The Support Problem That Gets Harder the More AI You Add

Here’s a paradox nobody running a support org wants to say out loud: every layer of AI you add makes the support work that remains harder. Not marginally harder. Structurally, compoundingly harder — through four mechanisms that reinforce each other, and that most deflection dashboards hide by design.

The backdrop is impossible to miss. Last week Bloomberg reported that Microsoft, Uber, Commonwealth Bank, and Hyatt have cut thousands of customer service roles they credit directly to generative AI. Read the fine print and one detail jumps out: the cuts are almost entirely contact-center and chat roles. Text work. The layoffs tell you what AI absorbed. They tell you nothing about what happened to the queue it left behind.

That residual queue is the story. Here’s why it gets worse the better your AI performs.

Mechanism one: the selection effect

AI doesn’t deflect a random sample of your tickets. It skims the cream — the password resets, the refund statuses, the questions with answers sitting in a knowledge base. Everything a customer can fully describe and a bot can fully resolve in text.

What’s left is everything that can’t. The router that won’t sync. A medical device blinking an error pattern the customer can’t name. An installation where the customer swears they followed the instructions and something is still wrong. Deflect 60% of volume and the remaining 40% isn’t 40% as hard — it’s a concentrate. The average difficulty of a human-handled ticket rises every time your bot gets better, by definition. I’ve written before that pay-per-resolution pricing only works when your problems are text-shaped; the selection effect is the same logic running in reverse. The bots keep the text-shaped problems. Your people inherit the physical ones.

Mechanism two: context stripping

Watch what a hard case looks like by the time a human sees it. The customer typed their problem into a chatbot. First, the bot tried two knowledge-base articles. Then the customer rephrased, and the bot tried again. Then an escalation to a second AI layer, maybe a “smart” triage agent, which asked three more clarifying questions. Then, finally, a handoff.

The transcript is now forty exchanges long and contains almost no signal. Instead, it documents what the problem isn’t — every article that didn’t apply, every guess that missed. Nobody ever captured the actual problem — the physical thing happening in the customer’s hands — because text was the wrong medium from the first message. And the customer arriving at your human agent is not the customer who started the conversation. They’ve been through two loops. They’re pre-frustrated, they’ve already explained themselves three times, and they will not do it a fourth time patiently.

So your hardest cases arrive with the worst context and the worst mood. That’s not an accident. It’s what the funnel manufactures.

Mechanism three: the broken learning ladder

This one is slower and nastier. Support expertise has always grown on a ladder. Junior agents learn on easy tickets, develop pattern recognition, and graduate to hard ones. The easy tickets were never just cost — they were training data for humans.

AI just ate the bottom of the ladder. The reps who should be developing into your tier-two escalation specialists no longer get a thousand easy reps to build intuition on. Instead, they land straight in the concentrate. Meanwhile the senior people who can actually handle a gnarly physical escalation are aging out or getting cut in the same rounds Bloomberg wrote about. Five years from now, plenty of support orgs will discover they never built an escalation bench — right as escalations became the whole job.

Mechanism four: measurement fog

And through all of this, the dashboards look great. Deflection rate: up. Tickets per agent: down. Cost per contact: down, because cheap bot resolutions stuff the denominator.

But almost nobody tracks the number that matters: cost per resolved case in the residual queue. That number is climbing. Longer handle times, more escalations per case, more repeat contacts, more truck rolls dispatched because nobody could see what was actually wrong. The aggregate metric improves while the part of your operation that touches your angriest customers and your most expensive failure modes quietly deteriorates. You’re navigating by an average that mixes two populations moving in opposite directions.

The instinct that makes it worse

When leaders finally notice the residual queue hurting, the reflex is more AI. Better routing, an agentic layer, smarter summarization of those forty-exchange transcripts. But the residual queue exists precisely because these cases resisted automation. Adding another AI layer just adds another loop for the customer to escape — mechanism two, again, on a bigger budget. AWS just moved its own first generation of enterprise AI services — Q Business, Kendra, Bedrock Agents — into maintenance mode after barely two years. The first build cycle is already over. Betting the hardest layer of your support stack on the next one is not a strategy.

What actually works on the concentrate

Three moves, none of them glamorous.

Instrument the residual queue separately. Split your metrics into two populations: what AI resolves and what humans resolve. Track cost per resolved case, escalation depth, and repeat-contact rate on the human side alone. If you can’t see the residual queue as its own system, you can’t manage it.

Give escalation agents eyes. Physical, visual, ambiguous problems dominate the concentrate — the ones text already failed. The single highest-leverage intervention is letting the agent see what the customer sees: a live video session through the customer’s phone camera, no app install, at the moment of escalation. It replaces the forty-exchange transcript with thirty seconds of looking. We built Viewabo for exactly this layer. I’ve laid out the reasoning behind treating it as the default first move in why remote visual support should be the first decision, not the last resort.

Rebuild the ladder around hard cases. If easy tickets no longer train juniors, something else has to: recorded visual sessions as case libraries, deliberate shadowing on escalations, pairing juniors with senior agents on live hard cases. Treat escalation skill as something you manufacture on purpose, because the old way of manufacturing it is gone.

The companies cutting thousands of text-support roles this month aren’t wrong that AI can do that work. But every one of them is about to learn that the org they kept is now doing a different, harder job than the org they cut — and that job needs eyes, judgment, and a bench. The paradox doesn’t resolve itself. You either instrument it and equip for it, or you find out about it from your churn numbers.