Two budget lines are moving in opposite directions right now. AI spend is climbing, and the money paying for it is coming out of infrastructure maintenance and legacy software contracts. Most of the coverage reads that as a story about AI enthusiasm. It is more useful to read it as a story about what the AI is being asked to stand on.
We have built and maintained production systems since 1998, and nearly every AI engagement we take starts inside a legacy estate that someone, at some point, decided could wait a year. So we watch this particular trade closely, because we tend to get the call after it goes wrong.
Where the AI money is actually coming from
A Goldman Sachs survey of CIOs, reported by IT Brew, puts numbers on the shift. Forty-two percent of respondents, up from 35 percent in the prior survey, now expect AI to exceed 10 percent of their IT budgets within three years. Overall IT spending dipped slightly over the same period. And two thirds of AI inference cost is being funded by reallocation rather than new money, with Goldman's analysts noting that "there may be some crowding out of category spend not tied to AI."
The categories being crowded out are not mysterious. They are the unglamorous lines: maintenance windows, patch cycles, monitoring renewals, the support contract on the system nobody wants to talk about. In a flat budget, an AI mandate has to eat something, and it eats the line items with the weakest internal advocates.
None of this is irrational on its face. A CFO looking at a maintenance line sees a cost that produced no visible change last year. The AI line has a board mandate behind it. The instinct to reallocate is not the problem. The problem is what the maintenance line was actually buying.
Deferred maintenance is a loan, not a savings
Cutting maintenance does not make the need for it go away. The estate keeps aging at exactly the same rate. What changes is that you have stopped paying down the principal, and legacy systems charge interest.
The interest shows up in specific, predictable forms. Dependencies drift further from current, so the eventual upgrade grows from a sprint into a project. Patch backlogs accumulate on internet-facing systems, which is how a skipped renewal turns into an incident that costs a multiple of everything it saved. Monitoring lapses quietly, so the first sign of a failing batch job is the business process it feeds going dark. And the people who understood the system move on, taking with them the only accurate documentation the estate ever had.
None of that appears on a budget line in the year you make the cut. It appears one to three years later, as unplanned work, at emergency rates, on a timeline you do not control. Deferral is not elimination. It is financing, at a rate no one would accept if it were written down.
The estate decides what the AI is worth
Here is the part the reallocation math misses entirely: the systems being defunded are the same systems the AI investment depends on.
Almost nothing on the AI line runs standalone. Agents read from the estate, write to the estate, and automate processes the estate implements. Their ceiling is set by what they sit on. Point an agent at a clean data model with tested interfaces and it compounds your delivery capacity. Point the same agent at un-normalized tables, undocumented batch jobs, and an integration layer held together by tribal memory, and it does not fix any of that. It amplifies it, confidently and at scale.
The practical failure comes in two flavors. The first is output quality: an AI system grounded in a degraded estate produces plausible answers assembled from wrong inputs, and nobody discovers the gap until a decision has been made on top of it. The second is integration cost: the business case assumed the agent could talk to your systems, and the discovery that it cannot, at least not safely, arrives after the spend is committed. Both failures get booked against the AI. Both were purchased by the maintenance cut.
This is the budget-side version of the argument behind our approach to sequencing AI work: remediation is not a competitor to the AI program. It is the first phase of it.
How to reallocate without eating the foundation
There are real savings inside most legacy estates. The discipline is telling the difference between cutting a system and merely defunding it.
-
Retire deliberately instead of deferring quietly. Decommissioning a system, migrating its last consumers, and ending its contracts is real, permanent savings. Skipping its maintenance while it stays in production is not savings at all. It is risk, rebooked under a nicer name.
-
Triage the estate before you cut it. Remediate, modernize, or leave alone is a decision to make per system, with a framework, not a percentage haircut applied across the board. A haircut cuts the load-bearing systems at exactly the same rate as the dead ones.
-
Fund remediation inside the AI program. If the agent roadmap depends on a data layer that needs cleanup, that cleanup belongs on the AI line, stated plainly in the business case. Hiding it in a shrinking maintenance budget guarantees it loses the argument it should never have been in.
-
Protect the maintenance you cannot see. Monitoring, alerting, backup verification, and support coverage fail silently when defunded. Nothing breaks on day one. The cost surfaces as incident response, which is the most expensive form of maintenance ever invented.
The question to ask before the next budget cycle
The CIOs in that survey are not wrong to fund AI. They are wrong if they believe the estate will politely pause its decay while the AI line proves itself. The honest budget conversation prices both sides of the trade: what the AI is expected to return, and what the deferred work will cost when it comes due, because it always comes due.
If you are staring at that trade and cannot say which of your systems can safely absorb a leaner year, that is the diagnostic to run before the reallocation, not after. An AI readiness assessment exists to name exactly that: the binding constraint between the estate you have and the AI you are budgeting for. It is a much cheaper way to find out than the incident. AI that ships, not AI that demos, starts with an estate that can carry it.