Agents are being pointed at the systems companies were too frightened to replace.

The most common enterprise AI project of the past year is not a chatbot. It is an agent placed in front of a legacy system to do what a trained clerk used to do, because replacing the system would cost more and take longer than the board will tolerate.

In narrow terms, this works. Order entry, reconciliation, claims coding and master data cleanup all yield to a well scoped agent with access to the same screens a human used.

What it defers

Every wrapper extends the life of the system underneath. That is the point, and it is also the problem, because the institutional knowledge of how that system behaves continues to retire.

Five years from now the organisation has an ancient core, an agent layer written against its quirks, and nobody who understands either. The migration did not get cheaper. It got harder and less visible.

The governance question nobody asks

When an agent makes a posting in a financial system, which control framework covers it. In most deployments the answer is the one written for the human process, which assumed a person could be asked why.

A defensible way to do it

Treat the wrapper as explicitly temporary, with a documented expiry and a migration budget attached at approval. Log every agent action against the same audit standard as a human user. And keep the interface boundary narrow enough that the eventual replacement has a clean seam to cut along.