UK retail studies find complaint chatbots win back buyers only when they can issue refunds

The work combined a qualitative study, two pretests, two randomised experiments and two secondary-data studies. The experiments used an incentivised purchase choice, while the field studies tracked repeat purchase, spending, time to next purchase and churn.

The advantage was not universal. Severe failures weakened the benefit of empowered AI, and strong customer relationships raised both the penalty for a low-authority AI agent and the reward for an empowered one. The authors recommend selective empowerment for routine failures and hybrid escalation to people for severe or data-constrained cases.

The research separates two features usually bundled together: whether the complaint handler is AI or human, and whether it is empowered, meaning it has authority to complete a remedy during the encounter. This lets the authors test each independently.

The evidence base spans seven parts in UK retailing: a qualitative study, two pretests, two randomised experiments and two secondary-data studies. Experiments measured an incentivised purchase choice; field studies observed real repeat purchase, spending, timing of the next purchase and churn.

The authors propose an authority-identity decoupling view: customers respond not only to who communicates but to whether the recovery system gives that agent the data access, decision rights and operational capability to act.

Their practical conclusion is selective empowerment for routine failures, with accountable hybrid escalation to humans for severe or data-constrained cases.