AI for Customer Support, Success & Experience
Support is where AI agents are most established. They answer from your help content, take actions such as refunds within limits, and work in chat, email and voice. Human agents move to the hard cases, with AI drafting beside them.
How AI is changing this work
- First contact is increasingly an AI agent that resolves the issue, not a form that deflects it.
- Actions such as refunds, bookings and account changes are done by the AI under written rules.
- Voice agents now hold natural phone conversations, replacing menu trees.
Your layers: the 4 that matter most
1. Layer 9 · Applications & Interfaces
Why it matters
The customer meets your AI here, in chat, email or on the phone. Tone, speed and the path to a person are the experience.
What to do here
- AI agent on every channel. One agent answers in chat, email and voice from the same knowledge and rules. Chat assistants
- Voice that listens. Spoken agents handle interruptions and accents, and replace phone menus. Voice & multimodal interfaces
- Assist for human agents. Suggested replies, summaries and next steps appear beside the ticket. Copilots
Examples as of 2026-11-11
2. Layer 8 · Agents
Why it matters
Answering is not resolving. Resolution needs an agent that can look up the order, apply the policy and act, within limits you set.
What to do here
- Procedures in plain language. Write each workflow, such as refunds under 30 days, as steps the agent follows and engineers can test. The agent loop
- Limits on actions. Cap amounts, require confirmation for account changes, and log every action. Agent permissions & sandboxing
- Clean handoff. Escalate on request, on repeated failure or on high value, passing the full history to the person. The agent loop
Examples as of 2026-11-11
3. Layer 7 · Orchestration & Tooling
Why it matters
The agent is only as right as your help content and as safe as its guardrails. Customers will try to talk it into things.
What to do here
- Answer from your knowledge base. Ground every answer in current help articles and policies, and fix the article when the answer is wrong. Retrieval-augmented generation (RAG)
- Test with real tickets. Replay past conversations before launch and after every change, and score the results. Guardrails & evals
- Expect manipulation. Customers and web content can try to override its rules. Screen inputs and keep limits outside the model. Prompt injection
Examples as of 2026-11-11
4. Layer 6 · Inference & Optimization
Why it matters
Support runs at volume and in real time. Cost per conversation and the pause before a spoken reply are both decided by how the model is served.
What to do here
- Speed for voice. A phone conversation needs replies in well under a second; choose models and hosting for that. Model APIs & routers
- Small model first. Send simple questions to a small fast model and only hard ones to a large one. Model APIs & routers
- Cache the fixed part. Policies and instructions repeat in every conversation; caching them cuts cost. Batching & KV caching
Examples as of 2026-11-11
High-leverage use cases
- Resolve common questions. Order status, how-to and policy questions answered at once, with a source.
- Take routine actions. Refunds, returns, bookings and address changes within written limits.
- Agent assist. Draft replies and summaries for human agents, which helps new staff most.
- Voice front door. A spoken agent that understands the request and resolves or routes it.
- Insight from conversations. Themes and product issues surfaced from all contacts, not a sample.
Start here: the next 30 to 60 days
- Week 1 to 2. Find your ten most common questions and check the help articles for each are current and clear.
- Week 3 to 4. Launch an AI agent on one channel for those topics, with an easy route to a person.
- Day 30 to 60. Review 50 AI conversations a week. Fix content first, rules second.
- Day 30 to 60. Add one action, such as order lookup or a capped refund, with logging and limits.
Risks and pitfalls
- You own what it says. A Canadian tribunal held an airline responsible for wrong information its website chatbot gave a customer. (Moffatt v. Air Canada: a misrepresentation by an AI chatbot (McCarthy Tétrault, on 2024 BCCRT 149))
- Telling customers it is AI. In the EU, people must be told when they are talking to an AI unless it is obvious. (Regulation (EU) 2024/1689 (the AI Act))
- No way out. Customers trapped with a bot leave. Always offer a person, and escalate on frustration.
- Talked into exceptions. Customers will test whether the agent can be argued into refunds or discounts. Enforce limits in the system, not in the prompt. (OWASP Top 10 for LLM Applications)
- Measuring deflection, not resolution. A closed chat is not a solved problem. Track repeat contacts and satisfaction.
Sources
A product named on this page is an example, not a recommendation.
- Fin (Intercom) (fin.ai) (2026-11-11)
- Zendesk AI agents (zendesk.com) (2026-11-11)
- Customer Agent (HubSpot) (hubspot.com) (2026-11-11)
- Sierra (sierra.ai) (2026-11-11)
- Decagon (decagon.ai) (2026-11-11)
- Glean (glean.com) (2026-11-11)
- Prompt Shields (Azure AI Content Safety) (learn.microsoft.com) (2026-11-11)
- OpenAI Realtime API (developers.openai.com) (2026-11-11)
- ElevenLabs Agents (elevenlabs.io) (2026-11-11)
- Prompt caching (Claude API) (platform.claude.com) (2026-11-11)
- Moffatt v. Air Canada: a misrepresentation by an AI chatbot (McCarthy Tétrault, on 2024 BCCRT 149) (2024-02)
- Regulation (EU) 2024/1689 (the AI Act) (2024)
- OWASP Top 10 for LLM Applications (2026)
- Generative AI at Work (Brynjolfsson, Li, Raymond) (2023-04-23)
Other roles
HR & People · Finance · Sales · Marketing · Product & Design · Engineering · Cybersecurity · Legal & Compliance · Operations