Data Quality: The Invisible Foundation of AI-Ready ERP Systems

Why the best demand forecast in wholesale fails on dirty master data — not on the AI. A status assessment drawn from 500+ projects across the DACH region.

AI × ERP series · Part 2
Part 1 showed why ERP data alone can't carry AI. Part 2 takes up the second way — data quality — using wholesale and distribution as the example: what "good data" concretely means, and why no model repairs poor master data.

The key answer in 60 seconds

No AI is better than the master data it computes on.

In wholesale, it is not the forecasting model that decides the hit rate, but the article master beneath it. Duplicates, inconsistent units of measure and missing attributes cause an AI to compute the wrong demand with full confidence — and the error scales across thousands of line items.

Data quality is therefore not an IT-hygiene topic, but the precondition of any AI-supported replenishment. Tackle it before the model and you earn the investment back; skip it and you automate your own mistakes. 33+ years of consulting experience, 500+ projects, 100% vendor-independent.

"Our processes run fine — so why does the forecast still produce nonsense?"

We hear this question especially often in wholesale. Orders, goods receipts and invoices flow cleanly through the ERP, the routines are well-drilled — and yet the first AI-supported demand forecast misses. The reflex is to tune the model. The cause almost always sits one level deeper: in the master data the model computes on in the first place.

That this is no fringe issue is shown by official statistics. According to the German Federal Statistical Office (Destatis), only 26 percent of German companies used AI in 2025. Among the companies that examined a deployment and rejected it, 44 percent cite difficulties with the availability or quality of the data — and 45 percent the incompatibility with their existing devices, software and systems. Two of the five most-cited obstacles therefore sit not in the AI, but in the data foundation and the system landscape beneath it.

What "data quality" really means for AI

"Good data" is not a feeling but something measurable. At its core it comes down to five dimensions: completeness (are all the necessary attributes filled?), consistency (do fields and systems not contradict each other?), currency (do the data reflect today's state?), accuracy (do they match reality?) and uniqueness (does each article exist exactly once?). An AI cannot guess any of these gaps — it amplifies them.

In trade, a standard for this has long been established. GS1 Germany defines binding validation rules for product master data through its Data Quality Gate and the industry standard GS1 DQX; for exchange via the GDSN in the German target market, validation has been mandatory since May 2023. More than 1,200 companies already use the service. The message behind it is unambiguous: in wholesale, master-data quality is not optional but obligatory — long before AI enters the picture.

Three master-data defects that hold back any wholesale AI

In project work it is repeatedly the same three defects that prevent dependable forecasts — and that no algorithm corrects on its own.

What stands out: none of these defects is an AI problem. They are master-data and process problems that stay invisible in day-to-day business for a long time — until a model computes them consistently and so makes them visible.

Where it tips over: when the forecast scales the error

As long as people do the planning, experienced staff quietly catch many data errors — they know that "this is really the same article" or "the pallet holds 48 units". An AI does not know that. It takes the data at their word and applies the error systematically across the entire assortment. A silent maintenance error thus becomes a loud planning problem: overstock here, shortages there, both at once.

Data quality is therefore not just a risk but a competitive factor. The German Economic Institute (IW) frames AI in 2025 as a competitive factor whose benefit depends directly on the data foundation — the edge comes not from the model, but from the data a company can give it. In wholesale this means, concretely: those who master their master data can stack forecasting, replenishment and automation on top of each other. Those who do not, automate their errors faster than the competition.

From our project work · B2B wholesale · DACH

A purchasing cooperative with 45,000 articles across ten branches — and no single version of the truth.

Starting situation: Master data — suppliers, products, customers — was created and changed independently at every branch, with no uniform standards and no central validation. Identical articles carried different names, attributes and classifications. Exactly the condition in which any AI-supported forecast or automated replenishment would have scaled the errors rather than the sales.

Approach: A master-data audit against completeness, uniqueness, currency and consistency; a central data model on the single-source-of-truth principle with mandatory attributes and validation rules for the full 45,000-item range; workflow-driven approval (decentralised entry, central validation, released with an audit trail); and a site-wide training and data-governance concept.

–73%error rate in master data

–15%quality-management effort

The point for AI: the clean data model is the precondition the planned ERP automation — machine-based planning, automated ordering, supply-chain workflows — now runs on. The forecast was never the hard part; the data underneath it was.

Read the full case study →

Recommendation: three steps to AI-ready master data

Before the first AI forecast starts, it pays to build a dependable master-data foundation. Three steps have proven themselves in practice:

• Step 1 — make quality measurable. Define per data dimension what "good enough" means, and measure the actual state on the critical assortment. Only a metric turns the gut feeling "the data are bad" into something you can steer.
• Step 2 — anchor ownership. Master data need owners in the business units, not only in IT. Who creates, who maintains, who approves? Without clear data ownership, every cleaned-up data set reverts to its old state within months.
• Step 3 — secure quality at the source. Check data on creation, not first at reporting. Mandatory fields, validation rules and an approval gate prevent errors where they arise — rather than cleaning them up expensively later.

Our assessment: these three steps are the precondition for the later parts of this series to hold at all — from cooperation in data spaces to the roadmap in Part 5. A good model on poor master data stays an expensive misunderstanding.

Next steps

When an AI-supported forecast or replenishment is on the horizon, it pays to look at the master data before the model is chosen. You can find more on our approach to digitalisation, ERP and AI integration in our overview of consulting services.

The start of the series: why ERP data alone can't carry AI — read Part 1.

Part 3 of the series: Sharing data without losing control — federated learning and data spaces (using manufacturing as the example). 

This applies most directly in wholesale & distribution, where grown master data is the usual starting point.

Dr. Harald Dreher Managing Director, Dreher Consulting · 33+ years of consulting experience in the DACH Mittelstand · 500+ ERP, digitalisation and AI projects guided · 100% vendor-independent · Personally available to leadership for an initial conversation. Book an appointment directly with Dr. Dreher →