Voice agents look sector-agnostic from the outside, but the constraints of the industry shape the design. Privacy in healthcare, time pressure in logistics and identity assurance in finance each bend the same technology in a different direction.
The three patterns below summarise the configurations we see most often in the field. Each is the same problem solved under a different constraint.
Healthcare: privacy and correct routing
In healthcare the call itself carries sensitive information. The sentence “you have a cardiology appointment” must not be overheard by anyone else in the room. So the agent shares nothing about appointments, departments or results until identity verification is complete.
- No clinical information is shared before verification completes
- Reminders go to the number already on file wherever possible
- Emergency language routes straight to a human or the emergency line
- The call summary is written in plain language rather than clinical terms
Another design decision concerns what the agent must not do: give medical advice. Directing a caller who describes symptoms to the right department is the agent's job; diagnosing or recommending treatment is not. That boundary belongs in writing, for both patient safety and professional liability.
Outbound reminders carry a different constraint: the person answering may not be the patient. The agent shares no detail until it has established whose record this is, and simply logs that the reminder was delivered. The script should therefore be short, neutral and free of personal detail.
Logistics: time windows and confirmation
Most logistics calls reduce to one question: when is my delivery arriving? The agent earns its place by connecting to the tracking system, giving a concrete time window and rescheduling when needed.
In logistics the good agent is not the fast talker — it is the one that states the right time window once.
Calls with drivers need a different design. A noisy cab, gloved hands and a short window of attention call for short sentences, unambiguous confirmations and the option to replay the message.
- Time-window confirmation and rescheduling live in one flow
- An address change becomes a task in the operations queue
- Driver calls use short sentences and spoken confirmation
- After a failed confirmation the call returns to the queue at second priority
Measurement differs here too. The metrics are first-call confirmation rate and repeat-call rate rather than containment. The point is not to end the call but to make sure the delivery actually happens.
Finance: identity, record and limits
In finance every sentence the agent speaks is recorded and may later become evidence. Responses therefore need to be standard, consistent and measured. Identity verification requires more than a single knowledge question.
The agent's authority also needs a written boundary. Raising a limit, moving money or changing a contract should not complete through the agent; the agent prepares the request and hands it to an authorised person. That boundary reduces both fraud risk and operational liability.
In financial institutions the retention period for a recording is set by sector regulation as well as by general data protection rules. Agent design has to line up with that period, and audio, transcripts and analytics copies should share one calendar. Otherwise data deleted in one system keeps living in another.
The common core
Across all three sectors the same four things decide success: identity verification, handoff rules, logging and measurement. The sector only moves the thresholds. When choosing a platform, look less at ready-made industry scenarios and more at how flexibly those four components can be configured.
Finally, adapt sector patterns rather than copying them. Appointments in healthcare, deliveries in logistics and balance enquiries in finance make up most of the volume. Perfecting that single scenario in the first release creates more value than shipping ten of them half-finished.
