Designing the handover after an AI failure

By the time a customer reaches a person, they may have already typed their order number twice, followed a link that took them back to the page they started on and explained that the parcel definitely isn’t behind the bin. The AI has tried a few standard answers. The customer has become shorter with every reply. Then the conversation is handed to a member of staff.

The member of staff appears to receive a simple question: where is the parcel? Their actual task is larger. They need to work out what has happened, decide which parts of the conversation can be trusted and recover enough confidence for the customer to carry on. They are doing this while two or three other chats are still open.

This is what disappears inside the phrase ‘human in the loop’. A person is present somewhere. We still don’t know when they arrive, what they can see, how much attention they have or what they are allowed to do once they take over.

A May 2026 working paper by Yiwei Wang and colleagues gives us a rare view of this work as it happened. The researchers studied an agentic AI on Alibaba’s Taobao platform. It spoke directly to customers and dealt with a limited range of standard service requests. Human workers watched the conversations, handled other chats at the same time and could step in when they thought it was necessary. Monitoring software could also alert them when it detected a technical problem or signs of frustration.

The research took place inside a real service operation. It followed 647 remote gig workers and 680,676 chats over 17 treatment days in August 2024. That scale makes the findings worth taking seriously. The deployment was still narrow. Only 5.8 per cent of the chats included in the analysis were considered suitable for the AI, and the paper is a preprint that has yet to complete peer review.

The customer had already been through the service

Workers were randomly assigned to carry on with all-human service or to supervise the AI while dealing with other customers themselves. Across the full set of chats, the AI arrangement cut average chat time by 3.2 per cent. Customer ratings and the rate at which people returned with the same problem within seven days did not change by a statistically significant amount.

Those figures would sit quite comfortably on a programme dashboard. The service became a little faster and quality held steady. A closer look shows how that average was made.

The small group of chats handled by the AI became 16.8 per cent faster. Their customer ratings fell by 0.412 points on a five-point scale. Meanwhile, the human-only chats handled by workers in the AI group became 1.8 per cent faster and received slightly better ratings. The researchers link this to attention. Workers switched away from their own customer conversations less often because the AI was carrying some of the other work.

People had different experiences inside the same overall result. Some received a fast automated service that they liked less. Others received more attention from a person. The stable average did not describe either experience particularly well.

The researchers then looked at 11,069 AI-led chats and matched each one with a similar all-human conversation. Technical alerts usually appeared when the request had moved beyond what the AI could deal with. Once a worker took over, these chats lasted 19.1 per cent longer than comparable all-human cases. Ratings and repeat contact were similar. The person was able to recover the service, although it took extra time.

Emotional alerts arrived after the software detected frustration or doubt. These cases lasted 40.8 per cent longer. Customer ratings were 0.928 points lower and repeat contact rose by six percentage points. Workers replied more slowly after they entered, sent fewer messages and did less information-seeking and solution work. Their measured empathy was similar to workers in the other escalation groups. They seem to have had fewer practical routes left by the time the alert reached them.

Workers who stepped in on their own judgement tended to arrive earlier. Their chats still lasted 9.5 per cent longer than comparable all-human cases and ratings fell by 0.524 points. The fall was smaller than it was after an emotional alert. Repeat contact was 3.2 percentage points lower, although the statistical evidence for that result was weaker.

The pattern makes intuitive sense. A person has more room to help before a customer has spent ten minutes going round in circles. The study cannot prove that timing caused the difference. Workers were randomly assigned access to the AI, while the moment and type of escalation emerged during the work. The researchers matched cases on twelve features, but an early self-intervention may still differ from a late emotional alert in ways the data did not capture.

There are other details to keep in view. Only 17 per cent of chats received a customer rating. Workers were paid by volume and managed an average of 3.14 conversations at once. Their ability to notice trouble was shaped by the rest of the job, not only by the quality of the alert.

The handover gives the person a different job

Monitoring an AI conversation asks for a different kind of attention. The worker needs to notice a weak signal across several live exchanges, judge when the AI has reached its limit and take responsibility for a conversation they did not begin. A neat summary helps them catch up. It cannot return the exchange to its starting point.

The AI has already asked questions, selected an explanation and possibly made a promise. Each action changes what the customer says next. A customer who has been asked for the same detail several times may leave it out when the person arrives. Someone who has already been told that nothing can be done may read the next answer as another refusal. The handover carries this history even when the case record looks complete.

Research on assistive AI makes the sequence easier to see. In a peer-reviewed field experiment at a meal-delivery company, Shunyuan Zhang and Das Narayandas gave some customer-service staff access to AI-generated suggestions. Responses became faster and customer sentiment generally improved, with larger gains among less-experienced workers. The AI worked behind the person. The worker continued to speak to the customer and decided which suggestion to use.

The same assistance could land differently after a chatbot failure. Very fast replies from the transferred human sometimes led customers to think they were still dealing with the bot, and sentiment worsened. A quick response usually looked helpful. In this part of the service, it became evidence that the failed interaction was continuing.

A second 2026 Alibaba field experiment looked at an assistant that offered diagnoses and possible solutions near the beginning of a chat. Average speed and customer ratings improved. Lower-performing workers gained most. Top performers received worse ratings and saw more customers return with the same problem. The process data suggest that the tool interrupted the way these workers managed several chats, although that explanation remains observational and the paper is also a preprint.

The studies are looking at different arrangements. One AI speaks before the person. Another sits behind them and supplies material. In each case, the AI changes the order of work and where the worker puts their attention. Worker experience and the customer’s recent history then shape the result.

A broader 2024 meta-analysis in Nature Human Behaviour reached a similarly untidy conclusion. It covered 106 experiments and 370 effect sizes from research published up to June 2023. Human-AI combinations performed better than people working alone on average. They performed worse than the stronger of the human-only and AI-only conditions on average. Results varied widely. Knowing that a human and AI are both involved tells us very little about the arrangement a team has actually built.

A handover needs more than an alert

The first design decision is where the AI should stop. Technical limits matter, and so does the state of the conversation. A customer repeating the same information, several failed solution attempts or a sharp change in tone may each tell us that the useful window is closing. Workers need a way to enter before the monitoring system reaches its formal threshold.

Every earlier alert creates work for someone. If staff are watching too many conversations, the service can make them responsible without giving them a realistic chance to respond. Alert volume, the number of chats being supervised and the time available to take over need to be tested together. An escalation rule that works with one live chat may fall apart with four.

Ruth Schmidt and Zeya Chen’s work on positive friction in human-AI interaction is useful here. An alert interrupts a worker and pulls them away from something else. That interruption earns its place when it creates a useful moment for judgement. A team can measure the work displaced by earlier alerts and the recovery those alerts make possible. Both belong in the decision.

The transfer itself needs to give the person enough state to act. They should be able to see the customer’s original request, what the AI tried, which claims it made and why the case was escalated. Uncertain facts should look uncertain. A generated summary can speed this up, provided the relevant conversation and records remain available when the worker needs to check them.

The customer also needs to know that the service has changed hands. A clear introduction from the worker can do more than a small ‘transferred’ label tucked into the interface. It gives them a reason to explain the problem again and separates the worker’s reply from the earlier automated exchange.

Authority matters once the person arrives. They may need to correct something the AI said, offer a remedy or move outside the standard route. A worker who can only apologise and paste another approved answer has been handed the emotional cost of the failure without the means to repair it.

Measurement needs to follow the same path. Average handling time can improve while one group of customers receives a worse service. Teams need to look at resolution, repeat contact and customer experience for the case types the AI handles, alongside the effect on the human work around them. Worker experience deserves attention too. A design that helps a new starter may interrupt someone who already has an effective way of managing the task.

I would treat the handover rule as an operating choice that will need revisiting. A 2026 review by Philip Cash, Ruth Schmidt and Weston Baxter found that design processes often call for iteration without telling people when to stop, repeat or change course. In this setting, the criteria can be fairly ordinary. A team might revise the threshold when late emotional alerts rise above an agreed level, when staff cannot respond in time or when repeat contact increases for a particular request.

What the research still leaves open

Seventeen days tells us very little about the longer experience of supervision. Workers may get better at spotting trouble. They may become tired of alerts or lose touch with the underlying service work. The study also says little about settings where a poor handover has legal, financial or clinical consequences. These were standard customer requests selected because the AI was thought capable of handling them.

A useful next experiment would randomise the handover rule itself. Some workers could receive earlier alerts when a conversation stops making progress. Others could use the later emotional threshold. Researchers could vary the number of chats being watched and the information available at transfer. The record would need to show what the AI did, when the person entered and what happened afterwards.

Teams do not need to wait for a large trial to learn something locally. They can record how the service works now, introduce one change and follow the cases through to resolution. Failed starts, late rescues and the work displaced by an alert all count. The comparison needs to stay close to the experience of the customer and the person expected to finish the job.

The current evidence gives us a strong reason to design the handover as part of the service. It does not yet tell us the right moment for every task, or how many conversations one person can sensibly supervise. Those answers will come from observing the work that happens around the alert, including the awkward parts that a dashboard tends to smooth away.

Human-AI Performance

By Lauren A Kelly

© 2026 Alterkind Ltd. All rights reserved.
Human-AI Performance™ is a proprietary methodology developed by Lauren A Kelly using my Behaviour Thinking® framework. All content, tools, systems, and resources presented on this site are the exclusive intellectual property of Alterkind Ltd.

You’re welcome to use, share, and adapt these materials for personal learning and non-commercial team use.

For any commercial use, redistribution, or integration into client work, services, or paid products, please contact lauren@laurenakelly.com to discuss licensing terms.

Icons by Creative Mahira, The Noun Project.

Thanks to Nicholas Edell, Valentina Tan and multiple VPs implementing AI for your feedback during development.

LICENSE
Based on work by Lauren A Kelly.
For commercial licensing contact: lauren@laurenakelly.com