Why Copilot stalls on legacy ERP data, and how Dataverse and Link to Microsoft Fabric each handle a different part of the fix.
Leadership wants Copilot answering inventory and margin questions by next quarter. You’re the one who knows the customer master has three versions of the same account, and half the item data lives in a spreadsheet purchasing keeps on a shared drive. The AI isn’t the hard part. The data is.
Key Takeaways
- AI projects usually fail because of the data, not the AI. Less than 10% of businesses say their data is AI-ready.
- Legacy ERP data is fragmented, and an AI tool pointed at that data answers confidently from whichever copy it finds first, meaning that if AI can’t find the right data, you can’t trust the AI results.
- Dataverse and Fabric do different jobs, and the order matters. Start with Dataverse, because Fabric on top of inconsistent records just gives you faster access to bad data.
- Know what’s ready today before you plan around it. Before choosing tools, answer three questions: which records the AI needs, who owns each one, and which system wins when two copies disagree.
The Gap That Never Makes the AI Roadmap
Gartner predicts that through 2026, organizations will abandon 60% of AI projects that aren’t supported by AI-ready data. (Gartner, February 2025) And very few companies are starting from a clean slate. In a 2026 study by Cloudera and Harvard Business Review Analytic Services, only 7% of respondents said their organization’s data is completely ready for AI. (Cloudera, March 2026)
Those numbers aren’t about weak models or bad prompts. They’re about plumbing. And, if you’re on an IT team, the plumbing has your name on it.
Why Legacy ERP Data Isn’t AI-Ready
Picture a typical distributor on Dynamics GP, NAV, or AX. The ERP has years of custom tables. The WMS, the e-commerce site, the CRM, and the EDI provider each connect through their own hand-built integration. Every one of those systems keeps its own copy of the customer, the item, and the vendor, and the copies drift.
Customizations make it worse. A field that means ship-to region in one report means sales territory in another, and the person who knew why left two years ago. AI can’t read that kind of tribal knowledge.
That’s the most common blocker the research found, too. In the same Cloudera and HBR study, 56% named siloed data or trouble integrating sources as a top obstacle to AI readiness. (Cloudera, March 2026) Point an AI tool at that estate and it will give you a confident answer from whichever copy it found first.
Layer One: Dataverse and the Common Data Model
The first fix is structural. The Common Data Model is a shared, governed schema used across Dynamics 365, and Dataverse is built on it. Customer, item, and vendor records get defined once. Dynamics 365 apps, Power Apps, and Power Automate flows then read and act on the same records, instead of each one keeping its own version.
How you get there depends on the product. Dynamics 365 Finance & Supply Chain Management (F&SCM) uses dual-write to keep its records in sync with Dataverse, plus virtual entities when you’d rather not copy data at all. Dynamics 365 Business Central connects to Dataverse through its built-in integration, where you map tables and set up synchronization.
This layer is about operations. When a rep, a buyer, or an AI assistant acts on a record, it’s the right record.
Layer Two: Link to Microsoft Fabric
Analytics and AI at scale need something different: history, volume, and connections across sources. That’s where Microsoft Fabric comes in.
For both F&SCM and Business Central, Link to Microsoft Fabric connects Dataverse data, including finance and operations tables, to Fabric’s OneLake through shortcuts. (Microsoft, November 2023) There’s no export job, no ETL pipeline, and no third-party tool for your team to maintain. Your data engineers, Power BI reports, and AI models all work from the same governed source. For IT there’s another benefit: getting to your own data is a supported, documented path, not a lock-in trap.
Moving forward, Business Central will be taking a different route. Microsoft is building direct data mirroring from Business Central online into OneLake, and today it’s still in preview. (Microsoft, October 2026) If you’re on Business Central, plan for it. Just don’t build a production AI roadmap on a preview feature until it’s generally available.
Which Layer Comes First?
It’s not Dataverse or Fabric. It’s Dataverse, then Fabric, and the order matters.
If your records disagree, Fabric just gives you faster access to inconsistent data. So the pattern we see work is to clean and govern the master data during the migration, land it on the shared schema, and then link it to Fabric for analytics.
Before you pick tools, answer three questions.
- Which records does the AI actually need?
- Who owns each one?
- Which system wins when two copies disagree?
If you can’t answer those yet, the AI project isn’t ready to start, and that’s useful to know early.
If the near-term ask is Copilot inside daily work, like a sales rep checking availability or a buyer reviewing vendor history, start with clean Dataverse records. If the request is for demand forecasting or analysis across systems, you’ll need Fabric too, built on those same clean records.
Start With the Data You Already Have
Western Computer has spent nearly 40 years and more than 1,750 implementations moving companies off legacy ERP. The AI projects that land are the ones that treat data migration as the first AI project, not a cleanup step to get through before go-live.
For the other risks an aging ERP puts on IT, and how Dynamics 365 handles each one, download our whitepaper, The Distributor’s IT Risk.
