The ERP reckoning: Decades of customization could block AI value

For years, ERP customization was the smart play. When SAP or Oracle systems didn’t bend easily to the way a factory ran shifts or a distributor priced products, CIOs built around it, modifying, extending, or hard-coding until the system matched the business rather than the other way around.

And, for some time, customization worked. It also quietly became one of the most expensive decisions many companies never explicitly signed off on.

There’s a correction under way in enterprise IT, however, and AI is what’s forcing it out into the open.

Angela Maragkopoulou, chief information and digital officer at Sunlight Group Energy Storage Systems, has been direct with her leadership team about what standardizing on an ERP system core buys them. “Don’t sell standardization to your board as technical cleanup,” she says of the case for a “clean core” ERP strategy. “Frame it for what it is: the price of admission to AI.”

That price isn’t one every CIO is willing to pay yet. Even as AI investment accelerates elsewhere in the business, the share of organizations that prefer a customized approach has barely moved: from 52% in late 2024 to 51% by August 2025, according to IDC’s most recent research. “It appears to be a strategic minority move today, concentrated in large enterprises with the technical debt pain, and not yet a broad-based CIO spending pattern,” says Mickey North Rizza, group vice president of enterprise software research at IDC.

James Baker, a partner in Bain & Co.’s Enterprise Technology practice, is seeing nearly the same thing among his clients: Companies, he says, are not yet “funding wholesale de-customization purely for AI.”

Until then, a competitive gap is likely to widen between those CIOs addressing a decades-old problem and IT leaders still deciding whether it’s worth touching at all. At stake is the ability to make the most of AI going forward, as grafting AI onto an existing mess quietly inherits every flaw in it, too.

Why the customization happened in the first place

No one got here by mistake,” says Maribel Lopez, founder of Lopez Research.

“You have to think of why they customized in the first place,” she says. “ERP companies have been trying to get customers to move to the cloud for years. One of the reasons they wouldn’t is they are loath to give up the customization. Customization equals specialized business process in ERP.” It was a rational choice at the time.

What’s changed is who, or what, is trying to use that customized system. Older reporting and automation tolerated customization because a human was still reconciling the exceptions. “AI agents, especially agentic AI, need a single, trustworthy version of the process to act on,” North Rizza says. “Customization fragments this version; … custom code, bespoke workflows, and legacy architecture become a ‘black box’ risk rather than just maintenance overhead.”

Lopez likens customized ERP to a bespoke bicycle. A bike built for a person five feet tall would be hard for someone six feet tall to ride. “You could get some things to work, but it wouldn’t work the same way,” she says. A vendor’s embedded AI is built for the standard bike.

Customization doesn’t make the system unrideable. But it means every mile will cost you more.

Where this shows up first is fairly consistent across the companies IDC tracks: finance close and reconciliation, procurement contract management, and supply-chain demand planning. That is, the high-volume, rules-based processes where agentic AI has the clearest early payoff, and the least tolerance for a fragmented version of the truth.

A CIO who reconciled this early

Sunlight’s S/4HANA implementation was, in Maragkopoulou’s telling, a standardization decision first and an AI decision in hindsight. The triggers were concrete: Its legacy SAP system was heading out of support in 2027, and running parallel systems across factories in Greece, Germany, and the US was expensive and fragile; moreover, leadership was eager for real-time, role-based reporting.

“AI wasn’t on the agenda when we scoped it,” Maragkopoulou says. “But standardization turned out to be the best AI decision we made, long before AI was the headline.”

Sunlight kept customization only where it was a genuine differentiator — a variant configurator that automates complex B2B order entry — and cut the rest, including what she calls a “convenience layer” of old screens and reports that had quietly accumulated without anyone questioning them.

The payoff came later. “The simplified data model and embedded analytics mean the data is consistent and trustworthy, which is most of the battle,” says Maragkopoulou. “A clean foundation makes AI possible, not automatic.”

Maragkopoulou’s advice to other CIOs is to standardize first, or don’t bother. “If you bolt AI onto a heavily customized landscape, you don’t get intelligence,” she says. “You get your complexity, automated and scaled.”

A realization, or a trap?

Not every CIO reads this the way Maragkopoulou does.

Scott Hicar, a fractional CIO who has lived through several SAP platform transitions (R/3 to ECC, and now ECC to S/4HANA), calls these shifts a vendor revenue strategy, not a strategic CIO choice. CIOs get pulled into them whether they intend to or not.

“ERP companies strategically, every five to 10 years, need to force a re-licensing or re-upgrade sort of event to generate enough top-end revenue to meet their plans,” he says. AI, in his view, is this cycle’s justification: new capability available natively only on the latest platform, paired with the pressure of an end-of-life deadline.

“It always made me mad as a CIO that my maintenance dollars were funding the R&D of a next platform I ultimately wasn’t licensed for,” he says.

Hicar isn’t just wary of being swept into an expensive re-licensing event on a vendor’s timeline. He’s also skeptical of CIOs stacking custom AI atop a legacy system rather than paying to standardize. That custom AI, he says, risks going obsolete, or simply being duplicated for free, once a company finally upgrades to the standardized platform.

“The shelf life of AI investments is really short, given the speed of innovation,” Hicar says. “It’s hard to make these one-off bets pay off before obsolescence.”

Even CIOs who accept the case for standardization may not agree on how to get there. Some go underneath, cleaning up the core system itself; others layer AI agents on top of legacy systems instead, leaving the customization in place. SAP HANA migrations alone average roughly 15 months, according to an IDC survey fielded in May 2025 — and that’s before a company has touched a line of custom code. Either path is expensive.

“The question,” Lopez says, “is where will people spend the money.”

The risk of waiting

The cost of sitting on the fence compounds. IDC’s CIO Sentiment Survey, fielded in April 2025, ties technical debt to 49% to 60% higher maintenance costs at large enterprises. North Rizza is frank that it “slows AI deployment and scaling.”

A November 2025 IBM study found that unaddressed tech debt can consume 18% to 29% of AI implementation costs and stretch a 30-month project to 36 months. Even before AI enters the picture, 80% of ERP transformations miss their goals, according to Bain’s AI-Enabled ERP X-Ray. Layering AI on top of an already unstable foundation, Baker says, “doesn’t solve the problem; it just makes it harder to unwind.”

For now, many companies are doing neither. They aren’t standardizing, and they aren’t sitting still. They’re piling AI onto whatever already exists, inheriting the same fragility that has always come with customization, just with higher stakes attached. Maragkopoulou is convinced that’s a mistake CIOs will come to regret. Hicar agrees the workaround is no safer; AI built to avoid the upgrade risks going obsolete before it pays for itself, he argues.

The bill is coming regardless. What’s left is a simpler question: whether CIOs choose when they pay it or let a vendor’s deadline choose for them.