Before You Add AI to Your ERP: The Data Problems It Will Expose First

Most small and mid-size companies now see AI features inside their ERP: demand forecasts, invoice capture, assistants that answer questions about stock or orders. The pitch is simple. Switch it on, and the system starts making predictions and suggestions from data you already have. 

The catch is in that last phrase. AI does not fix the data you already have; it reads it, learns from it and repeats its mistakes at speed. For many businesses, the first thing an AI add-on reveals is not a new insight but an old data problem that people had been quietly working around for years. 

The research points the same way. Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, and poor data quality was the first reason it listed (Gartner, July 2024). A later Gartner survey found that 63% of organizations either do not have, or are unsure whether they have, the right data management practices for AI (Gartner, February 2025). 

In ERP systems, three data problems come up again and again. Each one is fixable, and each is far cheaper to fix before an AI feature goes live than after. 

Problem 1: Master data that disagrees with itself 

Master data is the reference information everything else depends on: customers, suppliers, products, units of measure and price lists. In a growing business, it tends to drift. The same customer exists three times with slightly different names, a product is sold in boxes but stocked in units, and old items that were discontinued years ago are still marked active. 

People cope with this by memory. The sales team knows which of the three customer records is the real one, and the warehouse knows that “box” means twelve. An AI model has no such memory, so it treats every record as a separate fact. 

The result is predictable. A demand forecast splits one customer’s orders across three records and underestimates all of them. A purchasing suggestion recommends reordering an item nobody sells any more, because the system still says it is active. 

What to fix first: pick the master data your planned AI feature will actually use. For a demand forecast, that means products, units of measure and customers. Merge duplicates, retire inactive items and agree on one unit of measure per product before you train or switch on anything. 

Problem 2: History that does not mean what it seems to 

Forecasting and anomaly detection learn from transaction history. That history is only useful if it reflects what really happened in the business. In many ERP systems, it does not. 

Common examples include sales orders that were cancelled but never closed, stock adjustments posted months late to make the numbers balance, and a period when the company ran part of its operations outside the system. There are also one-off events, such as a large contract or a supply disruption, that look like normal demand to a model. 

A person reading the reports would know to ignore these. A model will learn from them unless someone tells it not to, and its forecasts will quietly carry those distortions forward. 

What to fix first: decide how much history the AI feature needs, often two or three years for seasonal products. Review that window for open documents that should be closed, late adjustments and known one-off events. Then document what you excluded or flagged, so the next person to look at a strange forecast knows where to start. 

Problem 3: Nobody owns the data 

The first two problems have a common root. In many smaller companies, nobody is clearly responsible for the quality of ERP data. Sales enters customers, purchasing enters suppliers, and finance corrects what it can at month end. 

Without an owner, every clean-up project slowly decays. Duplicates come back because there is no rule about who can create a customer. Units of measure drift because a new product was set up in a hurry before a trade show. 

AI makes this worse, not better. Once a model is in use, poor data no longer just produces a wrong report that someone can spot; it produces confident suggestions that people start to act on. If nobody owns the inputs, nobody is accountable when those suggestions are wrong. 

What to fix first: name an owner for each master data area that the AI feature depends on. Ownership does not need a new job title, just a person with the authority to set rules, approve new records and run a short monthly check. The NIST AI Risk Management Framework is a useful reference here, because it treats clear roles and responsibilities as part of managing AI risk, not as an afterthought. 

A simple readiness check before you switch anything on 

Before enabling an AI feature in your ERP, answer these five questions honestly. If the answer to any of them is no, fix that first. 

1. Can we name the specific decision this feature will improve? “Better forecasts” is too vague; “fewer stock-outs on our top 50 products” is a decision you can measure.
2. Do we know which data the feature uses? List the tables or record types, such as products, customers, sales orders and stock movements.
3. Has that data been checked for duplicates, inactive records and inconsistent units in the last three months?
4. Have we reviewed the history window for cancelled documents, late adjustments and one-off events?
5. Is there a named person who owns each data area, and who will review the AI output for the first few months?

Run the check on one feature at a time. It is far easier to prove value with a single, well-prepared use case than to switch on every AI option the vendor offers. 

Start small, measure, then expand 

Once the data is in order, run the AI feature alongside your current process instead of replacing it straight away. For a forecast, compare its suggestions with what your planners would have ordered for a few cycles. For invoice capture, have someone spot-check a sample of the documents it processes. 

Keep the measures simple and tied to the decision you named in the readiness check. If the feature improves that number and the people using it trust the output, expand it. If it does not, the problem is often still in the data, and you will know exactly where to look. 

The real value of the exercise 

Preparing data for AI is not wasted effort, even if you later decide a feature is not worth keeping. Clean master data, trustworthy history and clear ownership make every report, every month-end close and every future system change easier. 

AI in ERP can be genuinely useful for smaller companies. But the businesses that get value from it are usually the ones that treated the data first and the AI second.