AI Customer Support — AI Glossary

Reference details

Related terms

• Conversational AI
• RAG (Retrieval-Augmented Generation)
• Tool-Using Agent

How it works

1. Intent and scope are determined first: what the system will attempt and what it hands straight to a person. That boundary is a product decision and the main determinant of whether the deployment is a success.
2. Answers are grounded in retrieval over product documentation and policy, because a support answer that is not traceable to a source is a liability rather than a saving.
3. Tool integrations do the work that documentation cannot — order lookup, account state, subscription changes — and each one is also a permission decision about what the system may do on a customer's behalf.
4. Escalation is a designed path rather than a fallback, carrying the conversation, the retrieved sources and what was already attempted, so the human does not restart from nothing.
5. Confident wrong answers are the failure that matters, which makes abstention — declining and routing on low retrieval confidence — the single most valuable behaviour to tune.
6. Measurement must follow the ticket rather than the session, because a conversation that ends is not a problem that is solved and the difference between those two is where reported deflection rates go wrong.

Example: Why a 50% deflection rate is not a 50% saving

Take 10,000 tickets a month and suppose the system resolves half without a human: 10,000 x 0.5 = 5,000 deflected. Now suppose 10% of those were not actually resolved — the customer accepted an answer that was wrong, or gave up and returned. That is 5,000 x 0.1 = 500 tickets coming back, and a returning customer who was already given a wrong answer costs more than a fresh one, since the agent must undo the previous answer as well as solve the problem. At twice the handling cost, those 500 consume 500 x 2 = 1,000 ticket-equivalents. The net saving is 5,000 - 1,000 = 4,000 ticket-equivalents, not 5,000 — the headline deflection overstates the benefit by 25%. The sensitivity is what matters for design: at a 20% wrong-resolution rate the arithmetic becomes 5,000 - (1,000 x 2) = 3,000, so doubling the error rate costs a quarter of the remaining benefit. This is why abstention is worth more than coverage. A system that declines 10% more often and is wrong 10% less often reports a worse deflection rate and saves more money.

Commonly confused with

Deflection and resolution are different measurements, and support dashboards frequently report the first as the second. Deflection counts conversations that did not reach a human, which includes customers who gave up, were satisfied with a wrong answer, or went to another channel. Resolution counts problems actually solved, which requires following the ticket afterwards. The gap between them is populated entirely by failures, so a system optimised on deflection is optimised partly on customers abandoning the channel — and that is the metric most likely to look excellent while satisfaction falls.

When to use it

Reach for it when:

• High-volume, well-documented, repetitive inquiries, where retrieval has real sources and the questions recur.
• Account and order operations with safe tool integrations, where the answer requires live data rather than a document.
• As a first tier with a designed escalation path, which is the configuration that reliably reduces load.
• Where abstention is tuned deliberately, since declining is cheaper than being confidently wrong.

Reach for something else when:

• Complaints, cancellations and anything emotionally charged, where the interaction is the service and automation reads as evasion.
• Domains with regulatory or safety consequences for a wrong answer, absent human review in the path.
• Thin or stale documentation, since the system's answers inherit the source material's quality exactly.
• Measured on deflection, which counts abandonment as success and will look best when it is working worst.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to AI Customer Support — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path