How to audit AI agent handoffs to human agents

• Write down when your AI agent must hand off, then audit against that policy.
• Audit both directions: handoffs that happened and handoffs that should have.
• Grade loops, wrong answers and misunderstood requests that happen before the handoff.
• Check whether the human agent received context or made the customer start over.
• Track whether conversations recover after reaching a person, and route each pattern to the team that can fix it.

Handoffs are where a bot failure either gets rescued or gets worse, so they need their own checks. Customers expect the path to a person to work. In a Gartner survey of 3,566 B2B and B2C customers conducted in February and March 2026, 87% said it’s essential for companies using GenAI in customer service to provide an option to reach a human agent ¹. Gartner analyst Eric Keller adds that when customers are forced through multiple unsuccessful AI interactions before they can reach a person, they’re less likely to use that tool again ¹.

A containment rate on its own won’t show this. It can count a customer who gave up the same as one who got an answer, and it says nothing about whether the person who picked up the conversation knew what had already happened.

Use these five patterns as the backbone of your handoff rubric. Naming them helps reviewers and AI graders flag them the same way every time.

1. Missed escalation. The customer asked for a person, or the issue required one, and the bot kept trying.
2. Unnecessary escalation. The bot handed off a question it should have handled, which adds queue time and cost.
3. Loop or dead end before the handoff. The bot repeated itself or stalled while the customer rephrased the same request.
4. Wrong answer before the handoff. The bot gave incorrect information, and the human agent had to undo it.
5. Lost context after the handoff. The human agent didn’t see the transcript or intent, so the customer had to explain everything again.

A sixth outcome sits on the other side of the handoff: whether the conversation recovered. A handoff that ends with the customer’s issue resolved and their tone improved is a save. One that ends with the customer still frustrated is a failure even if every step looked correct.

Start by writing down when your AI agent must hand off, because you can’t audit against a rule nobody wrote. Gartner advises service leaders to set AI interaction policies, developing “guidelines for AI-led interactions, addressing data privacy, security, and escalation” ².

A practical handoff policy covers:

• Explicit requests. The customer asks for a person, in any wording.
• Topics the bot shouldn’t handle. For example, account security, payment disputes, legal complaints or anything your compliance team flags.
• Repeated failure. The bot hasn’t moved the conversation forward after a set number of turns.
• Customer frustration. The customer’s messages turn sharply negative.
• What travels with the customer. The transcript, the detected intent, the steps the bot already tried, and any identity checks already completed.

Keep the policy short enough that a reviewer can apply it to a transcript in a few minutes.

Audit the conversations that reached a person and the contained conversations that should have, because the second group is where missed escalations hide. If you only review handoffs, you’ll never see the customers the bot trapped.

• Handed-off conversations. Every conversation where the bot passed the customer to a human agent.
• Contained conversations with warning signs. Contained conversations that include a request for a person, repeated rephrasing, negative messages, or a reopened ticket on another channel soon after.

If you review by hand, take a fixed number from each group every week and keep the same mix so trends are comparable. If you grade with software, grade the conversations you connect and use the warning signs to prioritize what humans read.

Write each handoff rule as a pass/fail check with a plain-language definition and at least one example, so two reviewers reach the same answer on the same conversation.

For the bot’s side, these checks belong on the bot’s own rubric, because AI agents fail in different ways than people do. For the human agent’s side, add a check to your existing agent scorecard for picking up context without making the customer repeat themselves. Our guide on building support QA scorecards covers how to write checks agents trust.

The handoff isn’t finished when the human agent says hello. Grade the rest of the conversation on your human agent scorecard: did the agent read the bot transcript, correct any wrong information the bot gave, follow the process for the issue, and resolve it correctly under your policy?

Keep the bot rubric and the agent scorecard separate but view them side by side. That shows whether a poor outcome started with the bot, with the agent, or with the process in between.

Recovery tells you whether the handoff rescued the conversation. Compare the customer’s sentiment at the start of the conversation with sentiment at the end, and count how many conversations that turned negative with the bot ended neutral or positive after a person took over.

Compare conversations the bot finished with conversations a human agent handled. If bot-finished conversations end negative more often for one contact reason, that contact reason is a candidate for an earlier handoff.

An audit only pays off when each failure pattern lands with someone who can fix it.

Consider a hypothetical subscription company whose customers disputing a charge often end up angry even after reaching a person. The handoff audit shows a pattern. The bot treats “I was charged twice” as a refund-policy question and links the refund article three times before handing off. The human agent then opens with a generic greeting, and the customer explains the problem a fourth time.

The audit flags three failures in one conversation: a loop before handoff, a required escalation that came late, and lost context after it. Conversation design adds billing disputes to the handoff policy, operations passes the bot transcript into the agent’s view, and QA adds “acknowledges what the customer already told the bot” to the agent scorecard. The next audit checks the same three items on billing-dispute conversations.

A full handoff audit isn’t always necessary. A weekly manual spot-check fits when:

• Your AI agent handles a narrow set of simple requests and hands off everything else on the first turn.
• Bot volume is low enough that your QA team can read the handed-off conversations each week.
• You’re still in a short pilot and plan a full review before expanding the bot.

A structured audit fits teams where the AI agent handles a large share of contacts, where handoffs cross into outsourced or BPO teams, or where regulated topics make a missed escalation costly.

unitQ Support QA grades human agents on your scorecard and AI agents on their own bot rubric, side by side in one product, independent of the vendor that built the bot ³.

• An independent bot audit. It checks whether containment was real and flags wrong information, loops and dead ends, inappropriate escalations, misunderstood issues and inaccurate resolutions, and shows how often conversations recover after reaching a person ³. Bot Audit in unitQ Support QA is configured for your workspace during setup.
• Sentiment trajectory. Track how customer sentiment changes from the start to the end of a conversation, including conversations that recover from negative, and compare bot-finished with agent-handled conversations. Sentiment trajectory is set up per workspace; ask whether it’s available for yours.
• Missed steps on the human side. Flag missed steps that show up in the conversation, such as a skipped identity check, a required disclosure that wasn’t given, or an escalation that should have happened, checked against the process you define ³.
• Findings with owners. AI-generated findings come with recommended next actions, each tagged with a suggested owner such as operations, enablement, QA or conversation design.
• Evidence and coverage. Every scored miss links back to the conversation behind it, and unitQ Support QA scores the conversations you connect, not a sample.