--- title: "The input layer: where AI agents get what your company never wrote down | Spradley" description: "A company brain only knows what your company wrote down. The input layer is how the rest gets in: the reasons people carry and never put in a document." url: https://spradley.io/blog/input-layer-for-ai-agents --- # The input layer: where AI agents get what your company never wrote down A company brain only knows what your company wrote down. The input layer is how the rest gets in: the reasons people carry and never put in a document. Miguel Cabral Oliveira Co-founder & CTO Nov 9, 2026 ![A brain half filled with documents, with a wide stream of speech bubbles flowing in to fill the empty half](https://spradley.io/media/blog/input-layer-for-ai-agents/cover-brain.webp) Ask your company’s AI agent why a team is struggling and you’ll get a tidy answer. It will cite the project plan, the staffing numbers and a few chat threads. It will sound right. It will often be wrong, because the real reason was never typed anywhere. That is the gap in every company brain. This post explains what a company brain is, what a context layer does and which piece most companies are still missing. We call that piece the input layer. ## What is a company brain? Picture the colleague who has been at your company for ten years. They know where everything is, how the work really gets done and who to call when it goes wrong. A company brain tries to give an AI agent that same knowledge, by bringing what the business knows into one place an agent can use. Without one, an agent knows plenty about the world and nothing about you. It has never seen your customers, your rules or last year’s failed launch. ## What is a context layer? A company brain is the goal. The [context layer](https://slite.com/learn/context-layer) is how you get there. It connects to your documents, tickets, chat and other tools. When someone asks a question, it finds the right information, checks that the person is allowed to see it and hands it to the agent. A good context layer is worth having. It also has a hard limit: it can only find what already exists. ## What does a company brain miss? Your tools record what happened. They rarely record why. Here’s an invented example. Say two support teams merge. Three months later, tickets take longer to close. A leader asks the agent why. The agent reads the tickets and the staffing plan and blames volume. It’s a reasonable answer, and it’s wrong. Anyone on the team could tell you what happened. The new bonus counts closed tickets, so senior staff stopped picking up the hard ones. That explanation lives in conversations over lunch, so no context layer will ever find it. ## Why does nobody write it down? There are four reasons, and better search fixes none of them. - **Nobody asked.** No template has a field for “why our bonus is backfiring”. - **It’s awkward.** You won’t write “the bonus broke this team” in a channel your manager reads. - **It’s hard to put into words.** As the philosopher [Michael Polanyi](https://en.wikipedia.org/wiki/Tacit_knowledge) said, “we can know more than we can tell.” - **It costs something.** [Thoughtworks](https://www.thoughtworks.com/insights/articles/context-challenge-ai-agents) points out that experts are being asked to give away what makes them indispensable. Connecting more email and chat won’t help. Those only hold what people were willing to say in front of colleagues. ## What is an input layer? The input layer is how a company brain gets what nobody wrote down. Capturing tacit knowledge for AI agents is its whole job. The context layer reads what exists. The input layer asks for what doesn’t. A company brain needs both. A good input layer does four things: - **It asks.** One person at a time, with follow-up questions, because the real reason is rarely in the first answer. - **It makes answering safe.** If a remark can be traced back to the person, you’ll only hear the official version again. - **It turns answers into evidence.** A pile of transcripts is just more documents. An agent needs clear findings it can cite. - **It asks again.** When something is missing, it goes back to the people who know. The evidence step matters most. [Researchers](https://papers.academic-conferences.org/index.php/eckm/article/download/4875/4580) who tested an AI that interviews employees found that experts liked the interviews, but the summaries had gaps and some invented details. If an agent is going to repeat a finding, someone has to check it first. ## How does Spradley do this? Take that invented support team. Each person gets a personal link and talks one-to-one with an AI interviewer, by text or voice. They’re told it’s an AI. Each question depends on the last answer, so “it’s been harder lately” can lead, a few questions later, to the bonus. Spradley turns the answers into findings: what people see, what they think causes it and what they suggest. Each finding is checked against what people actually said. Now the leader can ask their Personal Agent why tickets are taking longer. The answer draws on what people said, with every point citing its source. If the reports can’t answer, the agent drafts a follow-up interview. Nothing goes out until the leader confirms, and it must reach at least five people. Your other agents can use this too. The [MCP server](https://spradley.io/integrations) lets an owner or manager read their reports from Claude or ChatGPT. It’s read-only and off by default. An owner turns it on, each person consents and reads are logged. ## How are people protected? Nobody mentions the bonus unless they feel safe doing it, so protection comes first. - Personal Agent has no route to one person’s answers. - Ratings, numbers and live views are never shown for fewer than five people. - A separate privacy judge reads each written finding and quote. It rewrites, holds or withholds anything that points at a person. - The agent refuses to identify someone or help retaliate against them. - Names and phone numbers are removed from the copy used for analysis and reports. There are limits, and you should know them. The original answer is also stored, and automated removal can miss a name. A company admin sees who has completed. Spradley’s platform admins can open a single transcript. Our [constitution](https://spradley.io/constitution) sets out what the AI will and will not do. ## What can’t an input layer do? It doesn’t replace your documents or your context layer. It adds what they can’t reach. It only captures what people can put into words. And answers go out of date, so ask again when things change. It also changes nothing if nobody acts. In Spradley, a leader can save a commitment from a write-up and get reminders at four and eight weeks until it is marked done. ## Where should you start? Pick one question your agent keeps getting wrong and work out what’s missing. If it’s a document, your context layer can fix that. If it’s a reason, someone in your company already knows it and nobody has asked them yet. Start with an [AI-moderated employee interview](https://spradley.io/articles/what-is-an-ai-moderated-employee-interview), or [see how an interview becomes context your agents can cite](https://spradley.io/contact). ## Frequently asked questions ### What is a company brain? A company brain is everything a company knows, brought together so an AI agent can use it. It usually draws on documents, tickets, chat and other tools, and it can only hold what someone wrote down unless you add a way to ask people. ### What is an input layer for AI agents? It is how knowledge that was never written down gets into the systems your agents read. It asks people directly, makes it safe to answer, turns answers into checked evidence and goes back when something is missing. ### How is an input layer different from a context layer? A context layer decides which existing knowledge reaches an agent. An input layer brings in knowledge that was not in any system. The first organises what you have. The second gets what you were missing. ### How do you capture tacit knowledge for AI agents? Ask people about specific situations and the reasons behind them, one at a time, with follow-up questions. Then turn the answers into findings an agent can cite, and check them against what was said. ### Can employees be identified when their answers feed an AI agent? In Spradley, Personal Agent has no route to one person’s answers, numbers are never shown for fewer than five people and a privacy judge checks each written finding. There are limits: original answers are stored, a company admin sees who has completed and Spradley’s platform admins can open a transcript. ## Key terms Company brain Everything a company knows, brought together so an AI agent can use it. Context layer The part of an AI system that decides which existing company knowledge reaches an agent. Input layer The part that brings in knowledge no system holds yet, by asking the people who have it. Tacit knowledge What people know from experience and have never written down. ## Other Blog [Explore all resources](https://spradley.io/blog) [ ![A person holding a tablet that shows a Spradley interview question](https://spradley.io/media/jIGKDMMJuCQ9Cs3URn9uyH7hHs.jpg) Innovation ### The research behind AI-moderated interviews for organizations How conversational AI bridges the gap between qualitative depth and organizational scale, and what the peer-reviewed evidence says about its validity. ](https://spradley.io/blog/research-behind-ai-moderated-interviews)