This interview is part of the SparrowDesk Spotlight series, where we sit down with the people rethinking how customer support actually works.

Ines van Dijk has spent nearly two decades in customer support as a frontline agent, a team lead, a QA specialist at Automattic, and eventually the industry's go-to QA consultant for companies that have a quality process in place and quietly know it isn't working.

More recently, she's turned her attention to research: what is AI actually doing to the shape of support work, and are we measuring any of it correctly?

We sat down with her to talk through both. Here's the conversation, lightly edited for clarity.

On scoring AI agents like human agents

Q: Most companies now run AI on the front line of support. Should AI agents be scored with the same QA rubric built for human agents?

"What I see happen quite often is that the QA scorecard is directly lifted from the human agents and applied to the AI. Which, if you think about it, is not the right way to go about it. It's a different element.

The way it responds to customers is completely different: it can look like human interaction, it can mimic human interaction, but it's not the same. And when humans are dealing with things that are much more difficult, the things you track with them should be completely different from what you're tracking with AI."

The uncomfortable implication: most support orgs haven't actually built a second rubric. They've just pointed the old one at a new kind of agent and called it coverage.

On deflection, and why it's the wrong number to chase

Q: Deflection rate has become the go-to metric for AI performance. What's wrong with optimizing for it?

"Deflection is essentially what percentage of tickets can the AI handle without any human touching it whatsoever. But what that means is that a lot of the interactions where the conversation could have been saved by a human stepping in, weren't. What I'm advocating for is decomposition, not deflection. Smart deflection means having certain markers in place to say, okay, this is a point where it's actually better to hand it off to a human, because even if the AI was technically capable of solving this without human intervention, with human intervention these cases resolve much better, or much faster, or with a higher CSAT score afterward."

Her point isn't that AI shouldn't close tickets solo. It's that "solo close" and "best outcome" are two different targets, and most dashboards only track the first one.

On why customers bristle the moment they spot a bot

Q: Why do customers get frustrated the second they realize they're talking to AI instead of a person?

"When a customer comes in, something's already broken. They've experienced damage to the trust they have in the company, or something has happened that made them have to ask for help, and asking for help is incredibly vulnerable.

The problem with an AI is that it can only give a categorical response, never a specific one. And humans are extremely good at telling the difference, even when we can't put our finger on why. We notice, fast, whether an answer is actually about us or just a template we've been slotted into. When the response reads as categorical, the customer doesn't think 'that was a bit generic.' They think 'my problem doesn't matter enough to this company to get a real answer.' And that lands on someone who's already feeling vulnerable, so it's a second hit on top of whatever brought them to support in the first place."

It's a distinction worth sitting with: an AI can sound empathetic. It cannot, in her framing, actually embody empathy, and customers in a vulnerable moment tend to notice the gap even when they can't articulate it.

On the part of the job that's quietly gotten worse

Q: Where has AI made a support agent's job harder rather than easier?

"It comes back to queue composition. AI snaps up the automatable bits,  it can maybe solve 70 to 80 percent of tickets.

But not a whole lot of volume expectations have changed. Humans are still expected to handle a comparable volume to what they handled before the AI implementation, except the cases they're now responsible for are far more cognitively demanding. I'm hearing reports all over the place: 'Everyone is tired. Everyone is burned out. I'm solving 70% of my volume through AI, but the remaining 30%, I can't keep people in their seats because the job is so horrible.' I don't think we've been presented with the bill for that yet. But I think it's coming."

She's also blunt about the second-order cost: teams that lean too hard into AI and lose their most tenured agents in the process aren't just losing headcount. They're losing the institutional knowledge that makes the next generation of agents good at the job, and that doesn't come back quickly.

On the biggest misconception about support as a function

Q: What's the biggest misconception people still have about customer support?

"We're still seen as bottom of the barrel when it comes to the organizational structure. My theory is that we're still essentially seen as secretaries: you're just sending some emails, you're just talking to people. That's a huge misconception, especially if you look at it from a strategic standpoint. The CX department is like the canary in the coal mine for organizational health as a whole, and there's business intelligence sitting there that companies would otherwise spend a lot of money to gather. If a product team wants to understand how customers are actually using what they've built, the first place they should stop is CX."

Ironically, she points out one silver lining of the AI shift: the budget scrutiny AI has brought to support has, for the first time, put support's numbers in front of the C-suite in a way that's hard to ignore.

On the metric she'd retire, and the one she'd replace it with

Q: If you could ban one support metric forever, what would it be?

"CSAT, without hesitation. It's fairly empty as a metric, it needs context. It's an emotional snapshot of one moment in time, and quite often it's not even about the actual interaction itself. It's about policy, or something that broke upstream, and we're not attributing that appropriately."

What she'd track instead: interaction tractability, essentially, how solvable was this conversation, given the conditions it started under? "The entry point is where the upstream failures will show up, the agent didn't cause that condition, they inherited it. Then there's the trajectory from that entry point to the end: am I going up or down in how solvable this problem is within this one interaction? That's where performance actually sits."

On the advice that's shaped how she thinks about quality

Q: What's the best advice you've ever gotten about measuring quality in customer support?

"That context is king. Rather than purely looking at the numbers, you need to understand what the actual context is. You could have your highest performer suddenly take a dip in their output, and that might be a reason to flag their performance and put them on an improvement plan. Then you start talking to them, and the context is actually that their spouse is really sick and they've been taking care of them. Context matters, and it makes it excusable."

Gift or threat?

Asked to sum up AI's impact on the industry in one word, she landed on evolving, not a dodge, she was quick to clarify, just an honest one. And asked whether AI is ultimately a gift or a threat to people building careers in support, she didn't hesitate:

"I think it's more of a gift, because I built my career before AI was even there, and I've seen the solutions it's brought to a lot of the problems I struggled against. But I do think we need to be very mindful of the negative impact it has, and I don't think we're quite ready for it, or equipped to face it. That boils down to support still being an underfunded department."

More from Ines

Ines van Dijk is the author of The Customer Support QA Playbook, available on Amazon, and publishes ongoing research at the Customer Support Research Lab on Substack. If any of the above struck a nerve, both are worth your time.