Joel Hellermark on Knowledge Access and Sana’s A-4 Agent

Sana founder Joel Hellermark opens the 2024 summit by recalling Richard Hamming’s question: what is the most important problem in your field, and why are you not working on it?

He recounts a story about Buckminster Fuller and Albert Einstein identifying access to knowledge as a “meta problem.” Whether or not the anecdote is historically verified, Hellermark uses it to explain Sana’s aim of helping people find and apply knowledge.

He names customers in energy, transportation, medicine and construction and reports that one user accessed Sana 2,700 times in six months. These are company examples and reported usage figures from the 2024 presentation.

Hellermark reviews the product’s development: searching knowledge spread across systems, asking questions through chat, and adding meeting notes to a searchable knowledge base. He says a free tier was introduced to broaden access.

He argues that assistants still struggle with stale knowledge, using tools, multistep problems and comprehensive retrieval. He says semantic similarity search alone may miss relationships that a knowledge graph can help represent.

His proposed shift is from a single-pass assistant to one that plans, retrieves information, checks results and tries again when needed. Sana calls the agent solution presented at the event A-4.

Hellermark describes an automatically derived company knowledge graph that users can correct by removing outdated information or adding missing facts. He presents this as part of the product design.

Model routing sends different questions to models suited to them, he says, aiming to balance quality and speed. He also describes tools that let an assistant work with applications such as Salesforce and recruiting systems.

Reflection and multistep reasoning let an assistant draft a plan, check intermediate results and improve its answer. He reports benchmark improvements for combinations of these techniques, though the figures are specific to the evaluation he presents.

Hellermark introduces supervised automation: the system anticipates a useful action and prepares it, while the user reviews and completes it. This is a proposed step beyond waiting for a prompt.

In one example, Sana finds notes from a meeting and prepares an update to Salesforce. The user can inspect the proposed action before completing it.

A home-screen catch-up flow presents summaries of missed meetings and suggested actions generated in the background. Users can inspect the summaries, ask follow-up questions and choose which actions to complete.

A construction-company example drafts a reply to a complex email by retrieving relevant information from several systems. Hellermark says the user can review or edit the draft before completing it.

He also shows a catch-up view synthesizing recently shared documents. The aim is to move from finding information toward helping people complete tasks that depend on it.

Hellermark previews a voice interaction built with OpenAI and jokes about giving the assistant a passive-aggressive personality. He discusses possible uses for sales representatives and researchers.

He reports early productivity gains of 30–50% across selected use cases, comparing them with historical technology adoption. These are figures presented by the company, not a universal estimate for all jobs.

Hellermark closes by encouraging organizations to use knowledge tools in support of goals such as medical research and renewable energy.