Context Decoded: Explaining the "Trillion Dollar Opportunity"

We recently ran a webinar on one of the biggest problems in enterprise AI: context. VCs have called it the trillion-dollar problem. This is a recap of what we covered.

What Is Context and Why Does It Matter?

When people first started using LLMs, they were toys. We tested jokes, rewrote emails in the style of famous authors and waited to see what odd thing they’d say next. That’s over. Now when people talk to AI, they’re handing it real work: “Turn this broadcast script into a web article.” “Draft an email to Sarah to go tomorrow morning.” “What’s the status of Project Alpha?”

The trouble is that none of these tasks stand alone. Each one sits inside layers of context: the person asking, their team, the project, the company’s standards and the market the company works in. Every one of those layers affects how the task should be done.

Ask a colleague to “draft an email to Sarah to go tomorrow morning” and they already know who Sarah is, what format she likes, what time zone she’s in and what “tomorrow morning” means at your company. Ask an AI the same thing and you get: “I’d love to help! Can you tell me who Sarah is? What’s the email about? When exactly should it be sent?”

That gap between what you say and what the AI needs to know is context. However good foundation models get, they won’t read your mind. They can’t know how your company structures deals, how your team runs projects, or that Sarah likes bullet points and hates long intros. Someone or something has to supply that information.

Current Approaches: Memory and Context Graphs

The market has settled on two main approaches.

Memory is the one you’ve probably seen in ChatGPT. The model stores facts about you, like “Mike prefers his coffee this way.” It picks things up from your conversations and tries to remember your preferences over time.

Context graphs work differently. They try to read everything going on across your company and build a large web of facts and relationships: “Sachin is the CEO of elvex. He has authority over these decisions. This account is connected to that opportunity.”

Both ideas make sense on paper. Memory copies how people remember things. Context graphs try to capture everything a company knows.

The Problems with These Approaches

Memory breaks down at the company level. Today it only works for individuals. There’s no group memory, team memory or company memory. Even if there were, how would they fit together? Does a team preference override a personal one? When? Nobody has worked out the rules, so there aren’t any clear ones.

Memory also gives users little control. You may be able to see what’s stored and edit it, but you rarely know what’s being added or why. And it misses a lot. If memory only forms from your own chats with the AI, it can’t see the work that happens in Slack threads and Zoom calls, or the decision your VP made last quarter about how deals should be structured.

Context graphs fail on transparency. They’re 100% opaque. You can’t see how the system builds up what it knows or how it draws conclusions, and you have almost no control over it. Context graphs do capture more than memory, but they still miss anything that happens outside the source systems.

Most context solutions work under the hood. elvex works in the dashboard, with you. Context should be a human asset first.

How We’re Approaching Context at elvex

We’re doing it differently, and we’re prototyping what we think is a better way right now.

Our starting point is that context should belong to people first. So it has to be transparent, simple, editable and consistent.

Transparent: You can see exactly what context the AI is using at any moment, whether it’s company, personal or project context. All of it shows up in the interface.

Simple: We treat context as documents. You don’t need to know anything about graph databases or memory architectures. You read a document and you edit a document. That’s all.

Editable: If you don’t want the agent to pay attention to something, you change the file. If you know how deals should be structured at your company, you write it down, and the system doesn’t have to guess from scraps scattered across your software. Most importantly, you don’t need to be technical in order to understand how this all works, or how to get it to do what you want it to do.

Consistent: This is about reliability. The system should do what you expect, and being able to understand and control what it’s doing is a big part of that.

We’re also taking a clear position on how the layers of context interact. When you’re working on a particular project, the system leans harder on that project’s context. When you’re doing general work, it draws on company and personal context. The rules are written out where you can see them.

In the prototypes, users can see their full stack of context at any time. Company context shows who’s editing it and what’s in it. Personal context shows what the system has learned about you. Project context tracks what you’re working on, who’s involved, and which tools and conversations matter.

We’re also prototyping ways to build context: the system interviews you, pulls context out of documents you give it, checks with you that it got things right, and shows you exactly where it saved what it learned. Nothing is hidden.

It’s a different way of thinking about the problem. Memory works from the bottom up and grows only out of conversations. Context graphs work from your integrations down, watching your software and guessing from how it’s used. We want to do both, but with documents you can see and edit instead of systems that run automatically where you can’t see them.

We haven’t figured everything out. But we’re convinced that people need control over the context that shapes how AI acts for them.

Bring your agent. Bring your team.