Beyond the glossy roadmap: Bridging the gap between agents and assets

A few months ago, I was reviewing an agentic AI roadmap with the CIO of a Fortune 500 insurer. He pulled up two slides side by side. On the left: A glossy roadmap for a new agentic AI platform — multi-agent orchestration, vector databases, the works. On the right: A 30-year-old loan origination system running on a mainframe his team had been “planning to retire” for the last few years.

“We’re spending $40 million on the left,” he said. “And the agent on the left can’t see anything on the right.”

That, in one sentence, is the architectural lie most enterprises are living with right now.

For the better part of a decade, we’ve split our IT strategy into two budgets: An innovation lab, where shiny things get built, and a maintenance core, where the actual business runs. We promote people for the first column and quietly outsource the second. Then we wonder why our generative AI pilots can’t make it to production.

I’ll say what most CIOs already suspect but won’t put on a slide: Your modernization strategy and your AI strategy are the same strategy. If you’re funding them as separate line items in 2026, you’re already behind.

What I got wrong about “lift and shift”

I’ll admit I was a believer for a long time. In 2022, my team finished a textbook migration off a 20-year-old Oracle Exadata footprint onto a Cloud native solution. Eighteen months, $20M, zero data loss and a 40% reduction in run-rate infrastructure cost. I had the slide ready for the board.

Then our first generative AI pilot landed, and I watched our agent fail to answer questions that the old database — with all its ugly stored procedures — could have answered in a single query. The “modernized” Spanner instance was clean, fast and — without the surrounding semantic layer — couldn’t answer questions the old database could.

The lesson I took away: Legacy systems aren’t technical debt. They’re encoded institutional memory — three decades of edge cases, compliance carve-outs, regional rules and “we tried that in 2007, and it broke everything” wisdom that nobody documented because nobody had to. You can rewrite the code. You can’t rewrite the knowledge.

Which is precisely why agentic AI changes the math. For the first time, we have a technology that doesn’t just tolerate that messy old context — it needs it.

Earlier this year, I spoke on this intersection of agentic cloud strategy and legacy modernization at Google Cloud Next 2026. What follows is the framework I shared for moving beyond the “lift and shift” mentality.

MCP isn’t a protocol. It’s a peace treaty

The Model Context Protocol gets discussed as a technical standard, and that framing undersells it. After watching teams burn quarters writing custom connectors between every model and every data source, I think MCP is closer to a peace treaty between two warring tribes inside the enterprise: The cloud-native engineers and the legacy custodians.

When an agent can pull a customer’s 2003 account history from a mainframe and combine it with a real-time fraud signal without somebody hand-coding the bridge, two things happen. The integration backlog stops growing. And — more importantly — the political wall between the two teams starts to crumble.

I’d argue this is the most under-appreciated CIO opportunity of the next 18 months. Not the agents themselves. The org chart you can finally redraw because the agents made the old separation pointless.

Why I stopped trusting RAG for legacy code

Here’s a take I’ve gotten pushback on: Standard retrieval-augmented generation is the wrong tool for legacy modernization, and we’re using it anyway because it’s familiar.

RAG is brilliant for documents. It’s a disaster for code. A monolithic codebase isn’t a haystack with a needle in it; it’s a tangled fishing net where every knot is connected to seven others. Ask a vanilla RAG system to help you refactor a billing module and it will confidently tell you the change is safe, right up until production goes down on a Tuesday morning.

Andrew Moore, the former head of Google Cloud AI and now co-founder of Lovelace, told CIO recently that “you cannot do safety-critical reasoning for agents purely based on trying to do the same sort of thing you normally do for chatbots.” That maps to my experience exactly. Chatbots can be wrong and embarrassing. Agents can be wrong and expensive.

GraphRAG — building a semantic knowledge graph of how the code actually behaves, not just how it reads — is the only approach I’ve seen that respects the structural reality of these systems. In 2025, I was advising on a mainframe modernization at a health insurer.  The CIO had given me the official application inventory: 47 systems, well-mapped, with named owners. We ran a graph analysis across the actual code and graph surfaced 53 systems.

The six extra weren’t ghosts. They were live, running and integrated — small applications written between 1998 and 2004, each owned by a business unit that had quietly built them, never registered them and treated them as “spreadsheets that happened to run on the mainframe.” One of them was processing about $35M a year in vendor rebates against pricing tables that no one in IT knew existed.

The graph didn’t just find undocumented dependencies. It found undocumented systems. You cannot modernize what you cannot see, and most enterprise application inventories are wrong by an amount that should embarrass us.

The hard part isn’t technical

Everything I’ve described is buildable today. The vendors are there. The protocols exist. The talent, if you know where to look, is hungry.

What’s missing is architectural courage at the CIO level — and I mean that specifically. It is much easier to fund a new AI initiative than to defend a modernization budget. The first gets a press release. The second gets a line item nobody reads. But if you keep funding them separately, you’ll end up with the same outcome I’ve watched at company after company: A sleek agentic layer floating above an integration layer that bleeds margin every quarter, while your competitors who did the unglamorous work compound advantage you can no longer catch.

If I could give CIOs reading this one piece of advice — and I realize this is the kind of thing that’s easy to say and hard to do — it’s this:

Take your top three modernization projects and your top three AI projects. Put them on the same slide. Give them the same executive sponsor. If you can’t, you don’t have a strategy. You have two parallel cost centers waving at each other.

The myth that legacy and innovation are opposites was always wrong. In the agentic era, it’s an unforced error. The question for the next 18 months isn’t whether to modernize. It’s whether you have the architectural clarity — and the political cover — to do it as one program instead of two.

I think most CIOs already know which side of that line they’re on. I’d just rather we stop pretending otherwise.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?