For most of the last decade, e-commerce teams have treated customer service as a math problem. Volume goes up, so you add agents, outsource a night shift, or bolt a chatbot onto the helpdesk to deflect the easy stuff. Every version of that answer accepts the same premise: that service is a cost that grows with the business, a tax you pay for getting bigger.

I’ve come to believe that premise is wrong. Customer service was never really a headcount problem. It was an automation problem we didn’t yet have the tools to solve, and the teams that understand the difference are about to pull away from the ones that don’t.

The queue is a symptom, not the disease

Look at what’s actually sitting in a support queue on any given morning. Where’s my order. I need to change my address. Can I return this. Is this back in stock. Very little of it is a conversation in the human sense. Most of it is an operational task wearing a conversation’s clothes: someone has to open the order management system, check the warehouse, look up the carrier’s status, read the returns policy, and only then write back.

As a store grows, that operational drag compounds, and it doesn’t grow smoothly. E-commerce now runs on two peaks, not one: the rush around Black Friday, and the returns wave that lands weeks later. Demand like that doesn’t just get bigger, it gets spikier, and headcount is the one resource that can’t flex on that timeline. You end up either overstaffed for eleven months or underwater for one. Neither is a strategy.

The support queue, in other words, is simply where a much larger operations problem becomes visible. It’s the symptom. The disease is that the manual work behind the buy button scales in lockstep with revenue.

Why “add AI to your helpdesk” hasn’t fixed it

The obvious objection is that we already have AI for this. We’ve had support bots for years. But most of them were built the way the last generation of software was built, an AI layer bolted onto a legacy system it can’t really touch. It can recognize that a customer is asking about a return. It cannot actually process the return. It can tell that someone wants to change a shipping address. It can’t reach into the order system and change it. So it deflects, the customer gets more frustrated, and the ticket lands back in a human’s lap anyway, now with an annoyed shopper attached.

That’s the gap that matters, and it’s worth being clear-eyed about it underneath all the AI hype. The unlock isn’t a smarter FAQ. It’s AI wired directly into the commerce stack, the OMS, the WMS, the ERP, the custom tools a brand has stitched together over the years, so that it can understand an order, a delivery, a return, or a product, and then complete the task end to end. The distinction that counts is between deflection and resolution. One makes the queue look shorter. The other makes the work disappear.

What resolution looks like at scale

We’ve spent the last two years testing that thesis in production. Today our platform handles roughly 250,000 customer conversations a month for around 70 brands across five countries, and resolves up to 90% of support tickets not by deflecting them, but by solving the underlying problem.

The result that surprised even us wasn’t the cost line, though that moved too; brands routinely take close to half of their support cost out of the equation. It was that service quality didn’t have to drop to get there, actually it improves for most brands. Customers get an accurate answer and a resolution to their problem instantly, at 3 a.m. on a Sunday, in their own language, and satisfaction holds or climbs. The trade-off everyone assumes, cheaper service or good service, turns out to be a false choice once the AI can actually act across your systems.

For what it’s worth, that same model is what let us in Mimir grow roughly sevenfold in a year and stay profitable without outside funding. I mention it only because it’s the clearest evidence I have that this isn’t a lab demo. When automation genuinely removes work rather than shuffling it around, the economics show up quickly.

Support is the starting point, not the destination

Here’s the part I’d push every retail leader to sit with. Once AI can take real action across your stack to resolve a support ticket, resolving tickets stops being the interesting question. The same capability reaches straight into the rest of your operations: modifying orders, managing subscriptions, catching a delivery exception and fixing it before the customer even notices anything went wrong. Support is just the most visible slice of the operational work that slows e-commerce teams down. It’s the wedge, not the whole.

The next phase of this industry won’t be won by whoever staffs up fastest for peak. It’ll be won by the teams that stop treating operations as the price of scale and start treating it as something software can carry.

What to do with this on Monday

If you run e-commerce operations, three shifts are worth making now, well before you evaluate a single vendor:

1. Change the metric. Stop measuring your service org by tickets-per-agent and start measuring resolution rate and cost per order. The first number rewards looking busy; the second rewards actually solving things.
2. Audit your queue. Sort a week of tickets into operational lookups versus genuine human-needed judgment calls. That ratio is your automation ceiling, and for most brands it sits far higher than they’d guess.
3. Judge AI on what it can do, not what it can say. When you assess any tool, the question isn’t how fluent the chat is. It’s whether it can meaningfully reach into your systems and complete the task. If it can only talk, it’s a deflection engine with better manners.

The buy button got automated years ago. Everything that happens after it is the frontier now. And for once, the technology is finally good enough to meet it.

About the author

Jørgen Vartdal Halse is co-founder and CEO of Mimir, an AI-native operations platform for B2C e-commerce based in Oslo, Norway.