Make legacy code AI-queryable: Vanguard's 7-day research loop became 15 minutes.

The numbers. Gareth Yoder, technical lead for Vanguard's cash program, describes a modernization effort that handles 30 million calls and half a million ledger entries per day. The turning point came when the team exposed its legacy mainframe codebase to AI tooling through an MCP server. Research that used to require a mainframe engineer and seven days of analysis now returns in about 15 minutes. Today 90 percent of the program's engineers use the AI tools daily.

The mechanism. The AI tooling did not start as a code generator. It worked as a queryable bridge into legacy code, letting developers, product owners, and architects ask questions directly instead of waiting on mainframe experts. The same MCP server exposed migration documentation from the company wiki. AI also translated opaque legacy status codes into human-readable cloud enums, which made the intent of the old system visible during the move. The faster loop changed the pace of the entire program.

The architecture. Three pillars ran in parallel: data, producers, and consumers. Data used change data capture with daily row and column audits. Producers moved through a gateway wrapping cloud APIs, with idempotency modeled on Snowflake-style IDs. Consumers used a schema registry across channels: GraphQL for real-time funds available, data-share views for batch, and Kafka for streaming with at least once delivery and idempotent retries. A coherent pattern, but kept at overview level.

The governance layer. More interesting than the speedup is Vanguard's approach to onboarding. With 35 producers and 50 consumers, the team required a three-character application code plus a written business-case justification for data access. AI compiled those descriptions into a use-case catalog and made SLA impact visible: a failing batch job allows hours to respond, while a real-time API needs minutes. Treating governance text as AI-readable data is a genuinely transferable idea.

The tradeoffs. The talk is a leadership-stage summary, so it skips failure modes, security boundaries, and cost. The "fix forward" rule, while sensible, throws off audit logs and requires whitelisting, which the speaker only mentions in passing. The 15-minute claim measures research turnaround, not total modernization time. One production case, no benchmarks. Still, the practical direction is clear.

The takeaway. Engineering leaders should copy three things: make legacy code AI-queryable, treat migration docs as a product, and use business cases to govern data access. The organizational work, onboarding dozens of teams and keeping observability honest, is what makes the AI speedup useful. The video is a good briefing, not a playbook.