Your ERP is 20 years old and AI needs to talk to it
Legacy integration patterns explained for the people who approve budgets — without rewriting the core or becoming hostage to a proprietary connector.
*Fourth article in our series on enterprise AI, legacy integration and information governance.*
Executive summary
The bottleneck in enterprise AI is almost never the model — it is access to data locked inside systems nobody wants to touch. The good news: integrating AI with a two-decade-old ERP does not require rewriting the core. It requires choosing the right integration pattern for each type of data, isolating the legacy behind a translation layer, and keeping original permissions intact. This article turns four technical patterns into decision criteria for whoever approves the budget.
Why legacy is the real project
An assistant that does not know your inventory, customer history or commercial policy answers well and helps little. All the value is in context — and context lives in the ERP, the CRM, the PLM, the quality system and network folders.
Those systems usually share three complicating traits: data models only the internal team understands, rigid maintenance windows, and limited capacity to absorb extra query load. Ignoring that produces the classic pilot that shines in a demo and takes the ERP down on the first business day in production.
Four patterns and when to use each
| Pattern | When to use | Advantage | Watch out |
|---|---|---|---|
| API gateway over legacy | Point queries needing live data | No data copy; permissions at source | Direct load on the legacy system |
| CDC (change data capture) | High volume, frequent reads | Spares the core; natural history | Data duplication needs governance |
| Event queue | Asynchronous, cross-department flows | Decouples pace, absorbs spikes | Higher operational complexity |
| Anti-corruption layer | Legacy data model is unintelligible | Translates into business vocabulary | Requires modeling investment |
In practice, mature projects combine patterns: CDC to feed the search index, a gateway for sensitive queries that need current-second data, and a queue to trigger processes.
The anti-corruption layer is the most strategic decision
The term is technical, the idea is commercial: instead of teaching AI the quirks of the old system, you build an intermediate layer that exposes clean business concepts — customer, order, item, policy — and translates to the legacy behind it.
The executive benefit is optionality. With that layer, replacing the ERP later does not invalidate everything built on top. Without it, each new AI application ties the company more tightly to the legacy vendor.
Permissions: the project's most expensive mistake
When giving AI corporate context, it is easy to create a technical user with broad access — and thereby let any employee obtain, through a generated answer, information they would never see in the source system.
The rule is simple and non-negotiable: AI answers with the access level of the person asking, not its own. That means propagating identity end to end and filtering results by permission before generation — a theme explored in *RAG is not magic*.
Cost: where the budget actually goes
A typical distribution for a first integration project:
- 1.Modeling and mapping legacy data — the largest and most underestimated slice.
- 2.Building the translation layer and connectors.
- 3.Security, identity propagation and audit.
- 4.Model and inference — usually the smallest part of the bill.
If a vendor proposal concentrates cost in item 4, the work in items 1 through 3 has probably not been confronted yet.
What to do on Monday
- Request an inventory of the three systems holding your most valuable data and the real state of their interfaces.
- Require the first use case to cross an actual legacy system — pilots built only on PDFs hide the problem.
- Define the permission inheritance rule before the first line of code.
- Budget explicitly for data modeling, not only for AI licensing.
Conclusion
Legacy integration is not technical debt to tolerate: it is the foundation of the return. Companies that invest in the translation layer buy more than the current use case — they buy the freedom to change tool, model and even ERP without starting over.
Further reading
Engineering track:
- An anti-corruption layer to connect AI to a legacy ERP — the technical deep dive on this topic.