Exact Online, right inside ChatGPT: DataFlowr is now an official plugin
Until now you connected ChatGPT to Exact Online through developer mode, with an address you entered yourself. From now on it is one button: DataFlowr has been approved as an official plugin in the ChatGPT plugin store. The video below shows the whole route, from installing to the first analysis in the conversation.
In the August release article we wrote that listings had been submitted to the three assistants’ vendor marketplaces, and that the card would become one button as soon as the first was approved. That first one is in: the DataFlowr listing in the ChatGPT plugin store has been reviewed by OpenAI and is live. This article, and the Dutch-spoken demo above, walk through what that changes in practice, and how voice fits in.
From entering an address to one button
Installing is now: open the DataFlowr listing in the plugin store, click Install plugin and sign in with your DataFlowr account. The route through developer mode, where you created a connector by address yourself, is no longer needed. For an administrator that saves instruction work; for everyone who is not an administrator, the connection becomes something you set up yourself in a few minutes, for the first time.
The sign-in deserves its own paragraph, because that is where the security sits. Access to the administration runs through Exact Online’s official authorisation: you sign in on Exact’s own screen, and the connection then works with exactly the permissions of your account. ChatGPT never sees a password for your books, and no copy of your administration moves anywhere: every question is a targeted query at the moment you ask it. Revoking is always possible, under Sessions on the connection page.
"The approval changes nothing about what the connection may do. It changes who can switch it on: installing is one button, and the permissions stay those of your own account."
What happens in the demo
The demo starts where every month starts: at the profit and loss statement. One message opening with @DataFlowr, so ChatGPT knows the answer must come from the administration, and the request for the figures with an analysis. Back comes a table with the period set against the same period last year: revenue, gross margin, operating costs, EBITDA and net result, with the change per line. Below it ChatGPT names what stands out: where the growth sits, what the margin is doing, which cost line is out of step.
The follow-up question shows why a conversation works differently from a dashboard: turn this into a management summary, with action points. ChatGPT builds on the figures already in the conversation, so no new export is involved. The action points are starting points for investigation, not conclusions: which margin deserves an explanation, which cost line you want to keep watching. The judgement stays with you.
Voice on top: speaking instead of typing
You can also speak to ChatGPT, and that makes this connection useful on the move for the first time. The route is then: you speak your question, ChatGPT turns it into a structured instruction, the plugin fetches the data from Exact Online or prepares an action, and the answer comes back as text you can have read aloud. Which invoices are still open, what is this month’s revenue, look up a customer: voice is the control layer, the work happens in the same secured connection, within the same permissions.
Being honest about where this stands today: ChatGPT’s dedicated voice modes do not support plugins everywhere yet. What works reliably is speaking your question into the ordinary conversation and having the answer read aloud. So test the voice mode in your own environment before building it into your working day. And mind where you are: having figures read out in an open space shares them with everyone within earshot.
Consulting, preparing, executing
Not every instruction carries the same weight, and the connection does not treat them the same either. The distinction, in three steps:
- Consulting. Balances, outstanding items, reports, customers. Reading changes nothing in the administration, and it is where almost every conversation starts.
- Preparing. A draft invoice or a proposed entry. The draft sits ready in Exact Online and waits for review; nothing is final yet.
- Executing. Posting, changing, sending. That comes with an explicit confirmation: the plugin summarises what is about to happen, with the amount and the administration attached, and you say yes or no.
Which instructions are available depends on your Exact Online modules and on the permissions of your account: what you may not do, the plugin may not do either. And everything can be traced afterwards, because the connection’s audit log records which assistant made which call.
What this is and is not
The video runs on a demo administration, so the amounts are not real; the route from installing to summary is exactly how it works on a real one. And a language model can misread a question: a period, an administration, a VAT rate. Critical outcomes you therefore verify in Exact Online itself, the way you would with the work of a new colleague.
The plugin store is a route, not an obligation. The same connection works in Claude and Microsoft 365 Copilot, because MCP is an open standard; the listings there are still pending. If you work with another assistant today, you connect through the routes in the documentation.
Try it yourself
Open the DataFlowr listing in the ChatGPT plugin store, choose Install plugin and sign in with your DataFlowr account. Then start with the question from the video: show the profit and loss statement and add an analysis. The installation is documented step by step, including what you need and how to revoke access again.
