All your company's knowledge in one agent

Most companies keep what they know in the heads of their people. This post explains how Hellomatik turns that knowledge into a digital asset: what goes into the base, how it is stored and how answers cite their sources. It also explains why each team talks to its own agent.

Why company knowledge is hard to reach

Company knowledge is usually available to whoever asks the right person, and unavailable the rest of the time.

Most companies answer internal questions the same way: by asking whoever knows. That is an important mechanism, but it does not scale. It depends on one person being available, it interrupts that person's work, and it leaves nothing written for the next time the question comes up.

We call this institutional amnesia. Every time someone has to ask how a thing is done or why a decision was made, that knowledge does not belong to the company. It belongs to a person, and it leaves with that person.

The cost of this arrangement has been measured. In 2018 Panopto estimated an average of 5.3 hours per week lost waiting for information a colleague has or redoing work without it. McKinsey had already measured in 2012 how interaction workers, its term for people whose work is mostly communication and judgment, spend their week:

In other words: roughly six hours out of every ten go to exchanging and locating information, and four go to the role itself. Those six hours are the part of the week the knowledge base is built to absorb: questions stop landing on whoever knows and start landing on an agent that has read the same material.

What goes into the knowledge base

The base is built from material the company already has. Nothing needs to be written from scratch.

Hellomatik accepts the formats companies actually keep their knowledge in: PDF and Office documents, plain text, spreadsheets, complete websites and video or audio recordings. Each format is processed according to what it is:

• Documents. PDF, Word and text files are read as they are and prepared for search.
• Spreadsheets. Tables are imported so that questions about them are answered from the cells, not from a model's memory. A question about a rate returns the figure stored in the sheet.
• Websites. A site is crawled page by page. Repeated navigation menus and footers are detected and removed, so they do not contaminate answers.
• Video and audio. Recordings are transcribed word by word with a timestamp on each word, which is what later allows an answer to point at the exact second of a video.

The practical effect shows up in the question itself. The team stops asking where the file is and starts asking how the task is done. The file arrives on its own, cited in the answer.

How answers cite their sources

The agent consults the base and cites it. It does not memorize your documents.

When someone asks a question, the agent retrieves the relevant fragments, answers with them, and attaches the sources it used. Each source shows the exact fragment that supported the answer. In the panel, the reference goes further: it opens the document at the cited page, or the video at the cited second.

This design has a practical consequence. Because the agent consults rather than memorizes, the base remains under your control: what you upload becomes answerable, and what you fully delete stops being answered.

The difference with a general-purpose model is where the answer comes from. A general model answers with what it learned from the internet. The agent of your base answers with what your company uploaded. When the base does not contain the answer, the agent is instructed to say so. That design also settles ownership: the knowledge lives in the base, outside the model. The model provides the engine; the memory stays with your company.

Why each team gets its own agent

Knowledge is segmented into spaces: separate containers, each with its own documents, its own data connections and its own agent.

A single shared pool would give every user access to everything, and most companies do not want that. Payroll data does not belong in the website chat, and marketing metrics do not belong in the support space. In Hellomatik the unit of separation is the space: a container that holds one team's documents, databases and integrations. The agent of a space can only read what that space contains. Access for people is decided by roles and per-space permissions; departments describe the organization but do not grant access.

Applied to a typical company, the segmentation looks like this:

The value of this segmentation is that it removes the dependency on specific people without creating the opposite problem: knowledge stops depending on who is on duty, and access stays limited to each team's own material.

How automations run on top of the base

The same knowledge that answers questions can also trigger actions.

Classic documentation explains how a task is done and waits for someone to read it. On top of the base, that explanation can become behavior: a flow that performs the task.

Automations in Hellomatik connect to the knowledge base directly. A chat menu option can open a specific document. A booking flow can capture a visitor's details and alert the team. An AI agent step inside an automation answers from the same base as the rest of the system. The mechanics of building those flows are a separate topic; the relevant point here is that the base is not a passive archive. It is the layer the rest of the product reads from.

Caveats and interpretation

Three limits are worth stating before adoption, not after.

The first and most important: organizing the spaces is human work. Deciding what goes into each space and who has access is the part of the system Hellomatik does not do for you, and it is the part that most determines answer quality.

The second: the agent answers with what the base contains, and when it does not know, it is instructed to say so. Those unanswered questions are recorded in the panel. The record is useful precisely because it is a list of missing documents, not a failure log.

The third: this post contains no savings or productivity figures of ours, because it does not measure them. The external figures we cite size the problem and they have aged (McKinsey is from 2012, Panopto from 2018): the mechanism they describe has not changed, but both predate current AI. The count that matters is the one you run with your own numbers, in your panel.

None of these limits weakens the central claim. The asset exists from the first document you upload; what grows with the work is its coverage.

The big picture

Your company already stores documents: in Drive, in Notion, in SharePoint, in shared folders. What is usually missing is not another place to store them. It is a system that understands that material, remembers it when people leave and applies it to real questions. The test is simple: knowledge a machine cannot use is not really organized.

That is what the knowledge base provides: ingestion of the material you already have, answers that cite their sources, a boundary that decides who sees what and one agent per team.

The practical starting point is small: pick one team, upload what it already has written, and ask.