Fabric Lab – Issue 01: Your Copilot Is Not Broken. Your Data Is.

𝗪𝗵𝗮𝘁 𝟵𝟰% 𝗼𝗳 𝗺𝗮𝗻𝘂𝗳𝗮𝗰𝘁𝘂𝗿𝗲𝗿𝘀 𝗴𝗲𝘁 𝘄𝗿𝗼𝗻𝗴 𝗯𝗲𝗳𝗼𝗿𝗲 𝘁𝗵𝗲𝘆 𝗲𝘃𝗲𝗿 𝗼𝗽𝗲𝗻 𝗮 𝗖𝗼𝗽𝗶𝗹𝗼𝘁 𝗹𝗶𝗰𝗲𝗻𝘀𝗲.

A COO at a $180M manufacturer called me last month.

His CIO had put nine months and roughly $400,000 into a Copilot rollout across the plant floor. Executive dashboards. Agent flows. The full deck. It looked like a win at the go-live meeting.

Then a shift supervisor asked it one question.

“How much aluminum coil do we have on hand?”

Copilot answered fast. It answered confident.

It was off by 62,000 pounds.

“We built the future,” the COO told me. “On top of yesterday’s mess.”

He is not alone. He is the 94%.

MIT NANDA published the number this year. 95% of GenAI pilots delivered zero P&L impact. HiveMQ found that 68% of manufacturers cannot scale a pilot past proof-of-concept. Dataiku says 71% of CIOs must show AI value by mid-2026 or lose the budget for it.

Under all three numbers, one pattern.

The record is broken. And no agent can fix a broken record.

This is Fabric Lab, Issue 01. I run a boutique Microsoft Data and AI shop that has spent the last decade cleaning up other people’s rollouts inside mid-market manufacturing. What I publish here is what most Microsoft partners will not.

Let me start with the one that matters most in 2026.

Why AI Stalls At The Plant

Every AI rollout I have watched fail inside a manufacturer failed at the same layer.

Not the model. Not the license. Not the change management deck.

The master data.

The part number that means one thing in ERP, a different thing in the MES, and a third thing in the tribal spreadsheet the warehouse manager keeps on his desktop.

The customer name spelled seven different ways across four systems.

The supplier ID that got merged in an acquisition six years ago and never reconciled.

The chart of accounts sitting at 340 line items because nobody had the appetite to retire the ones from before the ERP swap.

Copilot has no way to know any of this. Neither does your data agent. They read what you show them. If what you show them is a mess, they will hand you back a confident, articulate, well-formatted mess.

That is why 94% of manufacturers are running GenAI and only 2% have operationalized it.

The 92% caught in between are the ones who skipped the boring step.

Fix The Record. Then Turn On The Agent.

Here is the boring step, cut into five moves. Every one of them is a project you can start this quarter, without buying a single AI license.

1. Name your golden data source

For every domain that touches your P&L (customer, product, supplier, employee, inventory), one system is the master. Every other system reads from it. Write that down on one page. Get the domain owner to sign.

That is the hardest hour on the list.

2. Reconcile the attributes

The part numbers. The customer IDs. The GL codes. Run the match. Publish the exceptions. Then fix them.

Do not skip to step three because “the exceptions are manageable.” Those exceptions are exactly why your agent is about to lie about inventory.

3. Consolidate it in one place

This is where Microsoft Fabric earns its keep. Not because it is magic. Because it stops the copy-paste.

Every system, every night, into one lakehouse. One version of the truth. One place to fix a problem, not seven.

4. Instrument the drift

Data quality is not a one-time project. It rots.

Put a simple dashboard on top of the lakehouse that tells you, every morning: how many duplicate customers showed up, how many part numbers failed validation, how many GL entries hit a retired account. You cannot manage what you cannot see.

5. Now turn on the agent.

Copilot on this stack is a different product than Copilot on the mess. You get answers that match your books. Your COO stops hedging in the board meeting. Your plant manager trusts the number on the screen.

None of this is glamorous. None of it demos well at the vendor conference.

All of it is why the 2% who made GenAI work look like magicians, and the 92% who skipped it look like they wasted a year.

What This Means For You

You are probably one of two CIOs.

If you have not started with Copilot yet, the good news is you have not paid the tuition the 92% paid. Do the five moves first. Every dollar you spend on steps one through four buys you ten dollars of AI value later. Public Forrester Total Economic Impact studies on modernized analytics platforms report ROI ranging from roughly 200% to over 400% over three years. There is none published on running agents against a broken record.

If your Copilot rollout is not landing, you did not fail. You skipped a step nobody told you was a step. Go back and do it. The people telling you “the AI just is not there yet” have it exactly backwards. The AI is there. Your record is not.

Either way, you are not behind. You are exactly where the honest work starts.

The One Ask

If you are running or planning a Fabric or Copilot rollout inside a $50M to $500M manufacturer, I am opening five slots this month for a Data Readiness Audit.

Two weeks. Fixed scope. A scored assessment of your master data on the exact five moves above.

No pitch at the end. You get the scorecard whether you decide to work with me or not.

Reply here or DM. First five, first booked.

And if this issue landed with you, tell me in the comments. What is the one part of your record you already know is broken, and everyone at the table is pretending is fine?

I read every reply.

Next issue: The F64 Cliff. The Fabric pricing cliff Microsoft partners will not publish, and how to right-size before you sign.

Sources

Click here to read this article on Dave’s Demystify Data and AI LinkedIn newsletter.

Scroll to Top