Will AI replace your ERP? The IBM i answer is different

SAP says agentic AI will run your business end to end. McKinsey says ERP survives but people stop touching it. IT Jungle brought the question to the IBM i community in July, and the honest answer for most shops on this platform is that the debate is about somebody else’s software. Three quarters of IBM i shops do not run a packaged ERP at all.

The agentic AI argument everyone is having

The case for disruption goes like this. ERP has been the backbone of corporate computing for thirty years. Application logic is being commoditized, and some vendors are accelerating that themselves. SAP will no longer let customers modify its software, which turns business logic into a fixed product rather than something you own. Meanwhile monolithic code bases keep breaking apart into REST services. Once every function is an API call, training an agent to do what a human does on a screen looks like an implementation detail rather than a research problem.

SAP made this pitch official at Sapphire in May 2026 with its Autonomous Enterprise messaging: not better interfaces, not smarter assistants, but agents handling operational work end to end.

McKinsey’s version is more restrained and more likely. ERP does not disappear, it grows AI characteristics. Application logic still enforces business rules, systems of record still provide auditability, and humans still oversee outcomes. What changes is that users stop interacting with the system directly. Agents execute the workflow and people supervise it.

Forbes ran the sharpest counterpoint in May: deploying agentic AI inside ERP workflows without guardrails, immutable audit logs, and a defined security posture is a liability rather than a capability.

Both sides are arguing about packaged ERP. That is where the IBM i conversation goes sideways.

The statistic that changes the question

The 2026 IBM i Marketplace Survey from Fortra found that homegrown applications written in-house remain the most commonly run workload on IBM i, at 73% of respondents. Oracle and Infor products together account for 25%. Vertical write-in applications for medical, finance, warehousing, manufacturing and public administration make up more than a quarter of business applications on the platform.

That is the same survey that put the skills shortage above security as the community’s top concern for the first time since 2017.

Sit with the 73% for a second. When an analyst writes that AI will disrupt the relationship between companies and their ERP vendors, that sentence has no object for most IBM i shops. There is no vendor. There is a system somebody built in 1991, extended every year since, and named after a department that no longer exists.

So the question is not whether an agent will replace your ERP. It is whether an agent can do anything useful with a codebase nobody has fully mapped.

What “agents replace ERP” actually assumes

The disintegration theory rests on one specific assumption: that business logic is observable. Either it lives in documented APIs, or an agent can infer it by watching how people use the screens.

Both work reasonably well for packaged software. Vendor documentation exists. Screen flows are standardized because thousands of customers use the same ones. The rules are largely the vendor’s rules.

Neither holds on a homegrown IBM i system, and the reason is not that RPG is old. The reason is that the logic is not where an agent would look for it.

Your pricing rules are in a program that a night job calls, not in anything a user touches. Your real month-end close is a CL sequence with dependencies documented in one person’s memory. There is a field in a physical file that meant one thing until 2004 and something else afterward, and three programs know the difference. There is an exception coded for a single customer who is somehow still a customer. None of that is inferable from screen behavior, because none of it happens on a screen.

An agent pointed at that estate does not fail loudly. It produces confident, wrong output, which is worse, because in an ERP context wrong output is a shipped order or a booked journal entry.

This is also the honest reason why rip-and-replace migrations go badly so often. The undocumented logic is the system. Replacing the container does not extract it, and neither does an agent.

There is some irony in IBM’s own position here. Its Institute for Business Value published a report in May 2026 arguing that custom logic and inconsistent definitions across ERP systems create risk rather than speed in an agentic model. That is a defensible argument. It also describes roughly three quarters of the install base of IBM’s own midrange platform.

If you are on Infor or JD Edwards, check the roadmap

The 25% running packaged ERP have a more concrete way to check the claims: look at what the vendor is actually shipping, not what the keynote said. The Infor IBM i ERP roadmap for 2026 is public, and it is specific.

Infor’s roadmap for its IBM i ERP portfolio targets IBM i 7.6 support across XA, LX and System21 in 2026, with REST APIs arriving for all three in the same window. System21 version 3.3 is targeted for late 2026 with localization, Java modernization, NIS2 support and broader IDF integration. System i Workspace Liberty, a Java 17 compliant UI layer that moves green screen function into a browser based thin client, is expected in the third quarter. Published plans run out to 2034.

Read that list again. REST APIs, a Java refresh, and a browser client. That is plumbing. It is genuinely useful plumbing, and it is not an autonomous enterprise.

Oracle’s JD Edwards World and EnterpriseOne customers, meanwhile, have largely declined to move to other Oracle applications and continue running on IBM i, which is its own kind of answer. Oracle has since made its database better at accepting IBM i data, though the parts of the platform that actually hold shops in place are not the parts that release addresses.

If your vendor’s ten-year roadmap is about API surfaces and UI modernization, you are not going to wake up in 2027 with an agent running your order-to-cash. You may well get agents on top of that API surface, which is a different and much more plausible thing.

Where agents genuinely fit on IBM i today

Three different products get called “AI agents” for IBM i and they do unrelated jobs. Keeping them straight is most of the clarity here.

Coding agents write and refactor code. IBM Bob, Copilot, Claude. These are real and they work, with supervision. We covered what changed in Bob 2.0 and the short version is that RPG generation and documentation have gotten genuinely useful.

Process agents execute business workflow across applications. IBM positions watsonx Orchestrate here, with digital workers performing activities in ERP, CRM, HR and supply chain systems. This is the category the ERP debate is actually about.

Infrastructure agents run the platform itself, handling operations by conversation rather than dashboards.

Most of the excitement reaching IBM i shops is category one, and most of the disruption argument is category two. They get quoted in the same breath and they should not be. A coding agent that writes decent RPG tells you nothing about whether a process agent can safely run your close.

The number the bulls skip

Gartner forecasts that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, and that 33% of enterprise applications will include it by then, up from under 1% in 2024.

The same firm predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.

Both forecasts can be true. Adoption climbs while most individual projects fail, which is what every enterprise technology wave has looked like from the inside.

In an IBM i shop, “unclear business value” has a specific shape. Someone pilots an agent against an order entry process. It works on the clean path. Then it hits the customer with the special pricing agreement, or the warehouse that reserves stock differently, or the month where the fiscal calendar does not line up. Each exception gets patched by hand until the agent needs more supervision than the process it replaced. The pilot quietly ends. Nobody writes it up.

That failure mode is not caused by the agent. It is caused by the exceptions never having been written down.

What to actually do in the next twelve months

Two things, and neither requires a position on whether SAP is right.

Document the business logic. Not the code, the rules. What decides a price, what triggers a hold, what the exceptions are and who they exist for. AI is good at this now, which is a genuine change from three years ago. Pointing a model at RPG source and getting a readable explanation of what a program does works well enough to be a normal part of the job. Our IBM i AI Field Guide covers the prompts and workflow for exactly this if you want a starting point rather than a blank editor.

Build a clean API surface over the programs that matter. Not all of them. The twenty or thirty that carry your actual business. If you want the broader picture on sequencing, where to start with IBM i modernization walks through it, and IBM’s own Redbook on modernizing IBM i applications goes deep on the database and interface layers underneath that API surface.

Here is the test that makes these worth doing: both pay for themselves whether or not agents ever show up. Documented logic survives the developer who retires next year. A clean API surface makes integration, reporting, and any future migration cheaper. Compare that to a rip-and-replace program, which only pays off if the replacement works.

Bottom line

AI is not ending the ERP era on IBM i, mostly because three quarters of IBM i shops never entered it. What AI is ending is the era in which undocumented business logic was survivable. Agents make that liability expensive in a new way, on top of the way it was already expensive when the person who wrote it retires.

The move is not to wait and see which analyst was right. Document the rules, expose the programs that matter, and you are better off in every version of 2030, including the boring one where your ERP is still there and still running the business.

Frequently Asked Questions

Will agentic AI replace ERP systems?

Not in the next few years, and probably not the way vendors are pitching it. Gartner forecasts that 33% of enterprise applications will include agentic AI by 2028, but the same firm predicts more than 40% of agentic AI projects will be canceled by the end of 2027 over cost and unclear value. The realistic outcome is agents running on top of ERP, not instead of it.

How many IBM i shops actually run a packaged ERP?

About a quarter. The 2026 IBM i Marketplace Survey from Fortra found Oracle and Infor products in use at 25% of respondents, while homegrown applications written in-house remain the most common workload at 73%. Most IBM i shops are not in the market the ERP disruption debate is about.

What is Infor shipping for its IBM i ERPs in 2026?

IBM i 7.6 support across XA, LX and System21, REST APIs for all three, and System i Workspace Liberty, a Java 17 browser client that replaces green screens. System21 3.3 is targeted for late 2026. The published roadmap runs to 2034 and does not describe agents replacing the ERP.

What should an IBM i shop do about AI and ERP right now?

Document your business logic and build a clean API surface over the programs that matter. Both are prerequisites for any agent strategy, and both pay off on their own if the agents never arrive. Rip-and-replace migrations remain the expensive way to find out what your code actually does.