The problem
A professional services firm of sixty people had accumulated significant organisational knowledge across documents, emails, wikis and people's heads. Finding information required knowing where to look or who to ask. New team members took months to become productive because they did not know where knowledge lived.
The firm had tried a wiki. It was inconsistently maintained and rapidly became outdated. People reverted to asking colleagues — which worked but created constant interruptions and kept the business dependent on specific individuals knowing specific things.
What we built
We built an AI knowledge assistant connected to the firm's document repositories, email archives, project histories and communication channels. The assistant provided:
- Natural language question answering. Team members asked questions and received answers synthesised from the available information, regardless of where that information resided.
- Contextual surfacing. When someone opened a project, the assistant surfaced relevant precedent, similar past projects and key contacts.
- Knowledge maintenance alerts. The assistant flagged information that appeared to be outdated — documents referencing old processes, answers that contradicted current practice — for human review.
The results
Six months after deployment:
- Time spent searching for information reduced across the firm
- Colleague interruptions for information decreased significantly
- New team member productivity improved — they could access organisational knowledge directly rather than needing to learn who knew what
- Knowledge that would have left when people departed was captured and remained accessible
How it worked
The assistant did not require anyone to change how they worked. It indexed existing information wherever it lived — documents, emails, project histories — and made it queryable. People continued to create and store information as they always had. The assistant made it findable.
The critical design decision was to make the assistant proactive as well as reactive. It did not just answer questions. It surfaced relevant knowledge in context — when someone opened a project, when someone was preparing for a client meeting, when someone was researching a topic. This contextual surfacing was where the greatest value was created.
What we learned
The most important lesson was that knowledge accessibility is a maintenance problem, not a capture problem. The firm already had significant knowledge. It was just hard to find. Making it accessible required less effort than convincing people to document what they knew, and it produced more immediate value.
We also learned that trust in the assistant's answers built gradually. Initially, people verified everything. After a few months of consistently accurate answers, they began to trust the assistant for routine queries and reserved verification for higher-stakes questions.
For more on how AI assistants differ from traditional knowledge bases, see AI search vs knowledge bases. For the broader assistant category, see AI knowledge assistant.
This case study describes a composite of real implementations. Results vary based on the specific context and information landscape.