Support is the most common first AI project in any organization, and the most commonly botched. The reason is that teams start with the visible thing — a customer-facing bot — instead of the profitable thing.
Automate in this order
1. Classification and routing. Every inbound message gets a category, priority, sentiment, and owning team. Errors are cheap, volume is high, and the manual baseline is easy to measure. This alone often removes a full triage role's worth of work.
2. Draft responses for agents. The model writes a reply grounded in your help centre and the customer's history; the agent reviews, edits, sends. Handle time drops 30–50% in the deployments we have measured, with no customer-facing risk because a human always ships the message.
3. Knowledge retrieval for agents. An internal search that actually answers, over your docs, past tickets, and policies.
4. Autonomous resolution — narrow intents only. Order status, password reset, appointment changes, tracking. Things with a clean API call behind them and an unambiguous correct answer.
5. Everything else. Later, or never.
Most teams invert this and start at step 4 on all intents at once. That is why the bot gets pulled after a bad week on social media.
The pipeline
A support automation pipeline has a predictable shape:
- Ingest from email, chat, web form, and phone transcript.
- Normalize — strip signatures and quoted history, detect language, extract order numbers and account IDs.
- Enrich — join the customer record, entitlement, order history, and prior tickets.
- Classify intent, urgency, and sentiment with a small fast model.
- Retrieve relevant help-centre content and similar resolved tickets (RAG).
- Generate a draft reply or take the action, depending on intent and confidence.
- Route — auto-send, agent-review queue, or escalate.
- Log everything for evaluation.
Where it actually goes wrong
Deflection measured as containment. A bot that ends the conversation is counted as a success even when the customer gave up and emailed instead. Measure resolution — did the customer's problem get solved without a human — and track re-contact within 72 hours. Containment rates flatter; re-contact rates tell the truth.
Stale knowledge base. RAG retrieves whatever is written down. If your help centre is 18 months out of date, the bot confidently repeats 18-month-old policy. Fixing the knowledge base is usually a prerequisite, not a nice-to-have.
No graceful handoff. When the model cannot help, the customer must reach a human without restarting from scratch. The transcript, the detected intent, and everything already collected must travel with them.
Tone that does not match the brand. Generic model voice on a premium product reads as cheapness. This is a legitimate use for fine-tuning or, more cheaply, a carefully engineered prompt with real examples.
Realistic numbers
| Stage | Typical build time | Where value shows up |
|---|---|---|
| Classification + routing | 3–5 weeks | Triage headcount, faster first response |
| Agent draft replies | 4–8 weeks | Handle time, consistency, onboarding speed |
| Agent knowledge search | 4–6 weeks | Escalation rate, agent ramp |
| Narrow autonomous intents | 8–12 weeks | True deflection on 10–25% of volume |
Autonomous resolution on 10–25% of total volume is a strong, defensible outcome. Vendors quoting 60–80% deflection are almost always measuring containment.
Frequently asked questions
Will AI replace our support agents?
It reliably replaces triage and first-draft writing, not judgement or de-escalation. The pattern we see is flat headcount with substantially higher volume handled, plus agents moving up to complex and retention-critical work. Teams that cut headcount first and automate second usually end up rehiring.
How accurate does classification need to be?
For routing, 90% accuracy with a low-confidence queue is generally sufficient — misroutes are recoverable and cheap. For anything that triggers an irreversible action, such as issuing a refund, set the bar far higher and require confidence thresholds plus a spend cap.
Should the bot tell customers it is AI?
Yes, and increasingly you must. Several jurisdictions now require disclosure, and customers overwhelmingly react worse to discovering it late than to being told up front. Disclosure also measurably reduces frustration because it calibrates expectations.
What about phone support?
Voice adds transcription latency and error on top of everything else, so it lags text by a year or so in maturity. The reliable wins today are post-call: automatic summarization, disposition coding, and quality scoring — all of which remove real minutes per call without touching the live conversation.
How do we stop it from making things up?
Ground every response in retrieved content, instruct the model to answer only from that content, and have it explicitly say when it does not know. Then measure hallucination rate on a held-out evaluation set rather than assuming the prompt worked.