Why context beats the model
Stan Hansen is COO of Egnyte, which secures, governs and organizes unstructured business content (documents, spreadsheets, PDFs, models) for more than 23,000 customers. Many of them are in regulated industries like financial services and life sciences. He joined Egnyte in 2021 as Chief Revenue Officer and became COO a couple of years ago. Before that he spent 13 years in technology at Wells Fargo and held SaaS roles at Adobe, Domo and Pluralsight. Egnyte has done machine learning work such as data classification, OCR and computer vision for almost a decade, so it entered the LLM era with plenty of content and plenty of practice handling it.
Stan’s main point is that the model isn’t what matters. Egnyte’s first impulse was to build and control its own LLM, and that lasted about two weeks. It now partners with the labs that spend billions on compute and concentrates on what it already has: hundreds of petabytes of customer content plus the business context around it, meaning people, projects, permissions, relationships, workflows and industry knowledge. Stan says applying that context makes models 70 to 95% more efficient, which cuts token costs sharply. As he puts it, “if you have to use the most sophisticated model to run your business, something’s wrong in your business.” Tying yourself to one model is a trap anyway, since the leaderboard changes every few months. If the data is structured so any model can use it, customers can keep picking whatever tool is best at the moment.
The context layer matters even more once work moves from one agent to many. When Sachin asked about collaboration, Stan said orchestrating multiple agents, working with each other and with people, is exactly what Egnyte is focused on now and where he expects the next one to three years to go. He admitted they haven’t fully solved it, but they’ve made progress on agents that do work, check each other’s work and hand off to people. Doyle’s view was that people are settling on one harness that knows them, and that agents in different tools will talk to each other more and more over MCP while drawing on a shared source of truth. Stan wants Egnyte to be that source of truth. It has about 1,000 MCP connections, more than 200 of them added in the last 45 days. This is where the two ideas meet. When many agents act for many people, each agent has to know whose data it can see, what’s relevant to that person’s role, and what everyone else is working from. Without managed context, adding more agents just makes the mess bigger.
The rest of the discussion
- “The blue collar of tech”: why holding back on AI hype paid off. Egnyte chose to underpromise and overdeliver. For the first six months of the AI wave it sold very little AI, while competitors promised to change customers’ businesses within weeks. Stan says that put real pressure on the executive team. Then buyers tired of the hype as reports showed how few AI projects were succeeding, and Egnyte’s narrow, provable wins (like cutting 20 hours of repetitive work for one person) gave customers something real to take to their boards. Last quarter, 43% of Egnyte’s sales were for its next-gen AI offering.
- Customer examples: AI that pulls data out of documents and into the systems that need it. A financial services company spent hours each week pulling reports from franchise information for due diligence questionnaires and turning them into executive summaries. Egnyte helped it build an AI that extracts the information, and it’s now starting to pre-fill the output as well. Stan says that saves hundreds of hours a week and improves accuracy. At a wealth management firm onboarding a new client, the AI classifies and tags incoming documents, sets permissions, files them in the right folders, and sends the extracted data to investment tools, CRMs or ERPs.
- How Egnyte uses AI internally: an 80-page sales rulebook that answers questions on the spot. They started with repetitive sales work and adopted ZoomInfo’s AI features and Gong early. The clearest win they built themselves involves a 70 to 80 page sales rules-of-engagement document that covers partner channels, overlapping territories and headquarters with several locations. Two to four edge cases came up each week, and each took 30 minutes to an hour of research and about a week to settle. They loaded the document into a model, so reps describe a scenario and get pointed to the right rules right away. It took 30 to 40 hours to build, and Stan estimates it saves at least 100 hours of management time a quarter. They’ve since added account briefs and sales coaching that draw on Salesforce, Gainsight and Gong. Stan said MCP made a huge difference here, and Doyle said he barely logs into Salesforce anymore and uses elvex instead.
- Governance in regulated industries: apply existing security rules to AI, and hold model providers to them too. Egnyte operates under GDPR, Saudi Arabia’s PDPL, HIPAA, CMMC and more. Stan starts from the customer’s existing security and governance requirements and extends them to AI. Egnyte holds its AI features to the same InfoSec standards as the rest of the product (ISO 27001, for example) and requires the same from the model providers it uses. A small company couldn’t get those terms from a lab like Anthropic on its own, but Egnyte can, because it brings 23,000 customers. Sachin called it “reselling legal clauses,” and Stan agreed. The labs benefit too, because they manage one agreement instead of thousands of small customers.
- Winning over compliance teams: let them test permission-aware AI themselves. Sales teams want to “release the hounds.” Legal, compliance and the CFO’s office are cautious, and Stan thinks they’re right to be. What convinces them is seeing permission-aware AI in action: two people ask the same question and get different answers because they can access different data. Egnyte’s AI Safeguards product adds a second layer of protection, and compliance teams run their own tests. Those tests often turn up gaps in how the business already runs, which Stan counts as a plus.
- Where it’s headed: slower than the hype, focused on context, with data that stays where it is. Stan doesn’t expect the world to end in the next three to five years. He thinks adoption will be a bit slower than the hype suggests, because companies are thinking harder about risk and liability. He expects context to become everyone’s focus (two years ago almost nobody talked about it) and data to stay put instead of being copied into every new AI tool. Companies are already clamping down, he says, because data moving between systems and spread across them creates security problems.
- Sachin’s take: chat isn’t going away, so the real question is how agents and people talk to each other. People have predicted the end of chat as the AI interface since ChatGPT launched, and chat has only become more common. Humans do most of their work through chat-like channels (email, Slack, video calls), so Sachin expects agents to work the same way. What’s still open is how agent and human communication fit together, especially when every agent and every person has their own intent.
- Closing advice: don’t be afraid of the tools, build with them. Stan has five kids, all working now, and he tells them not to fear these tools. Learn them, build your own agents, connect your own data and find your own productivity gains. In his experience, confidence about the future comes from actually getting involved. He also gave a shout-out to his friend Dan Clark of Truyo, a previous guest on the show.