The promise of building a “brain” for your organization is pretty simple: imagine an internal knowledge system that pulls together data across all your services, so that everyone on your team knows what the organization knows, at the right level of access.
But what does it actually mean to build your organization’s brain? And how would you do it? If you run operations at a nonprofit, these questions are probably already on your mind. We build these systems for nonprofits at Sattva, and this is the first in a series of posts on how to do it well.
I’m going to give you a basic framework that can help rally your team around building a robust knowledge system. It’s called the 4 Cs: Context, Connections, Capabilities, and Cadence.
Context
Context has both an everyday meaning and a meaning specific to AI agent systems. In the everyday sense, context is the background information you’d want to bring in to any interaction. For example, an executive director, when weighing the merits of a new initiative, has context such as her organization’s theory of change, their 2030 strategic plan, the org chart, commitments made in various grants, and so on.
In the context of AI agent systems, context is the information that is loaded up every turn in conversation. As illustrated below, the agent retrieves context before doing any of its own reasoning or acting.
growing with use
As a result, context is important to get right: too much background context and it eats up the overall context window, too little and the agent fails to bring in background knowledge you’d want it to.
Connections
Connections describe the services you already use that contain the data essential to your organization. For instance, you may house program and relationship data in Salesforce, outcomes data in Snowflake, and everything else in Google Drive.
Historically, it has been very difficult to get these services to talk to each other, in part because companies generally want to lock in customers by keeping the data inside the system. One of the fortuitous results of the AI boom is that it is forcing a change to this approach. Now, companies know that in order to compete they must make data as available and interoperable as possible, primarily to agents. MCP, or Model Context Protocol, is the open standard developed to let AI models connect to services in this way.
Whether you connect to your data services via MCP, traditional APIs, or even scheduled data dumps, what’s important is to be clear about the source of truth. Consider a proposal that you draft in Google Drive, that gets iterated on via email, and ultimately is executed and saved in Salesforce. In this case, the source of truth for that contract is Salesforce. In general, a good internal policy is to have a single source of truth for the key concepts that your organization is built around; things like customers, outcomes data, financial data, etc.
Now, imagine you’ve got all your data sources hooked up, and you ask a question like “Given the commitments we made in our various contracts in 2025, which did we accomplish and which did we not?” This is the sort of question that would join data across multiple services. What is the right way to answer such a question? This gets into what I see as the heart and the hardest part of knowledge systems, which is retrieval. This will be the subject of a future post, but for now we can just say that there are several architectures: live queries via MCP or API, RAG, and compiled wikis, often used in tandem.
Capabilities
Once we have the elements of the brain in place — context and connections — we are ready to build out the capabilities of the agent system. In a nutshell, agent capabilities are aptly called “tools” — the agent has an ever-growing toolbox of tools, and its job is to correctly use the right tool for the given request. Tools can be very simple — look up outcomes data — or they can be complex and nested: prepare a quarterly board update that includes outcomes data and program updates. If you’ve taught an AI assistant “skills,” tools are the related idea on the action side: a skill is instructions the agent reads, while a tool is a function it can call.
It’s important as you build out your knowledge system that you make it easy for anyone to understand what capabilities an agent has; an agent should not only smartly use the right tool, but it should describe what tools it has, and what tools it doesn’t have. An important design choice is figuring out how you want the agent to act when no tool is available; for example, if I want to conduct a funder landscape analysis but no such tool for that exists, would I rather have my agent let me know it can’t do that, or would I like it to use a generic web research tool and give it its best shot? There’s no single right answer here but two values to enforce are transparency and observability — we should know what the agent is doing and how it’s doing it. Observability will be covered in a standalone post.
Cadence
Finally, we can determine the cadence with which the agent’s capabilities run. Consider your own practices: there are probably some things you do weekly at a given time, other things you do when asked to over Slack or email, and still other things you do when changes in the environment alert you that action is needed. At its best, your AI knowledge system works the same way. Under what conditions should your agent act? These are typically called “triggers” in agentic systems, and getting these right saves your team an enormous amount of time and cognitive overhead because you don’t need to manually nudge the agent every time you want something done.
At their best, triggers and effective decisions around cadence can actually level up the way your whole team operates: for example, imagine each person starting the week with a personalized brief, and ending the week with a personalized reflection, and those practices happening in the places your team already works, such as Slack or email.
Conclusion
I’ve described the 4 Cs linearly, but they are best understood as a self-improving loop: the artifacts produced by the agent’s capabilities feed back into the context, and the system grows with usage. A knowledge system is not a one-time implementation but a living garden that grows with and alongside the organization. And the promise of AI is that for the first time the garden doesn’t require constant, coordinated tending by the team.
All resources →