Daivio AI Analyst Launches for Microsoft Fabric: OneLake Data Residency and Security Questions

The idea is simple. You open a file from your Lakehouse, clean it up, ask a question in plain English, and get back an answer along with the SQL that produced it. You never leave the Fabric workspace. That promise is worth a close look, especially the company's claim that your data "never leaves Microsoft OneLake." Microsoft's own documentation on how Fabric workloads are built adds important context to that claim.

What Daivio says it does​

According to the company, Daivio runs as an item inside a Fabric workspace. Users sign in with their existing Microsoft Entra ID accounts, so there's no separate login. The announced features are:

• Bring data in: pick a Lakehouse file, Delta table, Warehouse or SQL database, or upload a file.
• Profile it: a dataset profile with insights for each column, a summary written by AI, and a scan that flags columns containing personal data.
• Clean it: eighteen one-click transformations, with a timeline that records every step.
• Ask in plain language: type a question and get the answer, the SQL behind it, a result table and a chart.
• Go deeper: a SQL playground, Jupyter notebooks, and a Save to Lakehouse option that writes results back to OneLake as Delta tables.

The Microsoft Marketplace listing for "Daivio SaaS" backs up most of this. It describes importing from a Lakehouse, Warehouse or SQL database, or uploading a CSV or Excel file. It lists column profiling that covers distributions, unique values, outliers, correlations and detection of personally identifiable information (PII). It also lists these features:

• guided cleaning steps such as trim, case changes, split, round, regex columns and fill missing, all recorded in a History view
• a SQL playground whose output can be saved to a Lakehouse as a Delta table
• scatter, line, bar and pie charts
• plain-language chat that returns the table and the SQL behind each answer
• a Jupyter notebook for code-first work

The listing does not confirm the "eighteen" transformation count. That number comes only from the company.

Section summary: The core features are described consistently in the announcement and the Marketplace listing. The exact transformation count, the accuracy of PII detection and the quality of the generated SQL are all vendor claims that nobody has independently tested.

Under the hood: what "inside Fabric" actually means​

Daivio's CTO, Ashutosh Mishra, says the product is built on the Fabric Extensibility Toolkit so it behaves like any native Fabric item, with the "same workspace, same identity, same capacity, same Microsoft OneLake." CEO Drar Amreh goes further and says the data never leaves OneLake.

Microsoft's documentation gives useful context here. Microsoft defines a workload as a partner- or customer-built web application that integrates into the Fabric portal, and the Fabric host loads your app according to its manifest, provides authentication tokens via Microsoft Entra ID, and exposes a host API to enable navigation, theming, notifications, and other platform interactions.

The key detail is where the code runs. Microsoft's overview says workloads are hosted by you, rendered in Fabric via iFrame - Your web application runs on your infrastructure but displays within Fabric. The architecture page adds that the app uses Microsoft Entra ID tokens (provided by the host) to call Fabric public APIs and, if needed, your own back-end services. The same Microsoft overview puts it bluntly: a workload is embedded in Fabric, but the vendor's code runs in the vendor's cloud.

This doesn't mean Daivio's data-residency claim is wrong. Microsoft does say workload items participate in collaboration and sharing, are discoverable in search, follow lifecycle operations and governance, and store data via OneLake while using Fabric public APIs for CRUD. Storing data in OneLake is one thing. Where the compute happens, and where your data goes while it's being profiled, cleaned or sent to an AI model, is a separate question. Running inside a Fabric iFrame doesn't answer it either way.

Permissions are declared up front. According to Microsoft, the workload's manifest declares the permissions it needs; the Fabric host obtains the appropriate tokens and enforces consent and access according to those declarations. Admins can review those requested scopes before approving the workload.

Section summary: "Inside Fabric" means a real, native-looking experience that uses your Entra identity. Under Microsoft's architecture, though, the app itself is hosted by the vendor. The OneLake-only claim is the company's own and needs checking before you use sensitive data.

Pricing and availability​

The company describes three tiers:

The Marketplace listing uses slightly different wording. It says to start on Free and upgrade to "Pay-Per-Use." Neither source gives the per-minute rate or explains what counts as "active processing." When the listing was checked, it showed a 5.0 rating from a single review, which says nothing yet about how widely the product is used.

One general platform limit applies: Microsoft states that the Fabric Extensibility Toolkit is not supported in Power BI Pro capacities. That's a platform rule, not a published Daivio requirement, so check the vendor's actual prerequisites against your capacity SKU.

The analyst's pitch versus the admin's checklist​

The problem Daivio targets is real. The company says analysts lose hours fixing headers, data types, duplicates and missing values, and that every file exported out of Fabric creates a governance question. The company offers no study to support the "hours" figure, but anyone who has received a CSV with three date formats in one column will recognise the complaint.

The more useful feature may be the audit trail. Returning the SQL and a notebook with every AI answer turns a black-box answer into something a colleague can review, rerun and check. Be clear about what that does and doesn't prove, though. Being able to read the SQL doesn't make it correct. An AI-written query that joins on the wrong key still looks tidy.

Before rolling Daivio out beyond a test workspace, Fabric admins should:

1. Map the data flow. Ask the vendor where compute runs, which region it runs in, whether any AI model processing happens outside your tenant, and what gets retained or logged.
2. Review manifest permissions. Check which Fabric API scopes the workload requests and whether they match what it actually does.
3. Test PII detection against your own policies. A vendor's personal-data scan is a starting point. It doesn't replace Purview classification or your compliance team's definitions.
4. Check the SQL. Treat AI answers as drafts until someone who knows the data confirms them.
5. Pin down billing. Get the per-minute rate and the definition of "active processing" in writing before the Pro tier ends up on a shared capacity.
6. Pilot on non-sensitive data. Use the Free plan in a sandbox workspace before letting it anywhere near production Lakehouses.

The bigger picture​

Daivio is one example of how Microsoft is positioning Fabric as a platform, not just a product. Microsoft's toolkit documentation openly invites solution builders to sell through the Workload Hub. The result is a growing marketplace of specialist tools that sit alongside Microsoft's own Copilot and Data Agent features in Fabric.

For customers, that means more choice, and each partner workload needs the same scrutiny you'd give any SaaS vendor with access to your data. Daivio's clearest selling point is packaging profiling, cleaning, plain-language querying, readable SQL, notebooks and Lakehouse write-back into a single workspace item. Whether it delivers on accuracy and data handling is something only independent testing will show.

Are you running third-party Fabric workloads, or sticking with Microsoft's built-in tools? Tell us in the forum.