At Pharmesol, we built and deployed agentic AI for pharmacy operations: voice and messaging that handles the high-volume, low-margin work quietly draining pharmacy teams.
Over my time there we nearly doubled our customer base, kept expanding scope with existing customers, and were acquired.
My role spanned design strategy, prompt engineering, and customer-facing work.
THE PROBLEMS WE KEPT HITTING AT PHARMESOL WEREN'T TECHNICAL. THEY WERE BEHAVIOURAL.
The hardest design problem wasn’t the AI itself. It was figuring out what to automate without burning the customer's time getting there.
THE CONSTRAINT
Pharmacy technicians are already stretched thin. Scope it wrong and you build the wrong thing. But the usual fix, long interviews and workshops, was itself a barrier.
We needed a scoping process that was light on the customer and flexed to their reality: their time, the mess of real workflows, and the gap between what people say they do and what they actually do.
WHAT WE BUILT
A scoping approach combining asynchronous inputs customers gave on their own time, a focused conversation with the operators closest to the work, and observation where it helped.
Not every customer needed every input. The process adapted to what was in front of us.
IT DOESN'T END AT LAUNCH
Once an automation went live, we monitored real calls for the things scoping couldn't predict.
The clearest example is the first comic below:
People wait patiently while a human types to look something up. The same wait from a machine feels intolerable. A caller will repeat their name three times for a person, but hang up if the AI asks twice.
The fix wasn't a faster AI. It was shaping the conversation: pre-empting where it veered off, guiding the caller forward, filling silence with the right signals of progress.
That monitoring loop fed straight back into how we scoped the next workflow, and pharmacies kept asking us to automate more.
TAKEAWAYS
1. Everything is a design problem, as our CEO liked to say. Latency, scoping, workflow selection, conversation experience. None of them look like design problems at first. Framing them that way is what lets you solve them with judgment instead of brute force. That’s the part that doesn’t scale automatically.
2. Designing AI is designing behaviour, not interface. What the system says, when, and how it holds a conversation.
3. Building AI products means working at the edge of the tools. New models, new prompting patterns, new evaluation approaches showed up weekly. We were automating our own work with the same kinds of systems we were building for pharmacies. The job was as much about learning fast as designing well.
To wrap, here's a comic about the importance of scoping right:
