Two years ago, putting a chat box on top of an ERP was a differentiator. It is now table stakes. Every major suite vendor ships an embedded copilot and a catalogue of task agents, and every one of them uses close to the same words to describe it — 'AI built into every workflow', 'agentic', 'in the flow of work'.
When every vendor makes the same claim, the claim stops carrying information. So we went through what each of them is actually shipping. Three things separate them, and none of them is the chat box.
One: whose data grounds the agent
An agent is only as good as what it can look up. SAP’s strongest card here is depth — Joule is grounded in SAP Knowledge Graph and Business Data Cloud, which means it understands business semantics rather than just tables. That is a genuine lead, and it is the reason a generic assistant pointed at SAP tables gives shallower answers.
The catch is what that grounding depends on. Joule’s value is tied to S/4HANA cloud and BDC adoption. If you are still on ECC or private cloud — and a great many companies are — you get considerably less from it than the marketing implies.
Two: who orchestrates across systems
This is the gap nobody has closed. Every vendor is strongest inside its own suite and weakest the moment a process crosses a boundary. Oracle’s agents are concentrated in Fusion. Workday covers finance and HR but has no supply chain or manufacturing. Infor is deep in its industries and thin on builders and marketplaces.
Meanwhile the actual enterprise runs SAP next to Salesforce next to Workday next to ServiceNow. A question worth asking — why is this order late — routinely spans two or three of them. No suite vendor governs agents across that boundary, because no suite vendor is incentivised to.
Three: how it is priced
Pricing is where the market is least honest, and where buyers get caught. The shapes on offer differ enormously:
- Microsoft publishes clear per-user list prices — Dynamics 365 Finance at $210 per user per month, Premium at $300.
- Workday sells Flex Credits, consumed by usage.
- NetSuite meters AI Units.
- Infor bundles into a flat-fee suite.
- SAP and Oracle state no AI pricing on their product pages at all.
Consumption pricing is the one to model carefully before you sign. It looks cheap during a pilot, when a handful of people ask a handful of questions. It behaves very differently once an agent is running against every invoice.
Agent counts are the wrong number to compare
Oracle leads on breadth with over 100 Fusion agents and a partner marketplace, and that figure sets the benchmark every other vendor gets measured against. It is also the figure we would trust least.
Customers consistently report that only a small fraction of available agents ever reach production. A catalogue entry is not a deployed capability. The number that matters is how many agents are running against live data with someone accountable for the outcome — and that number is far smaller, everywhere.
Unclaimed, by everyone: proven, audited return per agent. Vendors that lead with ROI dashboards are the exception, not the rule.
Where the real openings are
Reading across all six, the same four gaps appear — not because the vendors cannot close them, but because closing them works against how each is structured.
- Cross-system agents. The work spans vendors; the agents do not.
- Industry depth on SAP. Infor has proved buyers will pay for agents that already know their industry’s processes and regulations. On SAP, those packs are thin.
- The ECC and private-cloud majority. A large installed base is not on the platform the flagship copilot assumes, and needs value before the migration, not after it.
- Measured outcomes. Hours saved, days to close, error rate — per agent, audited. Almost nobody publishes this, which is exactly why it is worth being the one who does.
The threat worth naming
The structural risk to every ERP vendor is that the ERP becomes a back end. General assistants — Microsoft 365 Copilot, ChatGPT Enterprise, Claude — connected over MCP and open APIs let people ask their questions from whatever tool they already have open. NetSuite has already shipped an MCP connector for exactly this. Once ERP data is reachable that way, the ERP screen stops being where the value is decided.
Our read is that this is a shift to build with rather than resist. The durable position is not owning the chat window — the suite vendors will win that. It is owning what happens after the answer: the action, the approval, the record of who authorised it, and the number at the end proving it was worth doing.