ERP Data Strategy for the AI Era

You can't run AI on garbage data. Everyone knows this. Nobody's doing anything about it.

I've sat in three dozen meetings in 2025 where a company wants to add AI features to their ERP. Copilots. Agents. Autonomous workflows. The conversation goes great until someone asks: "What's the data quality?"

Long pause. "Um... we'll figure that out."

They won't. Or they will, but it'll take six months and twice the budget they allocated.

Here's what I'm seeing: companies that want AI-powered ERP features but can't move fast enough to build the data foundations. And I'm seeing the companies that did move on data strategy reaping massive advantages.

Why ERP Data Is Terrible

Let me paint the picture. Your company's been running SAP or Oracle for fifteen years. Or you're running five different systems that got "integrated" via middleware.

Your vendor master data has thousands of duplicate records. The same vendor appears as "ABC Inc," "ABC Inc.", "ABC Incorporated," and "ABC, Inc" (with a comma). Your analytics team has a reconciliation process to handle this. It's painful. It works 95% of the time.

An AI system sees four different vendors. It doesn't have your reconciliation process. It sees ambiguity. It halts or makes a bad decision.

Your product codes changed three times as you restructured the business. Old codes are still in historical data. New codes are in current data. A report pulls from both and gets confused. You know this. You document it. You work around it. An AI doesn't know any of this. It trains on the historical data, learns the patterns, and gets the product relationships wrong.

Your transaction data has flags and statuses that mean different things in different contexts. A "hold" on an order could mean "waiting for payment," "waiting for inventory," "waiting for approval," or "customer requested." Your team knows the context. Your ERP field doesn't capture it. An AI learns that "hold" means something, but not what.

Your transactional data spans fifteen years. For the last three years, you've been fixing data quality. But the first twelve years are legacy. An AI trained on all fifteen years learns patterns from bad data. If you exclude the first twelve years, you don't have enough data to train on.

Every company I've talked to has this problem. It's not unique. It's just not discussed loudly because the answer is expensive.

The Cheap Way (Doesn't Work)

Most companies try to add an AI layer on top of their existing ERP. Your data is what it is. We'll build a really smart AI that handles the ambiguity.

That's the strategy for 2025's failure projects.

An AI can handle some ambiguity. But if the ambiguity is structural—if the data itself is inconsistent—the AI either refuses to work or works confidently on wrong information.

You can't LLM your way out of bad data. You have to fix the data.

The Real Strategy

The companies shipping AI-powered ERP features that actually work are doing this:

Step 1: Audit what you have. Not a full data quality assessment (that takes forever). Just: for the data you want to power your AI features with, what's actually there? What's the quality?

For a copilot that helps with PO creation, you need: vendor master (quality?), product master (quality?), pricing (quality?), recent order history (quality?). Each has gaps and inconsistencies. Document them.

Step 2: Build a data layer. Not in your ERP. Above it. A normalized, queryable, consistent view of the data your AI needs.

This could be a data warehouse. Could be a MongoDB collection. Could be a GraphQL layer. The point is: you're creating a single source of truth that your AI reads from.

In that layer, you deduplicate vendors. You standardize product codes. You explicitly map historical codes to current codes. You add the context that clarifies what a "hold" status means in different scenarios.

Step 3: Establish governance. The data layer can't diverge from your ERP forever. You need a process where data flows from ERP to data layer, transformations happen, and the results are consistent.

This is an ongoing process. Your ERP changes. Your data layer needs to adapt.

Step 4: Gate your AI features on data quality. Your AI features don't work on data that fails quality checks. Not grudgingly. Actually not work. The copilot says "I don't have reliable data to help you with this decision. Ask me after..."

That sounds annoying. It prevents hallucination. It prevents the scenario where your AI confidently gives bad advice.

The Timeline Problem

Here's what I'm seeing in practice.

Companies that started building data layers in 2024 are shipping AI features now. Eight to twelve months for a real data strategy is possible if you're focused and you have good data engineering.

Companies starting in 2025 are looking at Q4 or Q1 2026 before they can do anything serious with AI. That's not because it's hard technically. It's because it takes discipline.

The companies shipping fast AI features right now? They picked a narrow domain where they could build a small, focused data layer. Not company-wide. Not enterprise-scale. A copilot for purchase orders. That's it. That data layer is manageable.

The Architecture That Works

The winning architecture in 2025 looks like:

• ERP system: Stores and manages operational data (unchanged)
• Data warehouse/lake: Normalized, consistent, governed (new)
• AI features: Read from data layer, not directly from ERP (new)
• Governance process: Syncs data ERP → warehouse, enforces quality (new)

The expensive part isn't the AI. It's the governance and the data layer.

Most companies underestimate this. They think "we'll use Claude or GPT as our AI layer, so that part's handled." The expensive part is the plumbing beneath it.

What I'd Do If I Were Starting Now

If I were building an AI-powered ERP feature in 2025, here's how I'd budget:

• 30% on defining what data I actually need
• 40% on building the data layer and governance
• 20% on the AI features themselves
• 10% on testing and iteration

Most teams allocate: 5% to data, 70% to AI, 15% to everything else.

That's backwards. The data strategy is the hard part. The AI is the easy part.

Companies that get this right—that invest in data strategy before racing to ship AI—are the ones that ship products that actually work. Not prototypes that fail in production. Real products.

The ERP AI frontier in 2025 isn't about better models or cleverer prompts. It's about companies finally taking their data seriously.

The ones doing that are winning.