9 Essential Pieces of an AI Context System

Exponential Operator is a 100% human-written newsletter by Sawyer Middeleer, founder of Revi Systems. If you were forwarded this email, join some of the sharpest AI operators by subscribing here:

Hi all,

People who work in AI are fond of saying context is everything.

Context is the user- and situation-specific data that enables an AI system to be most useful doing real-world work. As critical as it is, managing context is often an afterthought for teams that are just starting to use AI.

Missing context turns out to be one of the biggest bottlenecks for any organization that wants to work seriously with AI. Products like Claude, ChatGPT or your vendor's platform-specific AI tool all perform worse when they are isolated away from the context they need.

For example, let's say you launch a new product and ask AI to write website copy to promote it. The result can only be as good as the data the AI has:

This is doubly true with agentic tools like Claude Code that can take on whole workflows, not just individual tasks. In an ideal world, every AI has exactly the knowledge it needs to do its job, at exactly the right moment.

Inside an AI-native company, all of this knowledge lives inside a well-managed context layer.

In total, there are 9 key building blocks that make up an AI context system. I haven’t seen much written about context management in AI business operations, so I'll start with a survey.

While the term "AI context layer" makes it sound like a structure similar to a database or data warehouse, it's usually not actually a single piece of software. Often it's actually multiple tools organized together in a way that both humans and AIs on your team can use.

It's helpful to map a generalized version of this system out with two main pieces: the context itself and the processes that manage the context. I'm open to arguments that I'm missing pieces, but in my view, these are the essentials:

If your team uses AI heavily, you likely have bits and pieces of each already built out without realizing it. Hopefully illustrating them in a framework will help you identify the gaps you still have to fill.

Remember that an AI is trained on billions of webpages, books, YouTube videos, and forum posts. While it may be obvious to you how to measure "month-over-month revenue," it might be less obvious to an AI. Having one or more glossaries that help a general AI think about your business the same way you do will reduce the need for manual corrections by humans.

You likely have a lot of this information already stored inside an internal wiki, but I'd bet that quite a bit of knowledge is trapped inside employees' heads or inside an old spreadsheet that AI doesn't know about. Do yourself a favor and get it out of there and into an organized, well catalogued dictionary.

If dictionaries are the "what", playbooks are the "how." Every team has their own SOPs for how they do work. The best playbooks illustrate clearly what good looks like and why processes are set up the way they are.

Depending on your situation, playbooks might live alongside your dictionaries, or they may even be in the same document. What really matters is organizing the information so that it's easy for an AI. (and a human) to find and read. This information, too, will likely come from wikis, internal docs, and tribal knowledge that you already have.

This category might be the most important of any in an AI context layer. Business state comprises all of the data that anyone in your organization would use to answer the question of "what's happening right now" or even what happened in the past. It includes a very wide range of data, likely coming from multiple different sources: business metrics, project status, customer records, and more.

While you might need to create dictionaries and playbooks from scratch, business state context can be built by integrating with your core tech platforms (CRM, CMS, ERP, project management software, meeting notes, etc). This context usually lives alongside or within "systems of action" -- the AI-powered software tools that enable agents to actually do work within your organization.

Under appreciated but powerful in the right hands, external signals such as market trends, social media and other data vendors help AI develop a complete picture of the context in which a business operates. What's useful here will vary business to business and team to team.

This is one area where AI's strength relative to humans can shine through. LLMs can consume millions of words per minute and can often extract more useful insights from unstructured data like white papers, news and competitor intelligence than the average human can. There's value in giving AI a lot to read and letting it find useful signal in the noise.

Memory is a huge topic, encompassing broader and more technical aspects than I can write about today. All major agentic systems - Claude Code being the most popular example - rely on built-in memory to store specific knowledge with fast recall.

Most relevant for us, AI memory is built to include facts and decisions that come up in conversation between users and AI. These memories are linked to individual users and dates, enabling the AI to capture a memory like "Allie and John changed the project scope on Tuesday afternoon to align with leadership's new strategy" and then make use of it in the rest of its work.

The final piece is the one that in my experience is almost always overlooked by ops teams because it sounds esoteric and technical. Relationship graphs are actually very simple - they're just a way to illustrate the connection between entities, concepts and facts so that AI can understand them better. An org chart is a graph of people. A sitemap is a graph of pages in a website or wiki. Map out all the links and citations within a library of content and you have another, more interesting conceptual graph.

These are useful to AI agents because LLMs have a limited context window. They can't store an entire knowledge base in their memory. Having a graph that shows how topics, memories, people, documents, etc are related allows agents to find the right information in an efficient manner. Rather than getting in the weeds with a technical solution, I recommend starting off by making your knowledge base sitemap and org chart available to AI, then getting in the habit of internal linking between documents - that'll get you 90% of the way there.

Stale playbooks need to be updated. Broken integrations need to be fixed when issues emerge. New knowledge has to be added when big changes happen. Maintenance is all about keeping the context system fresh and accurate, which is a lot easier said than done, especially when your context layer gets really big.

Many maintenance workflows can be fully or partially automated on schedules or on event triggers. But you cannot fully get around the need for human contribution. If you've ever had to chase people down to update a company wiki you know that making this happen is more a cultural issue than a process issue.

Observability is just a fancy way of saying "tracking what your AI is doing". Usually this takes the form of logs captured in a text file or admin dashboard. Practically speaking, I mostly care about observability when something goes wrong. Every single time I'm working with an AI system that's not performing properly, I turn to AI session logs as a goldmine of information about what my AI was doing, when and why.

9 out of 10 times, a poorly performing AI is getting tripped up by bad context. Maybe it's looking in the wrong folders, maybe it can't read the data you expect it to. Whatever it is, it will show up in the logs.

The more comprehensive AI context is, the more likely it is to include private or proprietary data that requires rigorous access control policies. You obviously don't want ChatGPT informing your team about the upcoming reorg you're planning. The best way around these pitfalls is to make sure permissions are set properly and restrict organizational AI from accessing the truly sensitive information.

That’s all for today. Thanks for reading!

江湖再见🫡

-Sawyer