AI customer support, explained without the hype
Most "AI customer support" is a chatbot bolted onto an old helpdesk, priced per resolution so your bill climbs every time it works. Real AI support is different: the AI is the product. It reads the ticket, grounds an answer in your own documentation, and either resolves it or hands you a clean decision when it should not act alone. This guide explains how that works, where it fails, and what separates a support agent you can trust from one that invents a refund policy on a Tuesday.
What "AI customer support" really means
The term covers three very different things. The weakest is a deflection widget that searches your help center and shows articles, which mostly frustrates customers who already read those articles. The middle tier is a scripted chatbot with decision trees. The real thing is an AI agent that understands intent, retrieves the relevant facts from your live content, drafts a specific answer, and takes the action that closes the ticket.
Resolve is the third kind. Reva, the AI agent, resolves the issue end to end and cites the source it used, so the answer is not a confident guess. When it is not cleared to act alone, it does not send blind and it does not give up. It produces a decision card for a human.
How auto-resolution works
A resolution has four steps, and the order matters:
- Read the whole ticket, including account tier, history, entitlement, and sentiment. An answer that is correct in general but wrong for this customer reads as a lie.
- Retrieve or refuse. The agent pulls the most relevant content from your real docs. If nothing solid comes back, it caps its confidence and escalates rather than inventing an answer.
- Draft and verify. A separate supervisor model checks the draft against the retrieved facts before anything ships. Business logic like refund amounts and return windows is computed in code, never authored by the language model.
- Resolve or hand up. If confidence clears the bar, the agent sends and closes. If not, it escalates on a decision card with its reasoning and sources attached.
Why AI support fails (and how to prevent it)
The failures are well documented and mostly preventable. In 2025 a coding tool's support bot invented a nonexistent device policy and drove real cancellations. In a 2024 tribunal ruling, an airline was held liable after its chatbot made up a bereavement refund rule. Klarna replaced hundreds of agents, then began rehiring after admitting quality dropped on disputes and fraud.
The common thread is over-automation and ungrounded answers. The fix is not to turn the AI off. It is to ground every answer or refuse, verify with a supervisor, compute policy in code, and route the hard cases to a human. Escalation is a good outcome, not a failure.
The metrics that tell the truth
Deflection (the customer did not reach a human) is not resolution (the customer's problem was solved). Gartner reported in 2026 that only 9 percent of customers say a chatbot fully resolved their issue without a human. If you only track deflection, a bot that stonewalls people looks like a success.
Resolve reports verified auto-resolution rate separately from deflection, and treats the 72-hour reopen and recontact rate as the honesty north star. A resolution that bounces back within three days was never a resolution.
Frequently asked
Can AI really resolve most support tickets?
Yes, for the common, well-documented cases that make up the bulk of most queues. Resolve resolves 80 to 90 percent of tickets when the answer exists in your content. The remaining judgment calls, like disputes, fraud, and refunds over a threshold, are routed to a human by design.
How does AI support avoid making things up?
By grounding. The agent only states facts it retrieved from your real content or live system data, a supervisor model verifies the draft before it sends, and business logic like refund math runs in code. If it cannot ground a claim, it does not auto-send.
Does AI support replace human agents?
No. It removes the repetitive volume so your team spends its time on the judgment calls the AI hands up. Every decision a human makes trains the model to handle more of that pattern alone next time.