Ticket Triage Automation: What Production Deployments Show

Yes — with the cleanest human-versus-machine comparison on this site. Wayfair's supplier-ticket system identifies question type at 93% accuracy against a measured 75% for human agents on the same task, with language identification at 98% and supplier matching at 88%. Miro's bug-routing system reached over 75% top-1 accuracy and 95% top-3 across nearly a hundred engineering teams — a 70% improvement over their prior NLP approach — producing a six-fold reduction in team reassignments and a five-fold improvement in median time-to-resolution, from days to hours, at 53 seconds of classification latency per ticket. The baseline they were escaping had a price tag attached: misrouted bugs were costing an estimated 42 years of cumulative lost productivity annually.

The quieter wins compound the routing: Linear uses embeddings to surface similar and duplicate issues so the same problem doesn't get triaged five times; Adyen's smart routing feeds a copilot that made support agents measurably more efficient.

Read the mechanism honestly: triage is a decision made thousands of times a day under fatigue, and consistency is exactly what machines are for. The AI doesn't out-think the best triager on their best day — it out-consists them across every ticket, every day, in every language at once.

The organisation moves, and models that memorised it break. Miro's documented predecessor is the canonical case: a fine-tuned GPT model whose performance degraded quickly and demanded retraining whenever teams merged or responsibilities changed — impractical for a dynamic engineering organisation, and retired for exactly that reason. The replacement classifies against current organisational context at inference time instead of baking the org chart into weights, which is why it survives reorganisations that killed its predecessor. The design question for any triage system is therefore not opening accuracy but decay slope: what happens the week after the next restructure.

The second documented failure is agentic overreach. Wayfair's team tried a fully agentic design — a supervisor coordinating multiple sub-agents — and watched it fail on inter-agent communication errors and unnecessary repeated calls; in one case an agent simply decided not to invoke the ticket-update API, which in a triage system is the one job. The hybrid that shipped keeps deterministic structure around the model calls. Linear's build hit the infrastructure version of the same lesson: embeddings stored as database blobs worked in prototyping and posed performance risks at scale, with naive similarity queries timing out on tens of millions of issues. The pattern across all three: triage rewards boring architecture — current context in, one decision out, deterministic writes — and punishes cleverness in direct proportion to its sophistication.

Split by what the taxonomy is made of. If your triage categories are standard support taxonomy — billing, technical, account, urgency tiers — the helpdesk-native and platform routing this record documents (Adyen's LangChain-based smart routing among them) is bought capability, and buying it is sound. If your taxonomy is your organisation — a hundred engineering teams, a product graph, supplier relationships — the record's strongest systems are builds, because no vendor ships your org chart, and the moving-target problem above is yours to engineer for regardless of who sells the model.

The build rationale gets its most candid statement in this record from incident.io, who rejected available off-the-shelf tooling with reasoning worth quoting in any build-versus-buy meeting: relying on vendor marketing rather than first-hand experience risked adopting a product built for a different team's context — and would have cost them learning AI engineering from first principles. That's not anti-vendor sentiment; it's a team whose triage problem was genuinely theirs, choosing to own it.

Either path, the operational requirements converge: ownership data that reflects today's organisation (wire the HR and team systems in, don't snapshot them), a confidence threshold with a staffed human lane, and the reassignment feedback loop treated as the system's food supply — because in triage, every human correction is free training data, and the deployments that compound are the ones that eat it.