In-Meeting Knowledge Agent

Can your people pull the right answer, or the right file, during a live meeting, without breaking the conversation to dig through drives and email threads? Yes. We built a full operations agent that ran in WhatsApp, working in Hebrew voice notes and text: it knew what existed in the organization’s content library, found the right materials, videos, manuals, deliverables, and delivered them in-chat instantly, supporting selling, marketing, and in-meeting moments in real time. A salesperson mid-meeting recorded “send me the assembly video for this model”, and the agent returned the video in the same chat.

How it works#

The agent’s world is your content library, not the open internet. It is grounded in what your organization actually has: product materials, manuals, videos, deliverables, and it answers by finding and delivering the real item, not by paraphrasing from memory. That grounding is the whole point in a commercial moment: what reaches the customer is your approved material.

The channel is the one your people already use. No new app, no portal to remember. The agent lives in WhatsApp, takes Hebrew voice notes and text, which matters mid-meeting, when typing is awkward and a murmured voice note is not, and returns the material in the same chat. It handles the unglamorous production details automatically: compression, format constraints, delivering a file that actually opens on the other side.

It behaves like a colleague, not a search box. It interacts to clarify when a request is ambiguous, goes deeper when asked, and keeps pace with the conversation, which is what “real time” has to mean for this to work in front of a customer.

The same pattern fits desk teams too: DataWise built and demonstrated a conversational organizational knowledge agent for a multi-site operations company, with free-language access to organizational knowledge, demonstrated on the company’s real content.

From the meeting to the task list#

The designed next step for the same agent closes the loop after the meeting: a forty-minute meeting recorded with everyone’s consent, decisions separated from discussion, each task given an owner and a date, and the list sent to the people who own the tasks with the minute it came from. This step is a design pattern, not part of the built agent above. The weekly numbers that the meeting starts from are the subject of the weekly management brief.

What we’ve built#

• The WhatsApp operations agent described above: built. 30 people around the world worked with it, in 15 languages. Hebrew voice and text, material discovery and delivery from a real content library, real-time support in selling and in-meeting moments. Meeting recording and decision-and-task extraction are the designed next step, not part of this build.
• An organizational knowledge agent for a multi-site operations company: built and demonstrated on the company’s real content.
• An isolated assistant inside a closed network: built and demonstrated, chat over organizational knowledge and retrieval over internal documents, nothing leaving the network.

From the wider industry: the US Army’s CamoGPT, a government-hosted assistant over controlled internal content, reached roughly 75,000 users (per US Army reporting) and is in active daily use per DefenseScoop’s coverage. Grounded knowledge assistants are a production category. In Israel, Hebrew-native AI in chat is mainstream: Meta brought Meta AI with Hebrew support into WhatsApp, per Ynet; the enterprise version of that shift is the same experience over an organization’s own controlled knowledge.