← Back

Key Insights

by Yana Vlatchkova — 29 Jan 2024

Over the past 2 years, we talked with 100+ knowledge workers (Sales, customer support, product managers, founders, CXOs) and tested multiple prototypes. Here’s what we learned:

There’s documentation, yet people are still asking in chat. People want an instant, quick-to-read answer that is up-to-date and verified by a colleague. Here’s why:

  • Limited supply: Employees don’t find what they need. Only a fraction of the actual company knowledge is being documented and stored in the wiki. A lot of new insights discovered by employees end up undocumented. Because few have the permission, incentive or know where to put them. Thus Slack becomes a depository of some kind of this unstructured ad-hoc exchange of know-how.Most know-how is still stuck in people’s heads.
  • Speed & convenience: Employees have difficulty finding information inNotion/Confluence. Asking a question is a much faster way to obtain the know-how they need, tailored to their problem and level of understanding (e.g. sales asking product questions to the Product Owner need the information from the customer perspective.)
  • Assurance: Getting an answer from a colleague (especially someone with seniority) gives employees assurance that the information is reliable and up-to-date.

Employees and managers have resistance to adopting a new knowledge base.

  • People don’t want to go to “yet another place” to look for information. The natural place to ask a question is still in chat.
  • We designed a Slack Bot that doesn’t require employees to change habits.When a person asks a question in chat the bot auto-detects it and replies with a colleague-verified answer within seconds.
  • Managers don’t want to start a new knowledge base from scratch.
  • We solved that by turning the unstructured chat history into questions and answers that managers review and verify. Works the same with docs.

An AI-powered company knowledge base requires Information that to be cleaned and curated by people.

  • Simply plugging all company data into an LLM doesn’t guarantee the quality and reliability of answers. Because there can be contradicting or outdated information the AI will generate convincing answers dubbed “hallucinations”.
  • The AI is fantastic at extracting, organizing, and surfacing answers.
  • Information loaded into an AI knowledge base requires to be verified beforehand by company experts.

Customer-facing teams like sales, support, and customer success benefit the most from the speed of an AI-powered knowledge base.

  • It cuts down the waiting time for agents from hours to seconds. This is especially prominent for remote and async companies where customer and product teams are in different time zones.

There are 4 different types of questions employees ask.

  • Specific know-how: These are typically short questions with concrete answers that have been asked before >>> This is the category that we are starting with <<<
  • example: company policies, product features/bugs, sales know-how, and common support issues.
  • Status updates: These questions are time-specific and their answers are changing on a daily/weekly basis.
  • example: project progress, status on customer tickets, sales numbers, etc.
  • Solution: They are typically long-form and require a discussion between the person asking and the team before reaching a final decision.
  • example: uncommon support tickets, complex work-related problems, team decisions
  • Decision: People ask their team lead to take a decision not because it’s not in their own power, but in order to share responsibility