Every AI vendor selling into the supply chain market will tell you the same story: deploy our model, feed it your ERP data, and watch demand forecasting accuracy climb. What they rarely tell you is what happens when your procurement team runs SAP, your acquired subsidiary runs Oracle NetSuite, your logistics arm runs Microsoft Dynamics 365, and none of them agree on what a "product" or a "supplier" actually is.

This is the data harmonisation problem. And in 2026, as enterprises accelerate their AI investments across supply chain operations, it is quietly becoming the most expensive bottleneck that nobody in the room wants to own.

The Multi-ERP Reality Most Enterprises Are Living In

The first thing to understand is how common this situation actually is. Enterprise growth - through acquisition, regional expansion, or decades of tactical IT decisions - rarely produces a clean, single-ERP landscape. Persistent cost pressure, geopolitical instability, labour constraints, and demand volatility have become part of the operating baseline, and many organisations now recognise that traditional planning models, organisational silos, and technology stacks are no longer aligned with how supply chains actually need to perform.

The ERP market reflects this fragmentation directly. SAP holds a 6.57% revenue share of the $131B global ERP market, Oracle sits just above at 6.63%, and Microsoft Dynamics accounts for approximately 4%. But these are not mutually exclusive installations. Large enterprises routinely run more than one. A manufacturer might standardise on SAP S/4HANA for core financials and production planning while an acquired business continues on Oracle Fusion. A retail group may have Dynamics 365 at its head office and NetSuite embedded across its e-commerce subsidiaries.

41% of organisations already report active integration challenges between ERP and WMS systems - and that figure does not account for the hidden friction between ERP instances themselves, which tends to surface only when someone tries to build something on top of them, like an AI model.

Why AI Makes This Problem Impossible to Ignore

For years, multi-ERP fragmentation was manageable. Analysts pulled exports, reconciled spreadsheets, and produced reports that were slightly stale but directionally useful. Supply chain teams lived with the imprecision because the cost of fixing the data architecture was harder to justify than the cost of tolerating it.

AI changes that calculus entirely.

The challenge has intensified as leaders attempt to deploy advanced analytics and AI tools that require a foundation of high-fidelity data. Without addressing the underlying integration hurdles, these technologies cannot provide a measurable return.

This is not a minor caveat. Machine learning models for demand forecasting, inventory optimisation, and supplier risk scoring are only as reliable as the data they are trained on. When your training data contains the same supplier recorded under three different IDs across two ERP systems, with inconsistent lead times and mismatched unit-of-measure conventions, the model does not average out the errors. It learns it.

When different business units chase their own definitions of "customer" or "product", analytics become unreliable, workflows slow down, and costly manual reconciliation becomes the norm. In an AI context, this is not just operational inefficiency. It is model poisoning at the source.

The Four Layers of the Harmonisation Problem

Organisations trying to solve this often discover that what looks like a technical data integration challenge is actually four overlapping problems stacked on top of each other.

1. Structural Inconsistency

Different ERP systems use fundamentally different data structures. SAP's material master, Oracle's item master, and Dynamics 365's product catalogue each store nominally the same information - a product - in different schemas, with different field names, different validation rules, and different levels of granularity. Mapping them is not simply a matter of matching columns. It requires interpretive decisions that, if made incorrectly, produce silently wrong data rather than visibly broken data.

2. Semantic Divergence

Structural mismatches are at least visible. Semantic divergence is more insidious. Two ERP systems may both have a field called "lead time", while meaning entirely different things by it - one capturing supplier lead time, the other capturing the full order-to-receipt cycle. Both values may be populated, both may look reasonable, and the discrepancy may never surface until an AI model produces inventory recommendations that consistently miss for one product category and not another.

Data governance is becoming the new integration battleground. As AI works best with unified operational datasets, ERP providers and system integrators must prioritise data quality frameworks, semantic consistency, and lineage visibility.

3. Master Data Proliferation

The 2025 DATAVERSITY Trends in Data Management Survey found that 61% of organisations list data quality as a top challenge, yet only 15% report having mature data governance in place. In multi-ERP environments, the master data problem compounds with every system boundary. A supplier operating across five countries may appear under five different IDs, with five different payment terms, five different risk ratings, and no single record that can be trusted as the authoritative source.

Multiple records for the same supplier split spend across IDs, ERPs, and regions. Contract pricing, tiered discounts, and rebate thresholds fail to trigger. Sourcing teams lose a clear view of total supplier leverage. The problem is not just analytical - it has direct commercial consequences.

4. Governance Without an Owner

Even organisations that understand the problem struggle to fix it because data harmonisation sits in the space between IT, operations, procurement, and finance - and each function has legitimate reasons to defer ownership to another. Establishing who owns what data and which processes is critical; for manufacturing organisations, ownership may lie outside the data team with a business user whose work brings them into consistent, direct contact with the data. Without that governance clarity, harmonisation initiatives stall in committee.

What AI Is Now Being Asked to Do About It

The market is converging on three architectural approaches to address data harmonisation in multi-ERP supply chain environments, each with meaningfully different implications for AI readiness.

The Semantic Layer

A semantic layer connects analytic semantics - metrics, formulas, access controls - to enterprise-wide conceptual knowledge, allowing integration across silos such as ERP, CRM, and warehouse systems while enriching data with reasoning. AI agents or copilots can then query data with awareness of both the definition and the governance context simultaneously.

For multi-ERP environments, this allows a consistent business vocabulary to be imposed above the system layer without requiring the underlying ERP systems to be modified or migrated. The semantic layer acts as a translation engine - but one that can be reasoned over, not just queried.

Knowledge Graphs

An enterprise knowledge graph models entities - customers, products, contracts, events - the relationships between them, and the definitions that give those relationships meaning. Unlike relational data models that flatten information into tables and foreign keys, a knowledge graph preserves context: how entities connect, why definitions exist, and how state changes over time.

For supply chains specifically, knowledge graphs are particularly effective for multi-tier supplier networks with dependencies and logistics constraints. They allow an AI system to understand not just that two suppliers exist, but that one is a tier-2 sub-supplier of the other, that both share a logistics dependency, and that a disruption at one cascades to the other in a quantifiable way - a context that no flat data extract can convey. The knowledge graph acts as a shared memory and coordination hub: a digital nerve centre connecting specialised agents across departments and data systems.

AI-Driven Master Data Management

Rather than retrofitting AI capabilities to traditional MDM, new solutions are being built with machine learning, anomaly detection, and smart entity resolution at their core. This embedded approach delivers continuous learning, automated matching, and adaptive rules without heavy manual intervention.

In practice, this means AI systems that can identify "Siemens AG - DE" and "SIEMENS GMBH (Germany)" across two different ERP systems as the same entity, merge their records into a golden record, and propagate that resolution downstream - without a data steward reviewing every match individually.

The Agentic AI Complication

There is a further layer of urgency only beginning to register with supply chain and IT leaders: agentic AI.

Agentic systems already accounted for 17% of total AI value in the supply chain in 2025 and are projected to reach 29% by 2028. These agents reason through complex logic chains, query disparate systems - ERP, WMS, TMS - and trigger actions without constant human oversight.

When an AI agent is authorised to take operational actions - raising a purchase order, adjusting a safety stock level, flagging a supplier for review - the quality of the data it is acting on is no longer just an analytical problem. It is an operational risk. A harmonisation failure that would previously have produced a misleading report now produces a wrong decision, executed automatically, at speed.

Effective agentic deployment requires three foundations: a unified data layer connecting ERP, PLM, and market intelligence so agents act on a single source of truth; a hybrid workforce combining human expertise with digital agents; and governance that moves AI from pilot purgatory to production. The data layer comes first. Without it, agentic AI in multi-ERP supply chains is not a capability - it is a liability.

Where Organisations Are Getting Stuck

Despite strong awareness of the problem, most organisations invest heavily in data management yet remain dissatisfied with the outcomes. The primary barrier is the reliance on legacy systems that were never designed for modern interoperability. The common failure modes are well-documented:

The Big Bang Migration Trap

Organisations commit to a single-ERP consolidation, spend 18-36 months in programme planning, and find that by the time the programme is funded and scoped, an acquisition has occurred, or the target ERP has released a version that changes the migration path entirely. AI initiatives sit on hold throughout.

The Point Integration Spiral

Organisations build bespoke integrations between ERP pairs - SAP-to-Oracle, Oracle-to-Dynamics - and end up with a web of brittle connections harder to govern than the siloed systems it replaced. Each integration is understood by one person, documented poorly, and fails in unpredictable ways when either system is upgraded.

The Data Lake Deferral

Organisations ingest all ERP data into a central lake or warehouse, declare harmonisation solved, and discover that unresolved semantic inconsistencies do not disappear when data is co-located - they become embedded in every downstream model and report.

Clean data does not guarantee correct execution. Data that is technically valid may still fail under real-world constraints, producing unexpected downstream failures and partial results that are harder to diagnose than clean errors.

A More Practical Path Forward

The organisations making progress are typically doing three things differently.

They separate harmonisation from migration

Rather than waiting for a single-ERP future that may never arrive, they build a harmonisation layer above existing systems - defining canonical data definitions, entity resolution rules, and governance policies independently of what any individual ERP does. This produces AI-ready data now, without requiring system consolidation first.

They govern semantics, not just schemas

Technical data mapping gets the most attention, but semantic governance - agreeing on what terms mean and who owns those definitions - is where most value is lost or created. Leading organisations establish data product ownership at the domain level, so that the "supplier" definition used by procurement, finance, and supply chain is explicitly agreed, version-controlled, and applied consistently across every system that touches it.

They are instruments for AI, not just for reporting

Data quality requirements for a monthly management report are materially lower than those for an AI model updated daily and acting on its outputs. Organisations that treat AI readiness as a distinct data quality standard - separate from existing reporting standards - reach production AI deployments faster with fewer surprises at model validation.

The Bottom Line

The conversation in supply chain AI right now is dominated by model capabilities - which vendor's demand forecasting is most accurate, which platform has the best anomaly detection, which agentic framework can handle the most complex exception management. These are legitimate questions. But they are secondary questions.

The primary question - the one that determines whether any of those capabilities can actually be deployed at production scale - is whether the data those models will be trained and run on is harmonised, governed, and trustworthy across every ERP system in the estate.

Manufacturers face four consistent AI adoption challenges: fragmented data landscapes, limited in-house expertise, legacy system constraints, and a lack of measurable business outcomes. The first is the multiplier on all the others. Fix it, and the rest becomes tractable. Leave it unresolved, and even the best AI investment will underperform relative to its potential.

The data harmonisation problem in multi-ERP supply chains is not glamorous. It does not make for compelling conference keynotes. But it is, right now, the most important infrastructure decision a supply chain and IT leader can make - and the organisations that treat it as such will be the ones whose AI investments actually compound.

VE3 Global helps enterprises design and implement data harmonisation architectures that unlock AI readiness across complex, multi-ERP supply chain environments.