A knowledge-base chatbot built for 280 pharmaceutical sales representatives — giving them instant, approved answers to product and clinical questions right before or during a customer call.
The organisation's sales force covered six pharmaceutical products — including two recent launches — across 14 therapeutic areas. Reps were expected to answer highly technical questions from healthcare professionals, handle clinical data, counter objections, and stay within tight regulatory limits — often mid-conversation, with no time to look anything up.
The existing tool was a folder-based intranet — updated infrequently, hard to search on a phone, and structured around how the organisation filed documents rather than how a rep needed to find information. If they could not find the answer in 30 seconds, they guessed, escalated to a manager, or said nothing. None of those were good options in a regulated selling environment.
The brief: build a virtual assistant that reps could query mid-preparation, between appointments, or standing in a corridor — and get the right answer in under ten seconds, in plain language, with a source they could cite.
Before mapping questions or creating content, we did a full content landscape audit — because you cannot build a useful assistant on a disorganised knowledge base.
The question mapping process was the most critical — and most time-consuming — phase of the project. A virtual assistant is only as useful as its ability to recognise what a user is actually asking. To achieve this, we had to understand the full landscape of questions a sales representative might pose, in the language they would actually use, not in the language of product documentation.
We ran structured question-elicitation sessions with 40 representatives across four regions — asking them to recall every question they had ever been asked by a healthcare professional, every piece of information they had ever struggled to find, and every situation in which they had wished they had a faster answer. Sessions were recorded, transcribed, and analysed for patterns.
Reps do not ask "What is the mechanism of action of Product X?" They ask "How does it work in the body?" or "What happens when a patient takes the first dose?" The chatbot had to understand both — and map each to the same underlying answer. We catalogued over 340 natural language variants across the 580 distinct question types identified.
Four types of content — each designed for a different kind of question, each structured for optimal chatbot retrieval.
Microsoft Copilot Studio was selected for its native integration with the organisation's Microsoft 365 environment, its retrieval-augmented generation architecture, and its ability to restrict responses strictly to approved content — a non-negotiable requirement in a regulated pharmaceutical context.
The most common failure mode for chatbot projects is leading with the technology and retrofitting the content. The assistant's intelligence is entirely a function of the quality, structure, and completeness of its knowledge base. We spent 60% of the project on content — auditing, creating, tagging, and approving — before the Copilot Studio build began in earnest. The technology build took six weeks. The content foundation took fourteen.
This sequencing felt uncomfortable for stakeholders who wanted to see the chatbot working quickly. The pilot results — 94% response accuracy and a 71% reduction in information search time — vindicated the approach comprehensively.
A side effect of building the knowledge base was that it forced the organisation to resolve content conflicts that had existed for years unnoticed — cases where two product teams had different approved data points for the same clinical question, or where the intranet contained both a current and a superseded version of a product guide.
The chatbot could not have two answers to the same question. This forced decisions. And those decisions — made by medical affairs and product teams together — produced a governed, versioned, single-source content library that the organisation had needed for years but had never had a compelling reason to build.
Within three months of full deployment, the virtual assistant was handling an average of 1,400 queries per week from 280 representatives. The 94% accuracy rating — assessed through a monthly sample review conducted by the medical affairs team — held steady across all six product areas and all query categories.
The operational impact extended beyond individual rep performance. The regional sales managers reported that call preparation time dropped significantly, post-call query escalations fell by over two-thirds, and new representatives reached full product knowledge competency two weeks faster than the prior cohort — because they had a tool that could answer their questions immediately rather than waiting for their manager's availability.
"I used to spend 20 minutes before every important call digging through the intranet trying to find a clinical reference I half-remembered. Now I ask the assistant and I have it in 10 seconds. I have genuinely changed how I prepare for calls because of this tool."
— Senior Medical Sales Representative
Learning matters. Let's build something that travels with your people — and answers when it counts.