Better conversations need a human handover
What a peer-reviewed customer-support study tells us about AI assistance, answering automation and responsible receptionist workflows.
Evidence worth reading carefully
A study published in the May 2025 Quarterly Journal of Economics examined a staggered rollout of a generative AI assistant across 5,172 customer-support agents. Productivity, measured as issues resolved per hour, rose by 15% on average, with substantial differences between workers. This was assistance for human support agents at the studied business, not proof that an autonomous phone receptionist will produce the same outcome. [1]
Why the distinction matters
A support assistant can suggest a response while a person decides what to say. An automated receptionist handles part of the interaction directly. The evidence from one setup cannot simply be transferred to the other. A voice-automation pilot needs its own tests for understanding, accuracy, routing and customer experience.
Give routine conversations a clear structure
Start with a narrow set of intents: business hours, message taking, appointment enquiries and requests to speak to a person. Build from an approved knowledge source. If a question falls outside that scope, the useful action is to capture the request or transfer it, rather than produce a confident guess.
Design the handover before the greeting
Decide what happens when a caller is upset, repeats a question, gives conflicting information or asks for a human. Tell callers when they are interacting with an automated system. Pass only the relevant context to the staff member, and avoid making the caller repeat the entire conversation. These are suggested design principles, not findings claimed by the study.
What to measure in a local pilot
Count correctly routed calls, completed messages, misunderstood requests and successful human handovers. Review a permitted sample of interactions using an agreed privacy process. Staff time is useful to measure, but a shorter call is not automatically a better call.
Test ReceptionistX with a defined call flow
Book an appointment to explore a free ReceptionistX POC or trial. Agree a small set of answering and speech-automation scenarios, escalation rules and success criteria before the test starts. Keep the final decision about a broader rollout grounded in your own results.
Sources and context
Sources reviewed on 7 October 2026. External results are specific to their study or organisation and are not guarantees of PartnerX product performance.