How Plainwater AI works: from question to dashboard, with the gateway in between
Ruben, a BI consultant at Meerwijk Care Group, gets a question from the controller: why has the absence report been loading so slowly for two weeks? That used to mean an afternoon in query logs. Now he types the question into a conversation.
This page follows what happens next, step by step. Every step passes through one gateway that controls what reaches the model.
Updated .
What route does a question take?
A question travels from the conversation to an agent, which works on the source through the AI gateway, proves its proposal in a sandbox and puts the result in front of a person. Whatever lasts becomes a dashboard.
- Question. Ruben asks his question in the web app or in his own AI tool.
- Agent. A first responder answers simple questions straight away; for this job it starts an agent.
- Gateway. Every model request and tool call gets one decision under Meerwijk's rules.
- Source. The agent reads metadata, query statistics and missing indexes, with pseudonyms instead of real names.
- Sandbox. Its proposal runs in a disposable copy, is measured and compared with the original.
- Person. Ruben sees the verdict, reads the proposal back in real names and approves it.
- Dashboard. The measurement becomes a dashboard that refreshes every morning, with no model running.
What does the gateway do between agent and model?
The gateway gives every request exactly one policy decision and stops what is not allowed. A refused request never reaches the model provider.
Table names reach the model as pseudonyms, rows stay behind by default, and counts below ten become "<10". What Ruben reads back is in his own names again.
Why a sandbox before anything goes live?
Because a proposal is only worth something once it has been measured. The agent runs its change in a disposable copy, measures runtime and result, and compares them with the original.
The customer chooses what goes into the copy: synthetic data, synthetic data shaped to the real statistics, or real data. The copy is always cleaned up afterwards.
Which layers make up the platform?
The platform has nine layers, and each one answers its own question from a team working with data and AI. Together they form one environment.
| Layer | Question it answers |
|---|---|
| The AI gateway | What reaches the model? |
| Set up once | How does everyone work the same way? |
| AI for data development | How does AI help with engineering and analysis? |
| The AI layer on Yres | How does AI work with our data warehouse? |
| Dashboards | How does an analysis become a lasting report? |
| Security | What guarantees are there, and are they tested? |
| Ways of working | Can I use my own coding AI? |
| Any model | Am I tied to one vendor? |
| European | Where is the data processed? |
Who works in the platform?
Everyone who works with the data, in their own way. Ruben works in the web app, Mila in her editor with Claude Code, and the controller looks at dashboards and asks questions about them.
Administrators set the platform up once: sources, data access, rules, roles and the skills that capture how Meerwijk works. From then on that applies to every agent of every colleague.
Where it stops
Plainwater AI works on your existing warehouse, ETL and reporting. It does not replace them.
An agent proposes and proves; writing to a source only happens as a per-source permission, through a proposal a person approves.
And the gateway controls what the platform itself passes on. What you do with your own AI tool outside the platform falls outside it.
Frequently asked questions
Do I need a new data warehouse for Plainwater AI?
No. Plainwater AI works on your existing sources, such as SQL Server, Azure SQL, PostgreSQL and SAP HANA Cloud, and through connectors on tools like Power BI and Azure Data Factory.
Does an agent make changes to production by itself?
Only if that permission has been granted for the source, and then through a proposal a person approves. By default an agent reads and proves.
What does a dashboard cost in model usage?
Only when creating or changing it. Refreshing runs on a schedule without any AI model.
Can I see the platform at work?
Yes, in a demo we show a question from start to dashboard, on sample data.