Two voice skills in production, in 2018
Pharmaceutical multinational. Patient information leaflets consultable by voice, and voice reporting of medicine anomalies wired into tracking and stock systems.
- My roleProduct Owner — solution design, requirements, development coordination
- Duration2018
- AreasAmazon Alexa · Google Home · Conversation design · API integrations
In 2018 voice assistants were a living-room curiosity and generative AI did not yet exist as a market category. A pharmaceutical multinational chose to test them on two concrete problems: one for people consulting medicine information, one for people handling the medicines themselves.
Voice looks like the most natural interface there is, and it is the hardest to design. There is no screen, no history, the user cannot see the available options and does not forgive a misunderstanding. In pharma there is a further constraint that outweighs the rest: information about a medicine must be exact, always, and cannot be paraphrased.
- 01
Understand
Identifying the two use cases where voice genuinely solved something: hands-free consultation of medicine information, and anomaly reporting by people working in the field who cannot stop to fill in a form.
- 02
Design
Conversation flow design: how a question gets phrased, what happens when the assistant does not understand, how to confirm a critical piece of data without making the exchange interminable.
- 03
Build and deliver
Development of both skills and API integration with shipment tracking and stock monitoring services. Released to production on Amazon Echo devices and Google assistants.
- Voice skill for consulting the patient information leaflets of a selection of medicines: the user asks aloud, for instance, what contraindications a medicine has, and gets the correct information.
- Voice skill for reporting anomalies — for example a medicine exposed to temperatures outside the specified range — and for checking shipment progress.
- API connection to third-party services for tracking data and stock monitoring.
- In the same period, a prototype for real-time doctor–patient translation: the doctor speaks their own language during the consultation and the application translates for the patient in theirs, and back. Years before language models made the idea obvious.
This is why agentic AI was never a change of field for me. A voice agent in 2018 already raised the questions we face today: understand the intent, ask for confirmation when the data is critical, connect to real systems to do something useful, behave predictably when it has not understood. The models changed; the questions did not.
The client is not named, for confidentiality. Sector and scale are real, as is everything else here.
Recognise the situation?
If something here resembles your case, send me a couple of lines. If it is not my ground, I will say so straight away.