Resources
AI customer support: a practical guide
What it means in production, where the trade-offs actually live, and how to evaluate vendors without falling for marketing copy.
Two surfaces, one operating model
AI customer support comes in two shapes. AI Copilot assists humans with drafts, summaries, and next-step suggestions. AI Agent acts autonomously when confidence is high enough. Most production teams run both: Copilot speeds up the work humans handle, Agent removes routine work entirely.
The four layers that make AI safe to deploy
Vendors who skip these layers ship demos. Vendors who build them ship production systems.
- Confidence thresholds. Each AI reply has a confidence score. The system uses it to decide whether to send autonomously or escalate. Configurable per channel and ticket type.
- Approval gates. For sensitive categories like refunds over a threshold, AI drafts but does not commit. A human approves before the action runs.
- Audit logs. Every reply, action, and source citation is logged per ticket. Reviewers can reconstruct exactly what happened and why.
- Bounded tools. AI can only call a fixed list of pre-approved actions. The surface area of mistakes equals the surface area of permitted actions.
Pricing models
AI pricing has converged on three shapes:
- Per seat. AI billed per user per month. Punishes small efficient teams. Most legacy vendors use this.
- Per attempt. AI billed per interaction whether it works or not. Punishes you for cases where AI escalates.
- Per outcome. AI billed only when it fully resolves a ticket. Aligns vendor incentive with buyer outcomes. Support Oasis uses this model at $0.50 per resolved ticket.
How to evaluate a vendor in 30 minutes
Five questions that separate marketing copy from working systems:
- Can I see the confidence score for every AI reply?
- Can I require human approval for specific ticket types regardless of confidence?
- Can I see a per-ticket audit log of what AI did and why?
- What is the enumerated list of actions the AI is allowed to take?
- How is AI billed: per seat, per attempt, or per outcome?
Where AI shines and where it does not
Best fit categories: order status and tracking, in-policy refunds and replacements, password resets, sizing and product questions, basic account changes, meeting scheduling. Categories that need a human: complex troubleshooting, policy exceptions, sensitive personal information, anything legal-adjacent, customer escalations and complaints.
What to test before you commit
Connect the AI to your actual help center. Read 20 real drafts on real tickets. The drafts tell you everything: tone, accuracy, hallucination rate, source citations, escalation behavior. Run the cost calculator with your real volume to see the unit economics. Then decide.