AI won’t kill the ERP. It changes what the ERP is for. And that changes everything you need to know about modernising it.
ERP modernisation programs are among the largest operational investments an organisation will make. They're also among the most predictable in how they go wrong: customisation grows, integration complexity compounds, and the cost of future change quietly escalates until the next modernisation feels as urgent as the last one.
AI doesn't make this problem smaller. It makes it more consequential because AI fundamentally shifts the role an ERP plays in the enterprise. More work will be initiated through conversational, agent-driven workflows that span multiple systems. An AI agent handling a procurement request won't navigate five ERP screens to submit a purchase order. It will orchestrate across supplier data, contract terms, budget approvals, and compliance checks, then post the result to ERP once the transaction needs to be recorded and controlled.
The ERP isn't disappearing. But its role is shifting from where work happens to the anchor for transactional integrity, control enforcement, and exception handling.
That changes what modernisation programs should optimise for. Not perfecting how people perform tasks in one system, but enabling faster decisions, safe automation, and resilient controls across the enterprise.
If you are undertaking ERP modernisation right now, three deliberate choices will determine whether you get this right.
1. Where should customisations be built
ERP customisation was justified when ERP was where work happened. That logic weakens considerably when routine work increasingly starts outside it.
In most programs, custom requirements cluster into three categories:
-
Hard controls: statutory compliance, posting integrity, audit evidence, segregation of duties. These are non-negotiable.
-
Measurable differentiation: changes that improve customer outcomes or unit economics with a defined metric and payback. These need to prove their value.
-
Workarounds: compensating for broken processes, unclear decision rights, or poor data quality. These should be resolved at root cause, not built into the system.
Based on what we see, up to 50% of customisation requests fall into the workaround category when teams are honest about it.
The implication is clear: keep the ERP core clean and upgradeable. Reserve ERP changes for genuine control requirements. Deliver usability and workflow guidance above the ERP, where change is faster and cheaper.
2. What must be standardised now before you scale AI
Standardisation in an AI-enabled enterprise isn't about uniformity. It's what makes automation predictable, auditable, and safe.
The critical point most organisations miss: agents will amplify your definitions, your controls, and your data quality in whatever state they exist today. A human working in a flawed process might catch the inconsistency and escalate. An agent will faithfully replicate every bad definition and every ambiguous threshold across thousands of transactions before anyone notices.
Four foundations determine whether automation scales smoothly or scales risk.
-
Data and definitions When two divisions define "revenue" slightly differently, humans muddle through. An agent optimising against inconsistent definitions produces confident wrong answers at scale.
-
Control design Most control frameworks were designed with a human in the loop at key decision points. AI-initiated workflows compress or eliminate those loops. If your controls don't account for agent-initiated transactions, you'll discover the gaps in an audit finding rather than in your design phase.
-
Integration reliability Every integration point is a potential failure point. When something fails in an AI-orchestrated workflow, you need to detect it, trace what happened, and recover without manual reconstruction. If that's not true for your integrations today, scaling AI will expose it.
-
Identity and access In principle, AI agents should operate within the same authority and boundaries as the human roles they support. In practice, defining who instructs agents, what authority they operate under, and how their actions are logged against human accountability is still emerging. Getting ahead of this during design is significantly cheaper than retrofitting it after your first agent-related incident.
3. When do you use your ERP vendor's AI, and when not to?
Your ERP vendor has an AI roadmap. It's probably impressive but it may not be the right answer for every workflow you have.
Vendor AI makes sense when the workflow is ERP-native, the data and control points live inside the suite, and speed matters. It works well on home territory.
An independent AI layer makes sense when the end-to-end journey spans multiple systems, when you need a unified experience across silos, or when long-term optionality is critical — multi-ERP environments, frequent M&A activity.
The mistake we're seeing more often is defaulting to vendor AI because it's the path of least resistance, then discovering the most valuable AI opportunities sit beyond the vendor's platform. By then, switching costs are exponentially higher.
The bottom line
AI doesn't eliminate the need for ERP modernisation but it does raise the cost of getting it wrong.
-
Lock in heavy customisation in the core, and you'll make it harder to adopt AI-driven workflows above it.
-
Skip the standardisation foundations, and you'll scale risk instead of value.
-
Default to vendor AI without evaluating fit by workflow, and you'll discover the constraints when it's too expensive to change course.
Get these wrong and you will find yourselves modernising again in five years' time, for the same reasons, at greater cost.