Why Amrit Raj thinks AI support should measure completed requests

Amrit Raj, Co-founder and CEO at Flowcall
The Unpopular Opinion

Why Amrit Raj thinks AI support should measure completed requests

Amrit Raj, co-founder and CEO of Flowcall, works on an AI customer service platform for consumer brands. Raj argues that support bots should be judged by whether they finish a customer’s request, not whether they keep that customer away from an agent.

Amrit RajCo-founder and CEO at Flowcall
Published November 21, 2026
The takeaway

Measure whether a customer’s request was completed, rather than whether a bot prevented contact with an agent. Review chat samples and track whether customers return after being deflected to see where the bot is falling short.

01

What’s one opinion about your industry that most of your peers would disagree with?

Deflection is the wrong way to judge AI in customer service.

Most of my industry measures a support bot by how many customers it keeps away from a person. I think that number rewards the wrong thing. A customer asks for a refund. The bot sends a policy link. The dashboard counts a win. The customer writes back angrier, or calls.

The real test is simple. Did the request get finished?

02

What convinced you? Tell us about the moment, client or result that changed your mind.

It was one of our sales calls, with the person who runs support at a large consumer brand. Their bot showed about 85% of chats as resolved. Then they said it out loud: "We felt like we were resolving 85. Maybe the true result wasn’t actually 85."

That line stayed with me. Their number looked great, and they still didn’t trust it.

Nobody on that team was lying. The metric was counting the wrong thing. A customer who gave up looked the same as a customer who got helped. And the ones who gave up don’t vanish. They come back to an agent, or they post about it.

03

What’s the usual way people in your field handle this, and what do you do instead?

The usual way is to count what the bot kept away from agents. If the customer didn’t reach a person, it counts as solved. So bots are built to answer fast, send a link and close the chat.

We count it the other way. A chat only counts as handled by our AI if no agent had to step in. Every handoff counts against us. So our AI has to do the work. It follows the brand’s own rules, checks the order in their systems, and completes the refund, exchange or cancellation itself.

When it isn’t sure, it doesn’t guess. It hands the case to an agent with the full story attached. So the customer doesn’t have to start over.

04

Where has this view cost you, or where have you had to admit you were partly wrong?

It makes our number look smaller. We count every handoff as a miss. Next to a deflection number, our honest one can look weak on a slide. So I spend time on calls explaining the difference.

I’ll also admit this view has limits. Not every link is a dodge. Some questions are fine to answer with a link. Store hours, a size chart or a tracking page are real answers. My problem is with counting a refund request as solved when the customer is still waiting.

05

What should someone in your field stop doing, and what should they start doing?

Stop reporting deflection or containment as your main number. Stop counting a closed chat as a solved problem.

Start counting how many requests got finished with no agent stepping in. Read a sample of your bot’s chats every week, start to end. Then check if the customers who were "deflected" came back within a few days. That repeat number tells you more than any dashboard.

Thank you to Amrit Raj and the team at Flowcall for sharing an honest view with Leaders Perception readers.

Share this interview

Have a view your industry needs to hear?

Leaders Perception publishes interviews with business owners and leaders who see their industry differently. Five questions, about ten minutes, no calls. It is free and every completed interview gets published.

Get featured on Leaders Perception

Explore additional categories

Explore Other Interviews