Routing is the AI work that pays for itself fastest. It touches every conversation, including the ones the AI can't resolve. Even when the customer ends up with a human, routing well means the right human, with the right context, in less time than the team's average reassignment loop.
The trouble is that routing in Intercom has historically been built on rules. Rule-based systems plateau around 40 to 50% classification accuracy on real-world ticket queues. AI triage systems average around 89% accuracy on mature implementations. That gap is the case for moving the routing layer from "if-then" to "interpret-and-decide."
This piece is the practical version: what Intercom does natively, where AI changes the game, and how to build an Intercom ticket routing AI workflow that survives contact with reality.
TL;DR
- Routing is the highest-ROI AI work because it compounds across every ticket, resolved or not.
- Native Intercom workflows handle rule-based routing well. AI changes the game on intent, urgency, sentiment, and account-aware routing.
- Build the triage flow in five steps: capture, classify, enrich, decide, hand off.
- Track routing accuracy, time-to-correct-assignment, and reassignment rate. Speed alone is misleading.
- Carve out attribute-based exceptions for high-LTV accounts and compliance topics. They don't get triaged by the AI.
Why routing beats resolution as the first AI win
When teams plan AI automation on Intercom, the conversation usually centers on deflection (which tickets the AI can close without a human). Routing gets less attention because it's invisible to the customer.
The math runs the other way.
A 50% deflection rate saves human time on half the tickets. A well-built AI triage layer improves time-to-resolution on every ticket, including the half the AI can't deflect. The human who picks up an AI-routed ticket gets the right assignment, the right context, and a summary of what was tried. Their handle time drops by a meaningful margin even without the AI ever attempting resolution.
For teams whose knowledge base isn't ready for full deflection yet, AI routing is the work that pays back first.
OTO automates 77% of customer support across a high-volume retail operation with a 90%+ CSAT. Even before the deflection number climbed, the routing layer cut handle time on the half OTO's team still owned.
What Intercom does natively
The Intercom workflow builder ships strong primitives for rule-based routing. Worth knowing the floor before you build on top of it.
Rule-based vs AI-augmented routing
Rules cap around 40–50% classification accuracy on phrasing variation. AI triage averages 89% on mature deployments.
Inbox rules. Assign conversations to teams based on attributes: channel, subject, customer tier, business hours, conversation source. Fast to configure, deterministic, brittle when phrasing varies.
Workflows. Multi-step bots that can ask the customer questions, collect structured data, and branch routing on the answers. Intercom documents workflows in detail. Workflows handle structured intake well; they struggle on open-text questions where intent isn't obvious from the customer's first sentence.
Round-robin and load balancing. Distribute assigned conversations across a team. Works well as the final layer after routing decisions are made.
Custom attributes. Pass account data, plan tier, lifetime value, prior conversation history into the routing decision. The native engine respects attributes; the AI layer uses them.
The natural ceiling: rule-based routing breaks on phrasing variation. A customer who says "I want my money back" and a customer who says "this was charged twice, please fix it" are asking the same question. A rule looking for "refund" catches one and misses the other. An AI classifier reads the intent.
Where AI changes the game
Four routing decisions get markedly better with an AI layer.
Intent classification. The AI reads the conversation and assigns an intent label (billing, technical, cancellation, complaint) regardless of phrasing variation. Pattern-matching and RAG-based platforms typically report 60 to 80% classification accuracy; reasoning-first platforms push higher. Either is meaningfully above rule-based ceilings.
Urgency detection. The AI reads signals (account tier, language, sentiment, time-sensitivity cues like "outage" or "production") and assigns priority. Rule-based urgency systems require the customer to declare urgency explicitly, which most don't.
Sentiment-based routing. Customers in distress get routed to senior agents or a manager. Routine questions go to tier-1. Sentiment classification is one of the most reliable AI capabilities and it changes the customer experience materially.
Account-aware routing. The AI combines intent with account context (plan tier, LTV, open issues, churn risk) to make a routing decision that a rule couldn't. A churn-risk customer asking a billing question gets routed differently from a happy customer asking the same question.
The integration pattern: AI sits in front of native routing. Conversation comes in, AI classifies and enriches, the result drops back into the Intercom workflow which makes the deterministic routing decision. Both layers do what they're good at.
How to build the triage workflow (five steps)
A reliable AI triage workflow on Intercom has five stages. Build them in order.
Routing decision matrix
AI tags (intent, urgency, sentiment) feed a deterministic routing rule.
Step 1: Capture
The conversation comes in via Messenger, email, WhatsApp, or another channel. The capture step's job is to grab everything the downstream stages will need: customer message, conversation history, customer profile attributes, channel metadata.
Intercom does this natively. The workflow's trigger step gives access to conversation context and customer attributes; no extra plumbing is required.
Step 2: Classify
Hand the message to the AI for intent classification. If you're using Fin, this happens inside the "Let Fin handle" step. If you're using a third-party AI agent on top of Intercom, the classification call happens via the agent's API or app integration.
The output of classify is a structured set of tags:
- Intent: e.g.,
billing.refund,technical.outage,cancellation.account - Urgency:
high | medium | low - Sentiment:
positive | neutral | negative | distressed - Topic confidence: a number between 0 and 1 indicating how sure the model is
If confidence is below a threshold (typically 0.7), branch to a clarifying question step rather than routing on a guess.
Step 3: Enrich
Pull account-side context that informs the routing decision. The Intercom workflow can read custom attributes natively. For third-party agents, this is usually an API lookup to your CRM or product database.
What to pull:
- Plan tier and customer status
- Open issues / conversations in the last 30 days
- Account health signal (NPS, churn risk score, engagement)
- Recent product usage (logged in today, hasn't logged in for 90 days)
- VIP / high-LTV flag
Enrichment is the difference between routing every customer the same way and routing them in context.
Step 4: Decide
A routing decision matrix. The classify and enrich outputs feed a deterministic routing rule:
| Intent | Urgency | Account tier | Route to |
|---|---|---|---|
billing.refund | high | top 10% LTV | Senior support, escalation queue |
billing.refund | medium | mid-market | Billing team |
technical.outage | high | any | Engineering on-call |
cancellation.account | any | top 10% LTV | Retention team (always) |
complaint.distressed | any | any | Senior agent, sentiment flagged |
You can build this matrix as a branching workflow in Intercom or as a single decision step that takes the AI's tags as inputs.
Step 5: Hand off
The routing decision lands on a team or specific agent. The handoff carries:
- Conversation history
- Intent and urgency labels (visible on the ticket)
- Enriched account context (attached as conversation attributes)
- A one-line summary of the issue from the AI
The human picks up a ticket with a clean briefing instead of starting from scratch.
The summarization layer
The summarization step is the part that makes the routing layer feel like a different product to the human agent.
A good AI summary, attached to the conversation at the moment of handoff:
Customer Sarah K., Pro plan, 18 months tenure, NPS 9. Asking about a duplicate charge on May 8 for $99. Fin verified the charge in Stripe but cannot process the refund because the account is flagged for manual review. Customer is calm but stated this is the second time this has happened. Recommend prioritizing as VIP retention.
That summary changes the human's first message from "let me look into your account" to "Hi Sarah, I see the duplicate charge from May 8. Refunding it now and pulling the manual-review flag so this stops happening." Time-to-meaningful-response drops by minutes.
Fin can generate this summary natively. Third-party agents expose it through Intercom's conversation attributes. Whichever you use, treat the summary as part of the routing output, not a separate feature.
Measuring routing quality
Speed is the easy number to measure. It's also the misleading one. Four metrics give a real read on routing quality.
Routing accuracy. Of the AI's routing decisions, how many landed on the right team without a reassignment? The benchmark is 77% for industry routing accuracy, with AI-driven systems reaching 85%+. If your accuracy is below 70%, the classifier needs more training data or the rule layer needs refinement.
Reassignment rate. How many conversations get bounced from the initially routed team to another team? High reassignment is a sign the routing decision is wrong or the team's scope is undefined.
Time-to-correct-assignment. Total time from conversation start to the right human picking it up. Includes both the AI's classify+enrich time and the queue time before the human reads it.
Handle time on AI-routed vs. non-AI-routed tickets. The clearest read on whether the routing layer is producing value. AI-routed tickets should handle 15 to 25% faster than the baseline once the summary layer is in place.
Edge cases
A few situations where the standard routing flow breaks down, and how to handle them.
Multi-issue tickets. A customer who asks two unrelated questions in one conversation. The AI should pick the higher-priority intent and route, while flagging the second issue in the summary so the agent can address both.
Ambiguous intent. Confidence below threshold. Either ask a clarifying question (preferred) or route to a general queue with a "low confidence" tag so the agent knows the AI couldn't classify confidently.
VIP override. Top-N% LTV accounts can be routed straight to a senior agent without AI triage. Use an attribute-based bypass at the start of the workflow so the AI doesn't get a turn on these conversations.
Compliance and legal. Fraud, legal threats, regulatory edge cases. These get routed to a specialist queue immediately, no AI handling. The classify step still runs (so you have the labels) but the decide step bypasses the AI for these intents.
Returning customer with an open conversation. Don't route a new conversation when there's an active one with the same customer on a related topic. Merge or assign to the same agent. Intercom handles this natively with conversation rules.
What changes when routing is built right
The team feels it on the human side first. Reassignment rate drops. Handle time on the half that still goes to humans drops faster than the deflection rate climbs. CSAT on AI-routed-then-human-handled tickets often beats CSAT on AI-routed-then-AI-resolved tickets, because the customer gets a fast accurate response from a human who already knows what's going on.
The cost story compounds from there. Every ticket touched by good routing is a ticket the team handled with less context-switching, fewer cold starts, and less judgment-burning on routing decisions a machine can make. The savings show up in CSAT, retention, and a calmer support team.