# Company knowledge base > How your company works stops being a folder of documents nobody opens and a settings tree nobody can read. It becomes the system itself, so changing it takes days instead of a project Source: https://aipril.dev/docs/overview/knowledge-base ## What used to happen to it For thirty years the pattern has been the same. Consultants arrive. They interview the people who know how the work is done, and they write it down: process maps, requirement documents, specifications, a matrix of who approves what. Then that is handed to whoever configures the ERP, the low-code tool or the BPM suite, and they translate it into the settings the product supports. At the end of it your operating knowledge exists in two places, and neither is any use. **In documents nobody opens again.** They were true on the day they were signed off. Six months later the process has moved on and the document has not, and nobody knows which parts are still accurate. Nobody trusts it, so nobody maintains it, and it dies in a folder. **And in a settings tree nobody can read.** The system holds the truth, in the sense that it is what runs. It holds it as four hundred configuration screens, and no human being can look at that and answer "what happens when a discount over twenty percent is requested by a regional manager". The only people who can answer are the two who set it up, and one of them has left. The knowledge was never alive. Somebody captured it, translated it, buried it, and reconstructed it by interview the next time it was needed. ## Here, the description is the system Everything on this platform is a description of something real, in a form a person can read: the process and its routes, the record and its fields, the workbook, the catalogue, the report, the page, the agent and what it is allowed to do. Those descriptions are what the system is made of, not documentation *about* it, so there is nothing for them to drift from. So the question that used to require an archaeologist has an answer you can look up. *What happens to a discount over twenty percent?* Open the process. *Who may see the price?* Open the field permissions. *Why does this go to Legal?* The condition is written on the transition, in words. *When did that change, and who agreed to it?* In the version history, with the change request and the person who accepted it. ## And the platform knows the state, not just the shape The descriptions say how the company is meant to work. The platform also holds what is actually happening: which requests are open and with whom, what the balances are, what the catalogues contain, who works where and who is standing in for them, which numbers moved last month. That combination is what makes the assistant useful about *your* company. It reads your processes, your catalogues, your knowledge base and your data, and answers about the situation you are in rather than from a general model of how businesses work. The knowledge side of that is [Knowledge](https://aipril.dev/docs/harness/knowledge.md). ## Why this changes the arithmetic of change You built a CRM the way your sales worked. A year on, sales works differently: a new segment, a different qualification step, an approval that did not exist. In the old world this is a change project: find whoever configured it, work out what the settings mean now, estimate, schedule, regression-test, negotiate. Two things are true here that were not before. **The data is not inside the application.** Your customers, your catalogues, your org structure, your balances and your history live in the platform's own blocks: the data model, master data, the ledger, the process service. An application describes and uses them; it does not own them. So replacing the description does not mean re-entering the data. **And the description is the cheap part to change.** A new version, or a new application over the same model, gets written and reviewed instead of developed, and whatever still applies from the old one carries over instead of being reverse-engineered. So redoing a system stops being the thing you avoid for three years and becomes a thing you do when the business changes. The cost of changing your systems is the ceiling on how fast the company can change. ## The assistant proposes; you decide None of this means the platform tells you how to run your company. The assistant offers the shape it has seen work: the escalation, the rejection path, the control everybody adds after their first bad quarter. That is useful, because most processes are not special and people forget the same parts. But it proposes. Your discount policy has an exception that makes no sense to anyone outside your market and is the reason you hold that segment; keep it. Best practice here is an offer, not a shape the system imposes, which is the difference from products that arrive with somebody else's methodology compiled in. ## What this is for All of the above could read as a better low-code platform. That is not what this is. An AI agent cannot usefully act inside a company it cannot read. To do anything beyond conversation it needs a described world: what the entities are, what the processes are, who is responsible, what it may see, what it may change, what needs a person's approval, and a record of what it did. This platform makes exactly those things explicit, so that **agents can act in them safely.** People can read them too. The design goal is an environment where enterprise agents have somewhere real to work: a described operation, live data, actual permissions, and an audit trail that holds them to the same standard as everybody else. The applications are what you get; the described, governed operation underneath them is what makes the agents possible. ## Where to go next How an application is described, reviewed and built is [Built by AI, approved by you](https://aipril.dev/docs/apps/authoring.md); what one is as an object is [Under the hood](https://aipril.dev/docs/apps/under-the-hood.md). The retrieval side of the knowledge base is [Knowledge](https://aipril.dev/docs/harness/knowledge.md).