help. Next time they skip the chat and send an email, or go to a review site instead. Your deflection rate goes up, and the relationship gets worse in the same month.

So when I look at any AI support system, including the one I build, [DeskClone AI](https://deskclone.ai), I ask one question. Can it finish the job, or can it only answer the question?

Answering is the link to the help article. Finishing is the subscription cancelled inside the chat. It's the appointment moved, the return label sent, the invoice re-sent to the right address. Finishing is harder to build and harder to trust, so the business has to set the limits: which actions the agent may take on its own, how far it can go, and when it has to hand over to a person. And when it hands over, the whole conversation goes with it, so the customer doesn't have to explain it all again.

If you run a support team, there are two things I'd do this month.

First, read the conversations your tool marked as resolved. Twenty is enough to see a pattern, and it takes about an hour. Start with the ones that ended in a link. Keep a tally of how many actually ended with the problem fixed.

Second, start tracking repeat contact: the same customer, the same issue, within seven days, on any channel. A chat that ended with a link and came back as an email the next morning belongs in that number. Right now it probably shows up as a success.

One thing I don't know yet is what a good completion rate is. Every vendor defines "resolved" its own way, so the benchmarks don't line up, and I'd be careful with anyone who says they know. Your own count from those twenty conversations is a better number than any benchmark, because you know exactly how it was made, and it moves when you fix something.

The customer at 11pm didn't want an answer. They wanted it cancelled.

---

Aarohi Kulkarni is the founder of DeskClone AI ([deskclone.ai](https://deskclone.ai)), a platform for building AI support agents that clone your team's expertise and resolve customer issues end-to-end.