For most of the last decade, clean core was a CIO conversation about upgrade cost. Keep customizations out of the core, carry less legacy code into S/4HANA, and cut the maintenance bill every time SAP ships a new release. That case was strong enough on its own.
It no longer covers the full picture. In 2026, clean core decides something bigger: whether the AI agents built into your ERP can turn on at all.
The access gate nobody priced in
SAP and Oracle have both moved agentic AI from pilot to platform this year. SAP’s Joule now spans more than 40 purpose-built agents and over 2,400 Joule Skills across S/4HANA, Ariba, SuccessFactors, and IBP, executing multi-step workflows across modules rather than waiting for a person to approve each step. Oracle has gone further on volume, embedding more than 600 pre-built agents across Fusion Cloud ERP, HCM, SCM, and CX that carry out processes like invoice matching, period close, and candidate screening autonomously, checking policy and escalating only when needed.
These agents don’t sit on top of your landscape, they operate inside it. They read business objects directly and act within defined guardrails, using live data. That only works reliably when the objects they’re reading are the standard ones the agent was built against.
Two tracks, same release
This is already splitting enterprises into two groups, and it isn’t a future risk to plan around. It’s the outcome of architecture decisions made years ago, showing up in board-level AI roadmaps now.
Why messy cores actually break agents
The technical reason is simple once you sit with it. Custom code buried in core tables, modified user exits, and undocumented workarounds change how a business object behaves in ways an agent can’t reason about. An agent built to work against a standard Payables object doesn’t know what to do when your organization’s version of “Payables” has three custom fields and an approval chain only one person from the original implementation still understands.
Side-by-side extensibility on SAP BTP exists precisely to solve this: custom logic runs alongside the core instead of inside it, so agents keep interacting with clean, standard objects while your organization keeps the flexibility it built over the years.
The governance stakes went up too
The EU AI Act’s high-risk provisions become enforceable on August 2, 2026. Autonomous agents used in regulated decision-making fall squarely inside that scope, which raises the bar from “did someone approve this” to “can we reconstruct why the agent decided this.”
For the life sciences, BFSI, and other regulated sectors Varchai works with, that traceability bar was already the standard for GMP, FDA, and financial compliance. Clean core is what makes that same traceability achievable for AI-driven transactions, not just human-approved ones.
What AI-ready clean core actually requires
Five things separate a landscape that’s merely upgrade-safe from one that’s actually ready for autonomous agents.
The upgrade question was never the real question
Clean core was always worth doing for upgrade cost avoidance alone. But in 2026, it’s the difference between paying for AI agents you can use and paying for AI agents you can’t turn on. The question enterprises should be asking about their SAP or Oracle landscape this year isn’t “can we upgrade cleanly.” It’s “can we activate what we’re already licensed for.”
Find out what’s blocking activation in your landscape
Varchai runs landscape assessments that map exactly where custom logic is blocking agent activation, and builds clean core roadmaps that get you not just upgrade-safe, but AI-ready.
Talk to Varchai
