Every major ERP vendor — SAP, Oracle, Microsoft, Infor — has announced AI features in the last 18 months. Most of them share a common architecture: a large language model sitting on top of an existing data model, accessed through a chat interface or a new "AI assistant" panel. It looks impressive in demos. It underdelivers in production.

Here's why, and what a genuinely AI-native ERP modernisation looks like instead.

The Problem with AI-Bolted-On

Legacy ERP systems were designed around a fundamental assumption: humans are the primary consumers of data. The data models, the workflows, the reporting structures — all of it was built for human comprehension and human decision-making. Adding an AI layer on top of this doesn't change the underlying architecture. It just adds a natural language interface to the same rigid, human-centric data model.

The result is AI that can answer questions about your ERP data but can't fundamentally change how that data flows, how decisions get made, or how the system responds to changing conditions. You get a better search interface, not a smarter system.

Three Specific Failure Modes

  • Stale data: Most ERP systems batch-process data overnight. An AI layer on top of this is making decisions based on yesterday's reality. For inventory management, demand forecasting, or supply chain optimisation, this latency is often unacceptable.
  • Siloed context: ERP data doesn't exist in isolation. Procurement decisions depend on supplier data, market conditions, and logistics constraints that live outside the ERP. A bolted-on AI can only reason about what's in the ERP.
  • Rigid workflows: Legacy ERP workflows were designed for human approval chains. AI can't meaningfully accelerate a process that requires seven human sign-offs at fixed stages.

What AI-Native ERP Modernisation Looks Like

AI-native doesn't mean replacing your ERP. It means rebuilding the data and decision layer around it so that AI can actually operate on real-time, contextually rich data and take meaningful action.

Step 1: Build a Real-Time Data Foundation

The first step is extracting ERP data into a real-time streaming architecture. We use change data capture (CDC) to stream ERP transactions into a lakehouse as they happen. This gives AI systems access to current data, not yesterday's batch.

Step 2: Enrich with External Context

ERP data becomes dramatically more useful when enriched with external signals — supplier risk scores, commodity price feeds, weather data for logistics, demand signals from e-commerce platforms. We build data pipelines that join ERP data with these external sources in real time.

Step 3: Replace Approval Chains with Autonomous Agents

The highest-value AI application in ERP modernisation is replacing rigid approval workflows with autonomous agents that can make routine decisions within defined parameters and escalate exceptions to humans. A procurement agent that can autonomously approve purchase orders under £10,000 from approved suppliers, flag anomalies, and escalate edge cases — this is where the real productivity gains come from.

The Business Case

The enterprises we've worked with on AI-native ERP modernisation have seen procurement cycle times drop by 40–60%, inventory carrying costs reduce by 15–25%, and finance close cycles compress from weeks to days. These aren't AI demo numbers — they're production outcomes from systems that have been running for months.

If you're planning an ERP modernisation and want to understand what an AI-native approach would look like for your specific stack, we're happy to have that conversation.