Say a support leader switches an AI agent on across the whole queue on a Monday. It’s a made-up case, but not a far-fetched one. By Friday the ticket count is well down and the dashboard looks good. If anyone asked about AI customer support ROI that week, the answer would be a chart heading the right way.
Then someone sits down and reads a handful of the conversations where the customer never replied. Some look fine. Some look like the agent answered a slightly different question from the one that was asked. Nobody in the room can say whether those people were helped or just gave up and went looking somewhere else.
That’s where the numbers stop meaning much. On a count, a ticket that went quiet looks exactly like a ticket that got solved.
AI is not going to solve all support problems, and I don’t think it needs to. What matters is which ones it takes. I’d judge an agent on what it removes from the queue without hurting the customer’s experience, and ideally on whether it makes that experience a bit better. That holds whether you run it yourself or with an outsourced support provider, and it comes down to a few things worth settling before anything gets switched on.
Start with the three to five ticket types the agent can handle really well
The first thing I’d ask a team, or a support partner offering an agent, is which contacts it will take first.
I want to hear them named. Look at what comes in, find the most common repeating issues (the top three to five) and make sure the agent solves those really well. Once that’s in place and working, add the next one. Everything else stays with people, handled the way it was before. Start small, start simple, and build on that.
What worries me is “we’ll switch it on and see how many tickets drop”. People turn an agent on hoping it will just reduce the tickets, and it sounds practical, but it points the agent at the whole queue at once, including the contacts it’s least suited to. Success becomes a number going down. That’s the Monday-to-Friday case above. A better chart, and no idea what happened to the customers behind it. It’s a poor starting point for any claim about AI customer support ROI.
Before a ticket type goes on the list, I’d ask where the agent will truly reduce the flow of tickets without hurting the experience. If nobody can answer that honestly, the type stays with the team. If the answer is yes, and it might even make things better for the customer, it’s a candidate.
The common repeating issues usually pass, because your team already answers them the same way every time. The fix is known and you can check the result. They’re also the same top issues you’d measure before and after a fix, so you’ve already got a baseline.
It takes some pressure off, too. Nobody’s expecting the agent to solve everything, so nobody has to pretend it does. Pull last month’s tickets and count your top five before you talk to anyone about an agent.
A thank you is a result and silence is something to review
Next, I’d want to know how you’ll tell a solved ticket from one that just went quiet.
It’s fairly simple. If the person replies with a thank you, or anything else positive, that’s a good result. Something negative is a bad one. If they say nothing, the conversation gets reviewed to check whether the response really solved the issue and was a good answer. A separate conversation agent can do that review from the outside, so the agent isn’t marking its own homework.
People come in when the review comes back negative. Those are the conversations to learn from, so the team can work out what went wrong and make it less likely to happen again.
Count every quiet conversation as solved and you’ve found the quickest way to make weak support look efficient. The silence gets booked as a win.
I’d read the silent ones first. That’s where the honest picture of your AI customer support ROI is.
The agent should be an extension of how your brand already behaves
This one’s yours to answer before anything goes live. How do you want people to experience your brand?
Think about what it’s like when someone talks to your sales team, walks into one of your shops or goes round your website. You’ve built each of those around your values and goals, and the AI should be an extension of that. If it sounds like every other agent out there, it’s carrying your logo and not much else.
It matters most with a customer who’s already upset by the time they reach the agent, because a delivery is late or a vehicle is unusable. The agent has to be quick, and it has to sound like it understands and means it. I don’t want “happy to help” and “nice to hear from you” stacked in front of the answer. People don’t want fake, they want action.
Write down what action looks like for your most annoyed customer, and check the agent’s first reply against it.
Handover should sound like a colleague picking it up
When the agent can’t help, it should do what a person would: “I’m sorry, I need to pass you to a colleague so we make sure you get good service.” By the time it says that, the handover should already be done, and someone should pick it up fast. If the customer has to explain everything again, they’re starting over. Ask to see a real handover before you sign anything.
AI customer support ROI is the contacts a person no longer had to handle
Six months in, someone will ask whether it paid off, and I’d keep the numbers short.
I’d show how many conversations the agent held, the percentage that ended with a positive outcome, and how many were moved to a human. Every positive outcome that didn’t need a person is the return, because a person would otherwise have handled it. To keep the arithmetic simple, say it held a thousand conversations and six hundred ended well with nobody stepping in. That’s six hundred contacts your team didn’t have to take, and it’s AI customer support ROI in a form anyone can check.
I’d also want one central ticket system where everything gets logged, whatever channel it came through. That’s how you see who comes back another way, and how you speak to them knowing they’ve already been in touch.
A deflection figure on its own tells you how much left the queue. It says nothing about where the customer went next.
There’s one case I haven’t settled. A customer says thanks, the conversation gets marked positive, and the next week the same person emails or rings about the same thing. The central system will show me that. It won’t tell me how to count it. The thank you was real, but the problem clearly wasn’t solved, and I haven’t made up my mind which of those should win, at least not yet. Decide
whether that first conversation goes into your AI customer support ROI before someone asks you at the six-month review.
I started turning spanners at sixteen and went on to lead global support for some of the world’s most demanding industries. I’ve built a reputation for challenging the traditional ways of doing things, and pushing the limits of what’s possible.
From transforming operations at Volvo to building human-centred support models at Sigma Technology and now FIXATE, I lead with clarity, care, and courage. I see the big picture, sweat the details, and always try to do what’s right, even when it’s hard.
If something’s worth doing, I want to do it properly. And I’m probably already thinking about how to make it better.