Read 20 recent conversations that moved from an AI agent to a person, and record what each customer went through before reaching your team.
- Time to set up
- Half a day
- Cost
- £0
- Driver
- Process
The problem
Most AI reporting stops at the point where the AI agent hands over. The handover rate shows how often a transfer happens, but not what it was like for the customer. Zendesk, a customer service software provider, found in its CX Trends 2026 research that 74% of consumers are frustrated when they have to repeat information. A small sample of real conversations shows whether this is happening in your handovers, and why.
The fix
- Select 20 recent handovers. Take the most recent conversations from the last month that moved from the AI agent to a person, across more than one contact reason. If no AI agent is live yet, use conversations that reached the team from a web form, chat or phone menu. If you are still choosing a vendor, ask each one to run a handover end to end in your own helpdesk.
- Record five points for each conversation: how many times the customer asked or rephrased before the transfer, whether they asked for a person, whether the human agent received a summary of what had already happened, whether the customer had to repeat anything, and the time from asking for a person to the human agent’s first reply.
- Note what triggered the transfer: a customer request, a topic rule, or the AI agent being unable to answer.
- Look for patterns rather than individual errors. Group the findings by contact reason and by trigger. Several conversations with the same problem are likely to point to a single rule, setting or missing summary field.
- Agree one change and one owner. Choose the problem that affects the most conversations and assign a named person to fix it.
- Repeat the review monthly with a new sample, so the effect of the change is visible.
Tools to try
- Your helpdesk’s views or search, filtered on conversations that include both an AI agent reply and a human agent reply. Most helpdesks can mark these with a tag or a rule.
- A spreadsheet with one row per conversation and one column for each of the five points.
- An AI assistant such as Claude or Copilot to summarise exported transcripts and flag repeated information. Remove customer personal data before exporting, and check a sample of its findings by hand before relying on them.
How to measure it
Track the share of handed-over conversations in which the customer repeated information, and the time from asking for a person to the human agent’s first reply, each month. Report them alongside customer satisfaction (CSAT) for the AI-to-human group (see Split your CSAT by who handled the conversation). Both measures should fall after the change; if they do not, the next most common problem is the one to fix.
Why this matters: The AI handover problem
Want a second view on how your AI agent hands customers to your team?
Book an intro callSources
- Zendesk CX Trends 2026 press release (November 2025)