PerspectiveAI & automation

The AI handover problem

The handover from an AI agent to a person is where customer satisfaction is most at risk, and it is rarely designed with the same care as the AI agent itself.

By the time a handover happens, the customer has already tried once to solve the problem, has not succeeded, and is often asked to start again.

It is easy to overlook because most AI reporting ends where the AI agent stops. Containment, deflection and AI resolution rates describe the conversations the AI agent kept. The conversations it passed on are counted elsewhere, usually in the human team’s queue, and their experience is rarely reported as a single journey.

What good handover design looks like
  1. Clear triggers, by topic and by signal
  2. A short summary for the human agent, not a full transcript
  3. A visible route to a person from the first message
  4. Priority in the queue for handed-over customers
  5. The handover measured as a journey of its own
  6. One owner across the AI and human teams
How to design each one

What the evidence shows

Customers say clearly what they expect. In a survey of 5,728 customers conducted in December 2023, Gartner, a business and technology insights company, found that their most common concern about AI in customer service was that it would make it harder to reach a person. In the same survey, 64% said they would prefer companies not to use AI in customer service, and 53% would consider switching to a competitor if they learned a company planned to use it.

That expectation has not changed as the technology has improved. In a Gartner survey of 3,566 B2B and B2C customers, conducted in February and March 2026, customers said they expect the option to reach a human agent when companies use AI in customer service.

Repeating information is the most visible cost of a poor handover. Zendesk, a customer service software provider, found in its CX Trends 2026 research that 74% of consumers are frustrated when they are required to repeat information. A customer who explains the problem to an AI agent and then again to a person has done so twice before any progress is made.

Many customers also arrive at the handover already frustrated. Gartner’s 2023 research found that 45% of customers who started in self-service said the company did not understand what they were trying to do. These are the customers most likely to reach a person, and the ones for whom a second failure is most damaging.

Vendor research points in the same direction, although I could not confirm it at source. Aissist, an AI support software provider, reports that poor handovers significantly reduce customer satisfaction (CSAT) even when the AI agent’s answers were accurate. Klarna, often cited as an AI support success, illustrates the commercial side. In May 2025 its chief executive said publicly that customers would always be able to reach a person, after a period in which its AI assistant handled the majority of chats. Reports differ on whether this was a change of direction or an addition, but it shows that access to a person remained part of the service promise.

Why handovers fail

Context does not travel with the customer. In many deployments the transcript is not passed to the human agent at all, or arrives as a long block of text that the agent does not have time to read. The quickest option for the agent is to ask the customer to explain again.

The trigger is set in the wrong place. An AI agent that is measured on containment has an incentive to keep trying, so customers may need to rephrase a question several times, or ask repeatedly for a person, before they are transferred. Each attempt adds effort and reduces goodwill.

The handed-over customer joins the back of the queue. After waiting with the AI agent, the customer waits again with everyone else, often without being told how long the wait will be.

Measurement discourages good handovers. When the headline AI figure is deflection or containment, every transfer to a person counts against the AI agent, so the incentive is to reduce handovers rather than improve them. A handover that resolves the issue at the first human contact is a better outcome than an AI conversation that ends with the customer giving up, but most AI reports would record the opposite. The measures in Stop reporting deflection to your board address this directly.

Nobody owns the join. The AI agent is frequently configured by a digital or IT team, while the human team is managed within support. Each side measures its own part, so the handover itself has no owner and no target.

What good handover design looks like

Define when the handover should happen. Set triggers by topic, such as complaints, account closures and high-risk or emotional issues, and by signal, such as a customer asking for a person or rephrasing the same question. Decide in advance which contacts should go straight to a person without an AI attempt.

Pass the context in a form a human agent can use. A short summary of the issue, what the AI agent has already tried and which details have been verified is more useful than a full transcript. The agent’s first message should confirm that the customer does not need to start again.

Make the route to a person visible from the start. Gartner reports that many customers fear AI will become another obstacle between them and a person. A route to a person that is clear from the first message addresses that concern directly.

Prioritise customers who have been handed over. Give them a separate service level, or at minimum tell them how long they are likely to wait and through which channel the human agent will respond.

Measure the handover as a journey of its own. Report the handover rate, the time from request to reaching a person, CSAT for conversations that moved from AI to a human, and how often human agents ask for information the customer has already given, which a small quality sample will show.

Assign one owner for the handover across both teams, with authority to change the triggers, the summary and the queue rules.

Where to start

Review 20 recent conversations that moved from an AI agent to a person, and note for each how many times the customer asked for help, whether the human agent received the context and whether the customer had to repeat anything. If an AI agent is already live, take the sample from last month. If the vendor is still being selected, ask for the same handover to be demonstrated end to end in your own helpdesk, including exactly what the human agent sees.

Related driver: Process

Put it into practice: Review 20 handover conversations

Already running an AI agent? My two-week AI Support Health Check measures this for you.

See the Health Check

More from Practical CX

All Practical CX articles
PerspectiveAI & automation6 min read

The CSAT drop nobody puts in the business case

Most AI business cases account for savings but overlook a key cost: customer satisfaction. Savings are immediate, while satisfaction costs emerge later.

Read →