Responsible AI
FlowStudio AI (Flowra) · Last updated 3 August 2026
This page is about FlowStudio AI, the chat product at ai.flowstudio.app whose agent is called Flowra. FlowStudio MCP and the FlowStudio app work differently and have their own page — the important difference is that they do not run a language model, and this product does.
We run the model
Flowra is not a connector you point your own AI at. We operate the language model on our side, and our agent decides which tools to call and in what order.
The models run on Microsoft Azure OpenAI, inside our Azure subscription. Your prompts are not sent to OpenAI's consumer services.
Where that happens: East US. We are an Australian company and the rest of the product runs in Australia East, but the Azure OpenAI capacity we use is in the United States, so the text of a question and whatever flow data was read to answer it is processed there. If your organisation has data residency requirements, this is the part to check. The security review covers where everything else lives.
Which model handles a question depends on what you asked. Simple lookups go to a smaller, cheaper model; diagnosis and flow changes go to a stronger one. You can pick a different model in settings, and the credit cost changes with it — the picker shows the multiplier before you choose.
What we send it
To answer a question about a flow, Flowra sends the model your message, the conversation so far, and whatever it read from Power Automate to answer you — flow definitions, run history, action inputs and outputs.
That means the contents of your flows and their runs are sent to the model. If a run passed customer records through an action, and Flowra reads that action to diagnose it, those records are in the prompt. This is what makes the diagnosis useful, and it is worth knowing before you point it at a flow handling sensitive data.
It never trains on your data
- Your conversations are never used to train or fine-tune any model — ours or anyone else's.
- Your flow contents are never used for training.
- Azure OpenAI does not train on the prompts we send it.
There is a setting in the Conversations panel — "Let the FlowStudio team review my conversations to improve the product" — which is off unless you turn it on. Turning it on lets our team read your conversations to make the product better. It still does not train a model on them. Turn it off any time; it applies from that moment.
It asks before it changes anything
Flowra can read your flows freely. Changing one is different: it does not make a change unless you asked for it or agreed to it.
When it proposes a fix, you get a card with the change described and two buttons. Nothing is applied until you press Yes, go ahead. If you type a question instead, the offer disappears.
This is enforced in the product, not just written down here. Before any write reaches Power Automate, a separate check reads the recent conversation and decides whether you actually asked for it. If you did not, the write is refused, and Flowra answers your question and offers instead. Reads are never blocked by this.
Other checks refuse specific kinds of damage regardless of what the model intended: writing a flow definition that would delete existing actions, changing a flow's contract in a way that breaks its callers, replacing a live flow with an empty one, and several more. When one of these fires, the change is rejected and you are told why.
It works with your permissions, not its own
Flowra reaches Power Automate as you, through Microsoft delegated permissions you approve — or through the browser extension, where the access token stays in your browser and we never hold it. Either way it can only see and change what your own account can see and change. It has no standing access of its own, and revoking access in your Microsoft account or in Flowra stops it.
What it gets wrong
Flowra is a language model reading real data. It is usually right and it is not always right. In particular:
- It can misread an error and name the wrong cause, especially in flows with nested loops or child flows.
- It can build a flow that saves and starts but does not do what you meant.
- It can state something confidently that a later check contradicts.
- It only knows what the Power Automate API exposes. Some flows return definitions the API will not show in full, and it will tell you when that happens rather than guess.
It is built to report what it observed rather than what it expects: a change that saved is not reported as working until the run is checked, and a run that is still going is not reported as succeeded. If it tells you something you know is wrong, say so — it is instructed to re-check rather than repeat itself.
Treat it as a fast, well-informed colleague rather than an authority. For anything that matters, look at the run it points you to; it gives you the link.
What we keep, and for how long
Conversations are stored encrypted so you can pick up where you left off, and deleted after 90 days. We also keep usage numbers — credits, token counts, which tools ran. The full detail is in how we handle your data.
Telling us something is wrong
There is a Report a problem option in the account menu, which sends us the conversation you are in. Or email [email protected]. If Flowra changed a flow in a way you did not expect, tell us — that is a defect in the approval gate and we want to know.
Who we are
FlowStudio AI is built by Flow Studio Solutions, an independent Australian company. We are not affiliated with or endorsed by Microsoft. Power Automate, Azure and Microsoft 365 are Microsoft products.