A customer called three times about the same undelivered order. On the third call, an AI agent fielded it, collected the order number, confirmed the address, noted two previous failed resolutions, then escalated. The human agent picked up: "Hi, how can I help you today?"
That moment - the clean-slate handoff - is the design flaw that sits inside most AI-augmented support stacks right now. The classification problem is mostly solved. The context problem is not.
The gap AI triage actually closed
Old rule-based systems stall at around 40-50% triage accuracy. Trained AI models - ones that read the full ticket the way a senior agent would and learn from each resolution - reach 85-95% accuracy on mature deployments. That is a real gap closed. Roughly 15-25% of manually triaged tickets get reassigned at least once, and each reassignment adds about 47 minutes to resolution time.
For a team handling 2,000 tickets a month at a 35% misroute rate, that works out to over 540 extra hours a month - approaching $320,000 a year in avoidable spend.
Those numbers make the ROI case for AI triage almost trivially easy to close. Fix the routing, shrink the queue, recover the hours. Teams that have run this out have seen cases like Bolt's, where manual and rule-based escalation to engineering was producing thousands of misrouted tickets; average resolution time fell from 129.8 hours to 62.7 hours after deploying intelligent routing.
Fine. But routing accuracy is not the same as resolution quality.
What happens at the handoff
The most common failure in AI-assisted support isn't the AI getting the category wrong. It's the AI getting the category right, escalating cleanly to a human, and then dropping everything it knew about the customer in the process.
Metrics about automation rates, resolution speed, and agent hours saved don't capture what happens when AI hands a customer off to a human agent.
Those metrics matter, but they don't capture what happens at the transfer. The customer has already explained the problem. The AI has already gathered the context. Then the human picks up and starts from zero.
This is not a theoretical complaint. Zendesk's 2025 data shows 63% of customers abandon brands after one bad AI interaction. A cold handoff - AI to human, context lost - is precisely that bad interaction. The ticket gets "resolved" in the system because a human eventually handles it. CSAT tells a different story.
When the AI cannot resolve, the handoff to a human must include the full reasoning trace, customer sentiment, and attempted actions. Tickets that restart from scratch frustrate agents and customers alike.
The same problem compounds internally. When a support team has a shared Slack channel with engineering for bug escalations, tickets get duplicated between the support tool and Jira, and agents spend time writing escalation summaries that engineering re-reads. Manual or rule-based triage wastes one to two hours per ticket in back-and-forth.
The routing problem is a classification problem. The handoff problem is a memory problem.
Where the architecture breaks
Most support teams still route tickets using round-robin assignment or basic keyword rules. Round-robin distributes tickets evenly across agents regardless of skill, availability, or issue complexity. Keyword-based routing scans for words like "billing" or "refund" and drops the ticket into a matching queue.
When AI replaces those rules, the classification improves dramatically. But in many deployments, the AI still hands off a ticket as an object - an ID, a category tag, maybe a priority label - rather than as a conversation with history. The receiving agent sees the same thin context they always did. They just got it faster.
Routing is where most legacy automation stacks break. The moment your product surface area grows past a handful of features or your support org splits into more than two or three queues, rules-based routing becomes a maintenance nightmare of overlapping regex patterns and half-dead fallbacks. The result is chronic misrouting - tickets bounce between teams, agents reassign manually, and the customer waits.
The better architecture treats every ticket transfer as an opportunity to compress and transmit state, not just redirect a pointer. Every time a ticket changes hands, information is at risk. Structured handoff summaries help - and having AI generate them automatically means agents don't have to write them manually while handling the next ticket.
What good looks like
The teams getting this right are treating the escalation as its own engineering surface. Verizon treats every transfer as a data point. The system tags and tracks every handoff, feeding outcomes back into the learning algorithm. Issues that consistently require escalation get flagged for retraining. Customer signals that reliably predict the need for a human representative become a routing trigger.
That feedback loop is the part most teams skip. They deploy triage, measure routing accuracy, declare success, and move on. The escalation quality metrics - how often does a human agent have to ask the customer to repeat themselves, how often does an escalation reopen - stay invisible.
Most teams track first-response time and resolution time. Few track routing accuracy: did the ticket land with the right agent on the first assignment? Misroutes are invisible in standard reporting but account for a significant share of delayed resolutions.
The headline is consistency: a human triager's accuracy drifts with fatigue and shift handoffs, while a trained model holds 85-95% accuracy around the clock. That's why teams reaching full auto-triage redeploy their Tier-1 leads to quality review and escalation handling - the work that actually needs human judgment - instead of cutting the role.
A teammate like Beagle sits inside the Slack or Teams thread where these escalations often surface. When context lives in chat, it can travel with the escalation rather than dying at the ticketing system boundary.
The classification problem in support triage is mostly a solved problem at this point. What's left is the harder one: making sure the knowledge the AI accumulated during a conversation doesn't evaporate the moment it decides a human should take over.