Do you need to migrate before you get AI?

Why this decides business cases before they start

Most of the AI you see demonstrated for enterprise systems assumes you run the vendor’s cloud ERP. Most large enterprises do not, or not yet. A large share of the SAP install base still runs ECC or a self-managed S/4HANA. Plenty of Dynamics customers run Finance and Operations in their own data centres. Oracle’s E-Business Suite, JD Edwards and PeopleSoft estates are measured in decades.

So the practical question is a plain one: what can we get without migrating, what needs a cloud subscription, and what needs the cloud ERP itself? Here is the answer per vendor, from their own documentation.

SAP

SAP’s assistant is called Joule, and SAP’s own integration guide (the current one, dated August 2026) lists the systems it works with. They are cloud systems: SAP S/4HANA Cloud in both its public and private editions, SuccessFactors, and the rest of the cloud portfolio. The guide’s prerequisite is a licence for “an SAP product that supports Joule integration … in one of the supported data centers.” Self-managed S/4HANA and ECC on traditional perpetual licences are not on that list.

One consequence worth stating plainly: if you are on RISE with SAP (the private cloud edition) you are in scope, and you get SAP’s free tier of AI. SAP’s pricing page says Base AI “is included in standard cloud subscriptions and can be used without limits or additional cost.” Note the word base: the free tier covers navigation, answers grounded in SAP’s own content and simple transactions. The agentic capabilities are metered separately, which was the subject of Issue #7.

Now the part that changed this year. At Sapphire in May 2026, SAP stated: “SAP S/4HANA on-premises and SAP ERP Central Component (SAP ECC) customers are not excluded: those that commit to transitioning the majority of their current landscape to SAP Cloud ERP gain access to select AI scenarios, bridging the gap between their current landscape and their cloud destination.”

Read both halves. On-premise and ECC customers can now reach some AI capabilities, that is new, and it reverses SAP’s stated position from 2023, when new innovations were not to be made available for on-premise ERP. And the access comes with a condition.

One architectural point that matters for how you brief your team: even under this arrangement, nothing runs inside your on-premise system. Joule is a cloud service that reaches into it over a secure connection. “Switched on for ECC” is accurate; “running on ECC” is not. And the difference decides which team owns the work.

What an on-premise customer can use today, without any commitment, is SAP’s cloud platform for AI; the layer where document reading, the AI models and the custom-agent tooling live. SAP’s own reference architecture confirms these services work with on-premise systems including ECC through a secure connection. That is real capability: invoice and document extraction, answering questions from your own content, custom agents that reach into your systems.

Two caveats worth knowing before you budget it. These are paid platform services with their own entitlements, not something dormant in your existing contract: SAP’s documentation states the generative AI hub “is available only as part of the extended service plan”, and the document-processing service reserves AI-based extraction for its paid editions, metered in the same AI Units as the rest. An organisation that has never bought SAP’s cloud platform starts from zero here, not from a discount.

And there is a boundary that has moved this year and deserves a place in any build plan. SAP’s API Policy now states that, other than through SAP-endorsed pathways, SAP prohibits API use for “interaction or integration with (semi-) autonomous or generative AI systems that plan, select, or execute sequences of API calls”, and reserves the right to throttle or suspend access. The policy is not limited to cloud deployments. If your plan is a home-built agent calling SAP interfaces in a loop (which is exactly what the connector route makes technically easy), this is the document to read before the design review, and the question to put to SAP is which endorsed pathway your use case is expected to run through.

The clock underneath all this has not moved: “SAP will provide mainstream maintenance for SAP Business Suite 7 core applications until end of 2027. This offboarding phase will be followed by optional extended maintenance until end of 2030.”

Microsoft

Microsoft is the clearest of the three, because it publishes a side-by-side table. For Dynamics 365 Finance and Operations, the “AI and Copilot features” rows read: any Copilot capabilities — cloud yes, on-premises no. AI-generated summaries — yes, no. AI agents — yes, no. The reason is architectural: Copilot depends on Microsoft’s cloud data platform, which the on-premise deployment does not have.

Business Central, the mid-market product, is the same in one sentence: “Copilot is exclusive to Business Central online. Therefore, it isn’t available for other deployment types, such as on-premises or private cloud.”

Microsoft 365 Copilot (the one in Outlook, Teams and Word) is cloud-native by construction. It requires cloud identities and cloud-hosted mailboxes.

The on-premise path at Microsoft is therefore to build rather than switch on: Copilot Studio, Microsoft’s tool for custom agents, can reach on-premise systems through a secure gateway. Notably, Microsoft’s own documentation describes doing this against SAP specifically, with ready-made connectors into SAP’s standard interfaces. So a custom agent against an on-premise system is possible. An embedded Copilot is not.

Oracle

Oracle draws its line between product families rather than deployment types. The embedded agents belong to Fusion Cloud Applications: “Running on Oracle Cloud Infrastructure, Oracle AI agents are prebuilt with advanced security and natively integrated within Oracle Fusion Applications at no additional cost.”

For the on-premise products, Oracle’s own JD Edwards FAQ is direct, so it is worth quoting three answers in sequence. “Oracle AI is available only on OCI. However, JD Edwards customers with on-premises deployments can also connect with Oracle AI.” “You can connect your on-premises JD Edwards installation with Oracle AI. You do not need to migrate to OCI.” And on cost: “Oracle AI services are available on OCI through your OCI subscription. You will be charged separately.”

E-Business Suite has a documented path for asking questions of the system in plain language through Oracle’s cloud AI. PeopleSoft’s latest toolset, released on Oracle’s cloud in July 2026, adds plain-language search and an AI assistant for developers, “currently available on OCI, with on-premises availability to follow soon.” And the support horizon is long: Premier Support for E-Business Suite 12.2 now runs “through at least 2037.”

So Oracle’s on-premise customers have the most permissive answer of the three: connect, do not migrate — with a separate cloud bill attached, and without the embedded agents.

Hybrid systems

It would be easy to conclude the question has two locations: your ERP, and the vendor’s cloud. Most large enterprises have a third. The data that decides things increasingly sits in an enterprise data platform (Databricks, Snowflake and the like) that the ERP does not own. An assistant that only sees the ERP answers from a partial picture.

All three vendors acknowledge this and have built a door for it. SAP’s is Business Data Cloud, which includes an SAP-managed Databricks and connects non-SAP data to SAP data without copying it; SAP describes it as the layer that “equips Joule agents with a single trusted data layer”, and announced the equivalent connection to Snowflake in November 2025. Microsoft’s is Fabric, which brings Snowflake and Databricks data into Microsoft’s platform, from where a data agent can inform a custom Copilot, currently in preview. Oracle’s is the Autonomous AI Lakehouse, which reads data held in Databricks, Snowflake and AWS catalogues in place, alongside Fusion data.

The deployment question on this page and the data-location question are different questions, and the second deserves its own issue. For now, the practical point: when you classify a use case as embedded assistant or integration AI, also write down where its data lives. If the answer is “mostly in the data platform”, the vendor’s assistant is only part of the design, and the door above decides the timeline.

What it means for you, specifically