Back to Thinking

Does quieter customer support mean a better product?

Imagine a desk lamp. A nice one. A slim stem, a heavy base and a single button. Press it and the light comes on. Press it again and it goes off.

The product page said it was dimmable. But the button doesn’t turn. There isn’t another one. You press it a few times. On, off, on, off. The two of you have established a remarkably consistent relationship on this particular point.

You ask the AI assistant on the brand’s website. The answer arrives straight away: press and hold the button to dim the light.

You try it. It works. You say thanks and close the window.

From the support system’s point of view, that went well. The customer didn’t wait, got the right information and didn’t take up an employee’s time. The lamp is now at the brightness they wanted.

A few days later, someone else asks the same thing. Then another person. The assistant gives everyone the right answer. We can say the company is getting better at answering this question. We don’t yet know whether the lamp has become any easier to understand.

That’s the part that interests me as a product designer.

The value of AI support is easy to see. Someone setting up a lamp at home in the evening shouldn’t have to wait for the company to open the next morning. A support employee doesn’t need to spend the day describing the same button, either. I doubt anyone has “explain pressing and holding in eight languages” in their professional development plan.

But it’s worth looking at what this convenience changes inside the company.

When a support team keeps getting the same question, the repetition has a weight to it. Calls run long, queues build up, colleagues say “this one again.” In a well-run organisation, that information may eventually reach the people who designed the product. It doesn’t always. Sometimes support spends years explaining the same thing while the design team works on new colours.

Still, there’s something uncomfortable there. Something that takes up a person’s time, shows up in a budget and might get mentioned in a meeting.

When the cost of answering that question falls, some of the discomfort around the product can fade with it. The customer still needs an explanation. The company feels it differently.

I wouldn’t jump from this to the conclusion that AI distances companies from their customers. The same technology can also find patterns in conversations nobody previously had time to read in full. Intercom’s Topics Explorer, for example, uses AI to group support conversations by topic and show how much conversation volume each one generates. A tool like that can make visible something the team had only noticed in bits and pieces.

So it’s possible to see all those people asking about the lamp’s button in one place. It may even be easier than before.

What happens next depends on what the company wants to learn.

A team that notices the repeated question can improve the assistant’s answer. Write a shorter explanation, add a little video, adapt the response to the customer’s language. All useful work. All helpful to the person sitting at their desk that evening.

The information for the product team is slightly different: people know the lamp can be dimmed, but they can’t work out how to do it from the button.

That opens up more possibilities. A small card visible when the box is opened might be enough. A mark beside the button might help. Perhaps the interaction itself needs to change. Or perhaps the company decides that keeping the lamp’s simple appearance is worth a brief explanation on first use.

Not every repeated question calls for a redesign. There may be a perfectly reasonable aesthetic or manufacturing reason for having one button. Asking users to learn something isn’t automatically a design flaw. But there’s a difference between deliberately making that decision and quietly leaving support to deal with the question.

Then there are the words we use in reports.

“Resolved” is one of the more reassuring ones. It sounds finished. Yet you sometimes need to read the definition of what a support system counts as a resolution. In Intercom’s Fin reporting, the relevant AI answer resolution metric can include conversations where the customer didn’t ask to speak to an employee, alongside those where they gave positive feedback.

That kind of measure can help track an operation. But when a customer stops talking, we don’t know everything that happened in their home. They may have figured out the button. They may have found an answer somewhere else. They may also have decided they’d had enough for one evening.

Even when the customer explicitly says “thanks, that worked,” there may still be something for the product team to learn. In our example, the lamp works, the customer is happy and the conversation really is resolved. The next customer may still need exactly the same explanation.

I think that’s an easy thing to miss in support automation: successfully answered questions are worth looking at too.

If we only examine conversations the assistant couldn’t answer, we can find gaps in its knowledge. To understand where the product makes people work, we also need to read some of the conversations it handled correctly. Knowing the answer very well doesn’t explain why the question gets asked so often.

This is one of the things that interests me about aftercare when I think about Afterlayer. The needs that come up during setup, maintenance and repair carry very specific information about the product. We can see where people stop, which part they can’t identify and which explanation helps them continue. Those observations could inform the next product decision.

Of course, looking at visit counts and immediately drawing conclusions would be too easy. A help page may get a lot of traffic because the product sells well. The same person may keep returning because they don’t understand the explanation. We may also need to investigate what happens to people who never contact support. Bringing data together doesn’t automatically produce a good interpretation.

Still, if people repeatedly struggle at a particular step, it’s reasonable to expect that information to reach someone inside the company. It could be addressed in the next model, change the information packed with the current one or be left as it is for good reasons. At least the issue would travel a little further than the desk of the person writing the reply.

Back at the lamp, the AI assistant’s answer did its job. That evening, the light was adjusted and the person stayed at their desk. There’s no need to dismiss the value of that.

I’d also send a copy of that conversation to the team that designed the lamp, along with how often the question comes up. They might decide to try a small mark beside the button. Then we could see whether that mark actually tells anyone anything.