Ticket routing is one of those jobs that looks solved once you automate it. A message arrives, a model reads it, a queue receives it. The whole thing happens in under 30 seconds. The problem is what happens next, when the ticket is in the wrong queue and nobody is watching.
This is the real story of AI in customer support right now - not the chatbot that handles your password reset, but the routing layer underneath it, and what breaks when that layer is wrong about where a problem belongs.
Support teams handling more than 5,000 tickets a month spend roughly 30 percent of agent time on triage alone - just reading, tagging, prioritizing, and routing. That number doesn't include actually fixing anything. It's pure overhead. So the appetite for automation here is real and justified.
Manual triage sends between 23 and 40 percent of tickets to the wrong team before they reach resolution. That's not a trivial failure rate. And 15-25 percent of manually triaged tickets get reassigned at least once, with each reassignment adding roughly 47 minutes to resolution time.
AI triage mostly improves on this. A language model reads for meaning rather than matching strings, so a customer who writes "I was charged twice" and one who writes "there's a duplicate line on my statement" both land in the same billing queue. That's genuinely useful. Basic keyword automation achieves 65-75 percent accuracy; AI triage that understands natural language and intent reaches 90-95 percent, requiring human review on only a fraction of tickets.
But that remaining fraction is where things get expensive.
The cost of getting triage wrong isn't in the misroute. It's in the delay before anyone notices the misroute.
A misrouted billing ticket sits in a tier-1 queue while the customer churns. A high-priority outage report gets filed under "general inquiry" because the model never saw "the API returns 502." A VIP customer waits four hours because no one tagged the account tier on intake. Each of these is a silent failure. The ticket isn't lost - it's just in the wrong place, looking fine from the outside.
The escalation handoff makes it worse. Most enterprises track containment or deflection, but few track what happens to escalated conversations - whether the agent resolved the issue, how long it took, whether the customer repeated themselves. Without that loop, the AI learns nothing from its failures, and the escalation layer stays broken.
Accenture finds that 81 percent of customers want a smooth AI-to-human handoff, yet only 44 percent experience it today. The gap is structural. CRM integration gaps mean agents don't see customer history when calls transfer. Knowledge base disconnection means handoff summaries don't include what solutions were already attempted. Ticketing system silos mean case creation happens after the transfer rather than during, losing context in the transition.
The agent inheriting an escalated ticket often starts cold, which means the customer has to re-explain themselves. That experience - telling your story twice, to a human who should already know it - is where trust quietly breaks.
Klarna is the case study nobody should ignore
In early 2024, Klarna launched an AI customer service assistant in partnership with OpenAI that handled 2.3 million chats in its first month, covering two-thirds of all customer conversations, cutting resolution time from 11 minutes to under 2 minutes. The numbers were striking. The headlines were bigger.
But for simple queries - order status, payment schedules - AI matched human performance. For complex disputes, fraud claims, and hardship cases, AI resolution quality dropped noticeably. By May 2025, Klarna CEO Sebastian Siemiatkowski was reversing an AI-induced hiring freeze to bring on more human staff, telling Bloomberg that customers should always have a human presence to talk to if needed.
Siemiatkowski publicly admitted that the AI-driven transition had negatively affected service and product quality. The specific failure mode wasn't the AI being wrong on simple questions. It was the AI being routed to complex questions it wasn't equipped for - and having no clean way to hand those off. Customers reportedly had to re-explain themselves on escalated tickets, which tanked CSAT on exactly the cases that needed the most care.
Where the actual fix lives
NLP-based triage is still classification-only - it can't reason across systems, take action, or understand the full context of who the customer is or what's already known about their issue. That's a real constraint, and it matters most in two situations: when the ticket is ambiguous, and when the customer is already frustrated.
Sentiment analysis can distinguish a password reset request from a calm user from a password reset request from someone who has failed login six times and is now typing in all caps. Those should route differently. Most systems don't do this by default. It has to be configured deliberately.
The human inheriting the conversation should get the full transcript, the customer's account state, and the AI's reasoning about why it escalated. Without those, escalations damage the experience even when they're the right call.
A teammate like Beagle, sitting in the Slack channel where an escalated ticket lands, can surface that context - the thread history, the prior attempts, the account tier - so the agent starts informed rather than cold.
The teams that get this right treat the handoff as a first-class concern, not an afterthought. They define clear confidence thresholds. They audit a sample of AI-handled conversations every week. They watch the metrics most teams ignore: rage clicks, where a customer repeatedly submits the same message; and conversation depth before escalation, because if customers consistently need 8-12 messages to reach a human, the AI isn't filtering - it's blocking.
The question worth sitting with isn't whether AI should triage support tickets. It should, and it will. The question is whether your escalation path is good enough that the 5-10 percent of hard cases don't undo the goodwill the easy 90 percent earned you.
The shift highlights the need for customers to always have an option to speak with a human - and to use AI as a supplement, not a replacement, for staff. That framing doesn't make the technology less useful. It just makes it honest about where the limits are.
The ticket that lands on the wrong desk isn't a technology problem. It's a design problem. The routing model is the easy part. Knowing what to do when it fails - and making sure a real person gets the full picture when it does - is the work that's still mostly undone.