Author: @danbeksha; Compiler: Peggy, BlockBeats

AI is entering enterprises, but the real question is not "should we use agents?" but whether these agents can understand the company itself.

This article uses the author's 100 days after joining Ramp as a clue to discuss a lower-level issue: in a high-speed company, we cannot just rely on newcomers to slowly read documents, ask colleagues, and supplement context, nor can each AI tool work independently. What is really important is to build a continuously updated "company brain" that can store meetings, documents, Slack discussions, customer feedback, and product decisions so that both newcomers and agents can start from the same context.

When the context is systematized, onboarding is no longer just a long adaptation process, and AI is no longer just an isolated tool. The value of enterprise AI may not ultimately lie in how many agents are deployed, but in whether the company can first establish a trustworthy, readable, and reusable knowledge base.

The following is the original text:

In the 4×100 meter relay race, the outcome is often not determined by the entire race, but is compressed into a 20-meter handover area. Runners must complete the baton at high speed: if the baton is started too early, the baton will fall to the ground; if the baton is started too late, the baton handover will have to slow down, and the entire team will lose its advantage in an instant. If the handover action itself is not precise enough - if there is any error in hand position, angle, or timing - the result may also be a dropped stick.

A team can have the fastest player in the game and still lose in those 20 meters. Speed ​​is important, and so is handoff. What really determines victory or defeat is whether the two can be established at the same time.

Every job handover I have ever seen is essentially like a relay race, except that one runner is still at the starting block. New employees join the company on Monday and everything starts from scratch; however, the organization will not slow down and will continue to move forward at its original pace. As a result, newcomers can only rely on reading documents, lurking in Slack, asking the same questions repeatedly, and spending three months to figure out the organization's operating model until they finally become "useful."

We usually regard this gap as a matter of time, as if given enough time, newcomers will naturally catch up. But that's not the case. This gap will either be addressed by the system or it will persist.

Context is the real handover system of the organization

I've been with Ramp for about 100 days. Prior to that, I spent five years at Plaid, where I became familiar with every product, every customer story, and the context behind every decision. I can tell these stories without thinking. But coming to Ramp, I knew next to nothing about any of it.

The core of product marketing is storytelling. If you don’t know the characters, plot, and consequences, you can’t really tell the story.

From day one, my goal was to build an AI-native product marketing organization. But to do this without context, I first had to expand my base of knowledge—the “contextual layer” that underpins all my work.

Ramp is a company known for speed. There is no room for "slowly catching up next quarter." The company is releasing, iterating, and advancing every week. You either have to keep up, or you become an additional cost in running your organization.

Meanwhile, I was going through another layer of onboarding. Ramp was already fast, but AI was changing even faster, and I had to learn a new company and a new work environment at the same time. I'm not an engineer, and the last time I opened a terminal was in college computer science class. In other words, I have to make up for the organizational context and adapt to the new way of working with AI, and these two things are superimposed on each other, further amplifying the difficulty.

What finally freed me from this kind of pressure was not completing a specific article, product release, or workflow, but treating the "context" itself as the deliverable. As long as the context layer is built correctly, all subsequent work will become cheaper.

So I set out to build something truly scalable: a system that would help me make up lessons quickly in the same way that a good wiki helps researchers. By week three, it was drafting content based on my notes; by week eight, it was summarizing meetings I didn’t attend. Studying and tutoring are not going away, but as the system continues to fill up, they are starting to become less expensive by the day.

A personal version of this idea has actually been around for some time. Karpathy, the former head of Tesla AI and one of the founding members of OpenAI, wrote an article in April describing what he called a "personal LLM knowledge base": a folder that stores original input, including papers, articles, transcripts, and personal notes; an LLM that generates a wiki on top of these materials; and an editor like Obsidian as the front end. When the data accumulates to about 100 articles, LLM can answer complex questions around a personal corpus without the need for complex search techniques.

His judgment is: There is an opportunity here to birth a truly outstanding new product, rather than a bunch of improvised scripts.

The personal version exists today. But the corporate version isn't there yet. This is exactly the problem.

Roughly speaking, I built such a system in the first 100 days of joining the company. They are not yet sophisticated, but together they form the "connective tissue" within the organization.

The core is an Obsidian vault, read and written by Claude. Meeting transcripts, documents, public opinions, and personal notes that I have access to will all enter this knowledge base. When I ask, "What did Geoff and I decide about the homepage three weeks ago?" it looks to this vault for the answer, rather than relying on the generalized memory of the model itself.

To continue feeding this vault, Granola records every meeting by default and archives the transcripts overnight. So, the meeting I missed on Monday was available for query on Wednesday. To allow the rest of the company to keep up, I choose to work publicly - most of the content I'm building appears in #team-pmm or related release project channels before making it into the Notion docs. The build process itself is a synchronization mechanism.

On top of this vault, there is also a small named skill library that the agent can call on demand. One skill can generate an agenda based on the last four meetings I had with someone; another skill can scan Slack for a week’s worth of product updates and turn them into article topics. Each skill is about 200 lines of markdown and is used to replace a type of work that needed to be done manually in the past.

In addition, I built a dynamic product roadmap based on Ramp's internal application platform. It's reading from the same set of context layers, so it doesn't expire because it was never a static document to begin with. There’s also a morning summary that’s sent to my Slack DM every morning at 8 a.m.: what went live yesterday, where I got stuck, and what I need to respond to. These were sorted while I slept.

Taken alone, none of these things are amazing. But taken together, they provide a workable answer: What would it look like if a company had the kind of wiki Karpathy is talking about?

You might call it a wiki, a graph, a context layer, or the company brain. The name is not important, the function is. It must be able to absorb all the signals the company has generated: meetings, Slack discussions, documents, code, transcripts, customer calls, and key decisions, and keep it constantly updated without relying on human manual maintenance. It must also be the first thing every new employee and every new agent reads before they start working.

If a new employee joins the company tomorrow, what should he read on his first day? If the real answer is a Notion document from 2024, plus a defunct Confluence link, that's essentially asking him to take over from a standstill.

From point tools to corporate brains, the real gap in AI

Today, the main way for AI to enter enterprises still relies on forward-deployed engineers. Whether it is OpenAI, Anthropic, or a large consulting company, they will choose to build specific workflows on top of the model.

These jobs are real and valuable. But they are still stuck in the "chatbot era" of enterprise AI: narrow tools packaged around specific tasks, useful in isolation, but not connected to a system that can sustain compound interest.

The real "corporate brain" has not yet emerged. The customer service agent and HR onboarding agent may be built in different months and by different teams. They don't know each other what was decided at the last all-hands meeting, how the company understands its market, or what judgments the sales leader made at the last management offsite. Each agent is just a chatbot with specific responsibilities, but they don’t share the same brain.

This is the biggest gap right now. Outside the lab, few people are building products around this problem.

If you’re building a team or starting a company in 2026, the order of operations is already different than it was in 2022. Write the context file first and then install the tool. Document every meeting. Build the wiki first, then the dashboard. Deliver skills, not slides. Have new employees read the wiki on day one and start contributing to it on day two. Recruit and promote those who can keep the "company brain" running, and also reuse those agents who can actually read the company's brain.

Context is not a side project. It’s the infrastructure that makes all AI investments truly pay off.

I'm currently building out some of this in Ramp: a wiki, a skill library, an application that reads information from the same context layer, and an organizational mechanism for continually feeding it. It's still small and early days. If you're trying to build a corporate version elsewhere as well, I'd love to exchange experiences. The only thing more useful than one trustworthy brain is two brains in the same room.

Back to the relay. The real victory condition is not the cleanest handover, nor the fastest shot, but both happening at the same time in the same 20-meter stretch.

New hires read the company brain and start sprinting. The new agent reads the company's brain and starts working. New customers plug into the company brain and are up and running from day one.

We'll know we're doing something right when the word "ramp-up" no longer makes sense.