Welcome to the Plainwater AI knowledge base
Plainwater AI is one platform where your team works on data with AI: from a question about a table to a dashboard that refreshes every morning. The platform sits between your data and every AI model, and it watches what goes to that model. You decide per source what the AI gets to see; the platform enforces it on every request.
This knowledge base explains how that works. No marketing language, just the mechanics: which rules apply, who chooses them and where they are enforced.
In one sentence
AI for your data, without your data walking out the door.
What the platform does
- A gateway for every AI request. Every request to a model goes through one gateway. Each request gets exactly one policy decision: is this model allowed, does it fit the budget, does it match the customer's rules. A refused request never reaches the model vendor. See The AI gateway.
- Names become pseudonyms. Table and column names go to the model as pseudonyms. The answer is translated back into your own names for you. See Pseudonymisation.
- Data access per source. Per source you choose a mode: metrics only, encrypted values or open. The strictest rule always wins. See Data access.
- Working on data. Agents read metadata, check data quality, investigate performance and prove their proposal in a disposable copy before anyone runs anything on the real database.
- Dashboards that keep running. An analysis ends in a dashboard that keeps running on the source, not in a picture with baked-in numbers.
- Any model, any way of working. Claude, GPT, Gemini, Mistral or any other model with an OpenAI-compatible API. Work in the web app, together with your own coding AI, or through MCP only.
Where to start
| You are… | Start at |
|---|---|
| New and setting up the platform | Getting started |
| Responsible for privacy or security | The AI gateway, then Data access |
| A data engineer or BI developer | Work modes and The verification sandbox |
| Managing a team | Skills and knowledge and Environments |
| Looking for a term | Glossary |
| Looking for a quick answer | Frequently asked questions |
How this knowledge base is organised
- Getting started — five steps from zero to your first task.
- Concepts — the ideas everything rests on: the gateway, pseudonymisation, data access, work modes, the sandbox, skills and knowledge, environments and monitors.
- Dashboards and datasets — how an analysis becomes a dashboard that keeps running.
- Working with your own AI — the three ways of working and how to connect your own AI tool.
- Models — which models are available and how to choose per customer.
- Sources and connectors — databases, connectors and your own MCP servers.
- Security and privacy — the guarantees, the audit trail and European processing.
- Yres — the AI layer on the Yres data warehouse.
- Administration — roles, permissions and ground rules.
Three principles
- Rules are enforcement, not instructions. A rule that only lives in a prompt can be ignored by a model. That is why every rule lives in the gateway, where the model cannot get around it.
- Metadata first. For a lot of work an AI needs no rows: structure, counts and statistics are enough. Only when that is not enough, and only if the customer allows it, does more go to the model.
- Everything can be read back. Prompts and answers are stored encrypted, and whoever is entitled can read them back. Every read is itself recorded.
Rather see it than read about it?
Use Request a demo at the bottom of this page. We will show you the platform on a sample database.