AI
AI Is Changing Who Does What, Not Just How They Do It
Avanade has 700,000 people. Aaron Reich, its Chief Technology and Innovation Officer, expected developers to be the heaviest Copilot users. It turned out to be QA. That single data point captures the pattern across the firm’s global client base: AI is not just speeding up the same roles, it is shifting which roles carry the work.
“I will tell you, no one has figured out what that balance is.” — Aaron Reich
Reich calls the goal collaborative intelligence, matching human judgment to machine execution at the right points. The job description for a QA hire today still matches the one written three years ago. In 12 months, it will have to change.
The Legacy App Problem That Blocked Modernization
Legacy financial services applications often outlive the engineers who built them. Documentation is thin or absent. The codebase has run for 20 years and no one wants to touch it, because the cost of a mistake on a business-critical system is high and the cost of understanding it first is higher.
Joe Workman ran Avanade’s legacy app modernization practice before moving to GitHub services. His diagnosis: before writing a line of new code, a team spends the majority of its time figuring out what the existing system does. That is the problem he built tooling to solve.
Automating Discovery with a GitHub Knowledge Agent
Workman built a Node.js CLI called the GitHub Knowledge Agent. It takes a repo owner and repository name, picks a prompt template, creates a GitHub issue pre-loaded with that template’s criteria and deliverables, then assigns the issue to Copilot Coding Agent via the GitHub API. He walked through the GitHub Knowledge Agent CLI demo (12:27) against a financial services event-driven sample repo.
The agent ran 25 minutes and returned Mermaid ER diagrams, a capability matrix, an event catalog, and business rules formatted as feature requirements. Discovery used to consume 50% or more of project time. That time now goes to building.
From 3 Million Lines of COBOL to a Spec-Driven Lifecycle
An insurance client brought Avanade 3 million lines of COBOL managing its policy system. Cloud migration and faster feature delivery were the goals. A direct translation would produce the same undocumented mess in a new syntax.
“Our job as GSI is really to de-risk those those projects for our clients.” — Eric Bouguen
COMPASS maps files to business domains across COBOL, PL/I, Java, and C++ in weeks. It has handled up to 8 million lines for a single client. Eric walked through the COMPASS Woodgrove Bank modernization demo (21:29) showing domain extraction and per-rule docs. A Copilot agent called Campus Advisor answers business-rule questions on demand. Developers review and edit the generated C# specifications before any new code runs.
Notable Quotes
because we have this view across our business where we are 700,000 people Aaron Reich · ▶ 0:33
I will tell you, no one has figured out what that balance is. Aaron Reich · ▶ 5:16
Our system is able to ingest the whole million lines of code. Eric Bouguen · ▶ 22:48
Our job as GSI is really to de-risk those those projects for our clients. Eric Bouguen · ▶ 18:28
Key Takeaways
- Avanade’s QA team, not developers, became the top Copilot users, signaling which roles AI changes first.
- A 25-minute Copilot Coding Agent run can replace days of manual legacy codebase discovery work.
- COMPASS plus a spec-driven workflow turns undocumented COBOL into a maintainable lifecycle for decades ahead.