Vanguard says AI access to old code cut a seven-day research loop to 15 minutes

Gareth Yoder, who leads the cash program at Vanguard, said the turning point in modernizing the firm's cash ledger was letting AI tools answer questions about old code. He spoke on the AI Engineer show. By his account, Vanguard linked its legacy code to AI tools through an MCP server, a connector that lets an AI tool read a company's own systems. A research loop that once ran through a mainframe engineer for a week now takes about 15 minutes.

What was said

Yoder said the ledger's services can receive up to 30 million calls a day. The team ran three migrations at once: data, producers and consumers.

Early on, no AI tooling existed, so the team worked with mainframe engineers and architects to shape the new data schema. Yoder called the arrival of code access a turning point.

Before, an engineer had to dig into the code, analyze it and report back. Afterward, one prompt to an AI tool could cover that. Yoder said "the feedback loop was reduced from seven days down to 15 minutes." He said product owners, technical leads and developers all gained. His first of five lessons was to make the legacy codebase AI queryable.

Yoder also claimed that people can "tap into this common knowledge base, and now they have the same context in their prompt as an engineer that has been here for the last five years."

Why it matters

Our reading: the bottleneck was knowledge of what the old system did, not writing new code. That knowledge sat with mainframe engineers, so each question waited on one of them.

Yoder described a separate puzzle from the producer migration: data appeared in tables, and no one knew who wrote it. The excerpts do not show whether the AI tool helped solve it.

For managers, the question is where modernization waits on a single person. If product owners get answers directly, they must also decide how far to trust them. The skill may shift from knowing the old system to judging what is said about it.

The other side

These figures are Vanguard's own account, and the excerpts offer no outside check. The excerpts have a gap in the middle. In them, Yoder did not say how accurate the AI answers were or how the team validated them. They also do not say whether the 15-minute answers matched the depth of the seven-day analysis.

His claim that users get a five-year engineer's context is also untested here. The excerpts do not show how that context was validated, or what it meant for mainframe engineers' roles. What the "common knowledge base" contains is not clear from them either.

The controls he described sit elsewhere. Daily audits compared legacy and new databases during data migration. He called observability "non-negotiable" for knowing who produces and consumes data. Neither covers checking AI answers about old code.