The hardest thing in AI support isn't getting the answer right. It's knowing when to shut up.
A wrong answer, customers forgive. A confidently wrong answer that drags the conversation further off the rails for five more turns — that's the one that churns them. So the handoff part of "AI first, human fallback" matters more than the AI part.
Here are four patterns we've shipped and watched our customers ship. Each has its fit. Each has its failure mode.
Pattern 1: Confidence-threshold handoff
The simplest version. Every AI reply carries a confidence score. Drop below a threshold (say 0.7), auto-escalate.
Good fit: FAQ-heavy products. SaaS tools, online learning, docs-driven support. High KB coverage, and when a long-tail question shows up the AI basically admits it's uncertain.
Bad fit: E-commerce post-purchase. User messages tend to be multi-intent stews — "my order from yesterday hasn't shipped also does this product work with a Model A." The AI's confidence per sentence looks fine, but the stitched-together answer is wrong.
Pattern 2: Skill-tag routing
Tag the conversation (refund, technical, sales) and route each tag to the matching agent group. The AI handles the first response and any tag it can actually close, then hands off the rest.
Good fit: Teams with clear functional splits. Billing questions to Billing, integration questions to Dev Relations, etc.
Bad fit: Three-person teams where one person does everything. Skill-tags add routing errors without adding throughput — the AI will occasionally flag something it could have answered as "needs a specialist" and burn a handoff for no reason.
Pattern 3: User-initiated escalation
The most underrated one. Always put a "talk to a human" button under the AI's reply, and listen for keywords like "agent," "real person," "human."
Good fit: Basically every consumer product. Users' patience for AI is shorter than you think. Give them a clean exit and they'll actually try the AI first, because they know they can bail.
Bad fit: Almost none. Only rule — don't show the button before the user even asks their first question. It looks like the AI doesn't want to help.
Pattern 4: High-value customer bypass
Tag-based skip. VIP, high-ARR, renewal window — straight to a human, no AI gate.
Good fit: B2B and premium DTC with a real customer tier. A 5-figure annual contract shouldn't sit in the same AI queue as a free trial.
Bad fit: When "everyone is VIP." Usually that means you haven't actually tiered, you just feel bad serving AI to people.
How to actually combine these
Don't pick one. Stack them: filter by customer value first (Pattern 4), route by skill (Pattern 2), fall back on confidence when the AI flounders (Pattern 1), and always leave the escape hatch open (Pattern 3).
One more thing: the handoff must carry context. If the AI ran eight turns and then escalated, the agent's first screen should be what the AI asked, what the user answered, and where it got stuck. Making the agent re-interview the customer from scratch is the worst experience in the whole chain.