IntuigenceAI Industrial AI Workers Launches as a Microsoft Fabric Workload

The customers are refineries, chemical plants and energy operators. The people who will actually enable it, though, are Fabric administrators, and they have some decisions to make before anyone starts asking an AI where pump P-210 discharges.

What IntuigenceAI says it built​

The company describes its customers' problem this way: process and energy operators hold almost everything they need to know about their own plants and can query almost none of it. In IntuigenceAI's words, the record sits across thousands of engineering drawings, P&IDs, process flow diagrams and electrical schematics among them, alongside historian tags, datasheets, and shift logs. The company says those formats aren't ready for general-purpose AI.

There's also a staffing argument. The release says half the process-engineering workforce will reach retirement age this decade, and most of the knowledge they hold will be lost as it was never documented. That figure comes from the company. No source is given, and we didn't verify it.

The headline claim is speed. According to IntuigenceAI, the platform compiles a plant's scattered engineering record into a single knowledge graph in under 48 hours, work the industry still scopes in quarters. AI agents the company calls "Intuigents" then sit on top of that graph to answer questions and do engineering work.

Treat the 48-hour figure and the "executes engineering work" language as vendor claims. We found no independent benchmark, customer case study or third-party test that backs them up.

Summary: This is a knowledge-graph-plus-agents product for process industries. The speed and capability claims come from the vendor, not from independent testing.

Two deployment models, two data boundaries​

IntuigenceAI offers more than one way to deploy, and the data boundaries aren't the same in each. The product runs standalone in a customer's own cloud, and it also runs inside Microsoft Fabric, for operators whose data is already in Fabric OneLake.

For the Fabric version, the company says data customers store in Fabric OneLake stays in their own Fabric tenant. IntuigenceAI reads it in place rather than copying it, and raw drawings and files never leave the customer's tenant. The release also says:

• Access goes through Microsoft Entra ID, so answers respect the permissions the customer has already set.
• No customer data is used to train any model, including IntuigenceAI's own.
• Customers with stricter requirements can run the whole platform inside their own Azure tenant, model inference included, with per-tenant isolation and a complete audit trail.

Read that last point carefully. Full in-tenant inference is described as an option for stricter requirements. It isn't presented as how the standard Fabric workload behaves.

The phrase "raw drawings and files never leave the tenant" also leaves questions open. What about the derived knowledge graph, prompts and responses, telemetry, support logs and the inference traffic itself? Security teams should get those answers in writing before signing.

What you actually get inside Fabric​

The Marketplace listing calls the product IntuigenceAI Industrial AI Workers. Its listed functions are:

• chatting with PAM, the "Plant AI Manager," and with specialist synthetic engineers
• analyzing P&ID diagrams and equipment connections
• querying real-time telemetry from Eventhouse
• getting answers with source citations

It also lists two Fabric item types:

The press release adds more Fabric-specific features:

• live plant state through Fabric Eventstream and Eventhouse, so the company's Maven agent works from current conditions rather than a snapshot
• an industrial ontology grounded in customer data through Fabric IQ
• graph visualizations at plant level and per drawing
• voice and video calls with a synthetic engineer without leaving Fabric

The Eventstream and Eventhouse pairing fits Microsoft's own documentation, which says Eventstreams can send real-time events to an Eventhouse for querying with Kusto Query Language (KQL). That shows the design is plausible. It doesn't tell you which plant historians or control systems IntuigenceAI connects to, or how "live" live really is. The release gives no latency figure, and you shouldn't assume one.

Summary: You get two item types, a catalog and a board, plus agent chat with cited answers. The real-time and ontology features rely on Fabric services you may also need to license and configure.

How to enable and start the workload​

IntuigenceAI's documentation says the workload works with Lakehouse and can be used in any workspace once a Fabric administrator has enabled it. If a user sees "Contact your admin to configure this workload," the admin hasn't done that yet.

Microsoft has said that the workload hub, introduced with the Fabric Workload Development Kit, allows you to view all available workloads within Fabric. Each workload offers unique item types that you can create within Fabric workspaces. Microsoft Learn adds some rules:

• To add a workload, a user needs permission from a Fabric admin, and must be a capacity admin or allowed to assign capacities to workspaces.
• Workloads are assigned to specific capacities. Only workspaces on those capacities can create the workload's items.
• The first time someone uses a workload, they must consent to its Microsoft Entra ID application. Tenant admins can consent for the whole organization.

The vendor's documented first-run steps:

1. In Fabric, select Workloads in the left sidebar, find IntuigenceAI Industrial Workers and select Learn more.
2. Select Open in the Get Started panel and pick a workspace with a valid Fabric capacity.
3. Create a Data Catalog item to store and index plant files. Supported types are P&IDs as PDF, PNG or JPG; timeseries exports as .xlsx or .csv; and general documents such as manuals and SOPs.
4. Create an Intelligent Board, link it to the catalog, and start an agent session. Type @ to mention a specific AI worker.

When a file shows Success in the catalog, it's indexed and PAM can query and cite it. In progress means it's still processing. The vendor says big multi-sheet P&IDs can take several minutes, so refresh to check.

The documentation says you need a paid or trial Fabric capacity. It notes that free capacities "may have limitations," without spelling them out. The public material doesn't give a minimum SKU, regional availability, file-size limits or pricing.

What success looks like: a P&ID shows Success, and PAM can answer something like "what's upstream and downstream of HX-301?" with a citation you can click through to the source drawing.

Support and SLAs​

Support comes from IntuigenceAI, not Microsoft. Its documentation says you don't need a Microsoft support case for problems that start in the workload. Published initial-response targets are one hour around the clock for Severity 1 (workload unavailable, or suspected data loss) and four business hours for Severity 2. The company also points to a 99.9% monthly availability target. These are response targets, not fix times.

The quotes, and how much weight they carry​

Microsoft's Dipti Borkar, Vice President for Microsoft IQ and OneLake, is quoted in the release: "Fabric is at its strongest when partners build natively on it, operating on data that stays in OneLake and stays open." A former Saudi Aramco engineer adds a short endorsement.

A Microsoft executive quoted in a partner's release confirms the partnership exists. It doesn't validate the engineering. The same goes for the claim that IntuigenceAI is backed by Eric Schmidt's venture firm and has been running in production plants since 2024. We couldn't confirm either claim independently.

Analysis: who should care​

If your plant data already lives in OneLake, a native workload that reads it in place is a real draw. It means one less copy of sensitive engineering data sitting in some vendor's cloud. The harder question is trust. AI answers about piping and instrumentation are high-stakes, and citations only help if engineers actually check them. This is a tool to assist engineers, not one that signs off on management-of-change (MOC) paperwork.

Before you roll it out:

• Pilot it on one capacity, not the whole tenant.
• Read the Entra consent permissions line by line.
• Get the data-flow and inference boundaries for your chosen deployment in writing.
• Test the agents on drawings where your senior engineers already know the right answers.

IntuigenceAI is also at the European Microsoft Fabric + SQL Community Conference in Barcelona, which ends today. It's co-hosting an invitation-only Industrial AI Summit with Microsoft in Houston on November 4.