Legacy Systems vs AI: A Recipe for Failure

You've got a great AI idea. The team is motivated. The budget is approved. And then you discover that your core system has an API from 2008 and returns data in a format nobody remembers anymore.

Gartner predicts that more than 40% of agentic AI projects will fail by the end of 2027. Why? Costs keep climbing, and the new systems don't fit what the company already runs. Most people treat this as a technical problem. It's really a strategic one.

Why does legacy kill AI?

AI needs three things: data, speed and flexibility. Legacy systems offer the opposite of each.

Sometimes the whole legacy estate lives in spreadsheets. A medical-equipment distributor got tenders from dozens of sources, in .xls and .ods, and every source named the same columns differently. The pipeline that runs there today shows the model only the headers and one row of data. Plain code does the rest. When the model is unsure, the cell gets a colour and a comment, so a person knows what to check. I wrote up the results in the medical-equipment distributor case study.

3 integration strategies

1. Abstraction layer (Data Mesh)

Instead of touching the legacy, build an intermediate layer. An API Gateway that translates old formats into new ones. A Data Lake that pulls data from different sources together. Zhamak Dehghani called this Data Mesh: every dataset has a real owner, and you treat it like a product.

Pros: You don't touch the legacy. Cons: One more layer to maintain, plus cost.

2. Strangler Fig Pattern

Martin Fowler described this pattern: piece by piece, you replace legacy functionality with new microservices. AI runs on the new services while the old ones slowly die off. The name comes from the strangler fig, which grows over its host and replaces it.

Pros: Lower risk. Cons: Slow. Years, not months.

3. RPA + AI Bridge

You use RPA to "scrape" data from the legacy UI, AI processes it, and RPA writes the results back. It sounds ugly, but it works.

Pros: Fast. Cons: Brittle, needs maintenance.

"The wrong question is 'how do I integrate AI with the whole legacy system'. The better one: 'what's the minimum data access I need to get AI working'. Usually you need far less than you assume."

Audit before you start

The checklist below closes the data question from the list of five questions to answer before a rollout.

Before you start an AI project, answer these questions:

Case study: A client with a 15-year-old ERP

A manufacturing company wanted to deploy AI for demand forecasting. The problem: their ERP had no API, only a CSV export once a day.

The solution:

• • A script that automatically pulls the CSV at 6 a.m.
• • ETL into a Data Lake (PostgreSQL)
• • An AI model trained on yesterday's data
• • Predictions emailed to the planners

It's not a textbook architecture. But it works and it delivers value.

Eight months, up to −23% in costs. You'll find the full arc of a deployment like this in the logistics case study.