Alf was built for a fast-moving, patient-first environment - an online medicinal cannabis clinic where people often arrive with pressing questions and a need for straightforward, reliable answers.
Because Alf’s scope was focused on general FAQs rather than transactional tasks, it was a natural fit for a purely Generative AI model. This gave Alf the flexibility to handle the unpredictable, varied ways patients phrase their concerns - without the brittleness of a rigid intent library.
The Challenge:
One of the trickiest problems to solve was a direct consequence of that flexibility: Alf was initially framed simply as a “Telehealth” agent, which meant it treated cannabis-related questions the same as any other controlled substance query and refused to engage - even though answering those questions accurately was the entire point of the product. At the same time, compliance rules meant Alf could never actually use cannabis-specific language in its replies to users.
Alf
The Solution
The fix was a two-stage architecture: reframing how the model understood its own role - explicitly as an agent for a medicinal cannabis Telehealth platform, not a generic one - so it would engage with these questions rather than refuse them, paired with a separate compliance-filtering agent that rewrote any cannabis-specific terms into neutral language before a response ever reached the user. An earlier, simpler attempt at this - a static word list - was dropped quickly, as it couldn’t keep pace with misspellings and the constantly evolving slang patients actually used.
The design work also centred on shaping a voice that felt warm and credible in a sensitive healthcare context, and on defining the boundaries of what Alf should and shouldn’t attempt to answer.
Impact
Alf’s approach proved well-suited to the task, maintaining a containment rate of between 92% and 94% - a strong indicator that patients were consistently finding the answers they needed without needing to go further.
Alf’s greeting and user prompting using buttons
Alf in Motion