AI Customer Support Agents That Actually Resolve Tickets

Deflection Was Never the Goal

The chatbots of 2018-2023 were deflection machines: FAQ retrieval with a conversational veneer, measured on "containment" — a metric that counts a frustrated customer giving up as success. Customers learned to type "agent" immediately, and they were right to.

What changed is tool use. A modern AI agent doesn't just answer questions about your refund policy — it can look up the order, check the return window, issue the refund, and email the confirmation. Resolution, not deflection. Across our deployments, well-scoped agents resolve 40-70% of inbound volume end-to-end, with customer satisfaction scores matching or beating the human baseline on those categories.

The Architecture: Answers Are Easy, Actions Are the Point

A production support agent has four layers:

• Knowledge: retrieval over your help centre, policies and product documentation — built with the disciplines in our RAG in production guide, so answers cite sources and refuse gracefully when unsure.
• Tools: authenticated, narrowly-scoped API access to the systems where resolution actually happens — order management, billing, CRM, identity. Each tool does one thing: get_order, issue_refund(order_id, amount ≤ policy_limit), update_address. (We build these as MCP tool servers — see our MCP architecture guide.)
• Policy: the rules engine that decides what the agent may do autonomously, what needs customer confirmation, and what routes to a human. Refund up to £50: autonomous. £50-£250: customer confirms. Above, or any hint of fraud: human.
• Escalation: a first-class handover, not a failure state — full conversation summary, entities extracted, actions already taken, sentiment flag, routed to the right queue. Human agents describe good handovers as "arriving mid-conversation with a briefing".

Guardrails That Earn the Right to Act

• Hard authorisation outside the model. The model requests actions; a deterministic policy layer approves them. Prompt injection can make a model want to refund £10,000 — it cannot make the policy layer allow it.
• Identity before action. Read-only answers may be low-friction; anything that changes state requires verified customer identity via your existing auth.
• Full audit trail. Every tool call logged with inputs, outputs and conversation context. When finance asks why a refund happened, the answer is one query away.
• Honest uncertainty. An agent that says "I'm not certain — let me get a colleague" preserves trust. One that guesses confidently destroys it, once, permanently.

Measure resolution, not containment. A ticket is resolved when the customer's problem is fixed and they don't come back about it — anything else is theatre with a transcript.

The Rollout That Doesn't Burn Trust

Never launch on 100% of traffic. The sequence that works: shadow mode (agent drafts, humans send) for two weeks to measure quality on real conversations; then autonomous on the top 2-3 intent categories with tight action limits; then expand category by category as resolution and CSAT data earn it. Weekly review of escalations and edge cases feeds the knowledge base — the agent improves on a cadence, not by accident.

Honest Economics

For a team handling 5,000+ tickets a month, the maths is straightforward: at 50% end-to-end resolution and a £3-6 fully-loaded cost per human-handled ticket, an agent typically pays back its build cost inside two quarters — while cutting first-response time from hours to seconds on the automated share. The teams that win don't cut headcount; they redeploy humans to the complex, revenue-touching conversations that deserve them.