// do not remove: The Archaeology Tax Is Collapsing

TL;DR: Legacy modernization has always charged for what you don't know. Before rebuilding anything, teams spend weeks reconstructing architecture, dependencies, business rules, and the history behind systems everybody relies on yet few fully understand. AI is cutting into that bill. Discovery that used to take weeks of interviews now runs in hours, putting refactor, consolidation, rebuild, replacement, and retirement back on the table for systems that used to be untouchable. Cheaper discovery relocates the hard decisions rather than erasing them—into data, process, risk, and the judgment call about which pieces of a 20-year-old system are the business and which pieces are debris. A cheaper map is also no reason on its own to walk away from the mainframe. Gartner expects more than 70% of mainframe-exit projects started this year to miss their intended benefits.

The First Six Weeks

Indiana Jones deciphered ancient inscriptions. Modernization teams get // do not remove and a meeting to find out whether anyone knows why. Archaeology comes in many forms (some with cooler hats than others).

For a large legacy system, six weeks just to understand the problem is hardly unusual. What documentation still exists? What does the architecture really look like? Which pieces can stay? Does anyone still need the thing everybody is terrified to touch? That's how Sparq's engineering leaders describe the traditional opening move. Derek Perry calls it "the archaeology standpoint."

Six weeks in, you have a map. Modernizing anything is a separate meeting, on a separate calendar, that hasn't happened yet. That price tag shaped years of scope decisions: when excavation costs this much, teams keep the hole small.

The Shovel Got Smarter

This is where AI gets legitimately useful.

Legacy code contains more than syntax. It contains business rules, dependencies, old architecture decisions, and a decent amount of institutional memory that never made it into Confluence. Derek's point in Sparq's modernization work is that AI can work backward from the code toward its semantic meaning, pulling out business requirements that humans could recover manually if someone had enough afternoons to lose.

The market is already proving the scale of that idea. Morgan Stanley told CIO Dive its in-house platform, DevGen.AI, has reworked more than 17 million lines of legacy code and saved developers more than one million hours, with some week-long coding tasks dropping to half a day.

Sparq built Blueprint.IQ against the same problem and stress-tested it inside The Shop before it went near a client repository. It reads legacy repositories and produces an application dossier covering architecture, dependencies, business context, code health, security findings, and modernization options, in hours.

It's a version of the John Henry problem: the machine makes the person with the hammer faster, not obsolete.

The Menu Got Longer

There are plenty of ways to deal with an old system. Keep it running. Rehost it. Replatform it. Refactor pieces. Rebuild it. Replace it. Consolidate it with something else. Retire it and enjoy the small miracle of having one less application.

A few of those options used to die in the budget meeting before the architecture meeting. AI-assisted development has made replacement and rebuild credible in situations where greenfield development once cost too much to entertain.

The market is moving the same way, despite the LinkedIn chorus insisting AI would simply kill the mainframe. Kyndryl found 80% of organizations changed their mainframe modernization strategy in a year, while the share of workloads planned to move off the platform dropped from 36% to 28%. The real pattern looks more like triage than exodus: get smarter about what to move and what to leave alone.

Less theology, more engineering.

Some Artifacts Earn Their Keep

This is the part legacy conversations tend to flatten. Old code can be ugly, brittle, poorly documented, and full of decisions worth keeping.

A custom application often exists because the business had a reason to be different. Somewhere inside that 15-year-old order-management system may be pricing logic, exception handling, routing rules, or customer behavior that never existed anywhere else. The people who designed it leave. The requirements document goes missing. The code keeps showing up for work. Eventually, the implementation becomes the record.

Good archaeology helps separate the accumulated rubble from the business logic that earned its place. Blueprint.IQ's early use on legacy systems is instructive here: business-rule mapping and documentation drew more attention than the vulnerability scans, for applications where that knowledge had largely lived in people's heads.

Finding the artifact is increasingly cheap. Deciding what it's worth is still somebody's job, and no dashboard is coming to fill it.

Then You Hit the Data

And here the curve gets a lot less dramatic. Brittle applications teach people to work around them. Fields get repurposed. Definitions drift. Records diverge. Somebody exports something to Excel because changing the application would take six months, and seven years later that spreadsheet is part of the operating model.

Sparq COO Brian Carter calls data the Achilles heel of modernization for exactly this reason. AI can find inconsistencies and make the analysis much faster. Data cleanup, ETL, master data work, and the business decisions behind them improve at a much more normal human pace.

A model can show you that three definitions of "customer" exist. The meeting where Finance, Operations, and Sales pick one is still on the calendar.

The money is moving out of discovery and into decision-making.

A Cheap Map Is Not a Safe Exit

Cheaper discovery is not a green light to leave the mainframe.

Gartner expects more than 70% of mainframe-exit projects started this year to miss their intended benefits, largely because leaders are overestimating what generative AI tooling can do. Vendors are folding AI into their offerings regardless of whether it improves the outcome, and a mission-critical application carries the same risk to move whether the discovery phase took six weeks or six hours.

Knowing what a system does faster doesn't reduce the risk of moving it. Those are two separate bets.

The Questions Worth Asking Before You Lock the Next Scope

Before the next legacy conversation turns into a six-week discovery sprint, a few questions are worth putting on the table:

How much of your last modernization budget went to finding out what the system does, versus deciding what to do about it? If most of it went to the first question, AI-assisted discovery is worth piloting on the next system.

Which applications is your team avoiding because nobody's confident they understand what's inside them? That's the archaeology tax showing up as avoidance instead of a line item. It's worth pricing out directly.

If discovery got dramatically cheaper tomorrow, what decision would still take a committee? Data definitions, ownership of a workflow, which pieces of a system are load-bearing: those conversations don't move faster just because the map does. Knowing that in advance changes how the project gets scoped.

Are you scoping a modernization decision or an exit decision, and does everyone in the room agree on which one it is? The Gartner and Kyndryl data above point to two different risk profiles. Confusing them is how a discovery win turns into a migration write-off.

Execution is the peer review that matters here. Ask these before the roadmap gets built, not after.

See what Blueprint.IQ could find on one of your own systems →

Meet the archaeology tax's mascot. Sparq's We're Into That series just put Legacy Stack, the system this whole post is about, on a couch to work through his commitment issues. Watch the first episode →

Frequently Asked Questions

What is the "archaeology tax" in legacy modernization? The archaeology tax is the cost of paying senior engineers to reconstruct what a system does, its architecture, dependencies, business rules, and history, before anyone can safely decide what to build, rebuild, or retire. For a large legacy system, this discovery phase has historically taken weeks and consumed the time of the people an organization can least afford to pull off other work.

Can AI document a legacy codebase that has no existing documentation? Yes. AI can work backward from the code itself toward its semantic meaning, reconstructing architecture diagrams, dependency maps, and business logic directly from what the application does in production, rather than from documentation that may never have existed or stopped being accurate years ago. Blueprint.IQ compresses work that used to take weeks of interviews and tribal knowledge into a structured dossier delivered in hours.

How long does an AI-assisted legacy code assessment take? The analysis itself typically runs in hours. Total turnaround, including setup and reviewing the results with stakeholders, is usually measured in days rather than the weeks a manual discovery process requires.

Does cheaper AI-assisted discovery mean it's safe to migrate off the mainframe? No. Discovery and migration carry different risk profiles. Gartner expects more than 70% of mainframe-exit projects started this year to miss their intended benefits due to overestimating generative AI's capabilities, and Kyndryl's own research shows the share of workloads organizations plan to move off the mainframe has fallen even as more of them adopt AI-assisted modernization strategies. Understanding a system faster doesn't reduce the risk of moving it.

What's the difference between AI-assisted discovery and AI-assisted migration? Discovery is about understanding what a system does: its architecture, dependencies, business rules, and risk exposure. Migration is about acting on that understanding, moving, rebuilding, or retiring the system. AI has meaningfully compressed the cost of discovery. The judgment migration requires, especially around data and exception handling, still runs at a normal human pace.

How does Blueprint.IQ fit into this? Blueprint.IQ is Sparq's agentic codebase intelligence accelerator, built for the discovery phase specifically and stress-tested inside The Shop before it reaches a client repository. It ingests legacy repositories and returns a structured application dossier, architecture, dependencies, business context, code health, security findings, and a phased modernization roadmap, in hours, giving engineering and operations leadership a complete, accurate picture of what they're working with before committing to a modernization path.