ERP AI: SAP Joule and Oracle Fusion Agents, and What Customers Switched On

Oracle says it shipped more than 1,000 AI agents into its Fusion applications in a single year. SAP says Business AI was included in two thirds of its Q4 2025 cloud order entry. Neither company has published how many customers have actually switched an ERP AI agent on. The closest independent number comes from SAP's German-speaking user group, and it is 3%.

What ERP AI vendors count on an earnings call

ERP AI is the label for assistants and agents built into enterprise resource planning suites. Those are the finance, procurement, supply chain and HR systems a company runs its books on. An assistant answers or drafts when asked. An agent takes a step on its own, such as matching an invoice or posting a journal.

Every number a vendor reports about these products falls into one of three buckets. Supply metrics count what the vendor built. Sales metrics count what the vendor attached to contracts. Activation metrics count what customers actually turned on and kept running.

The first two are easy for a vendor to produce and easy to grow. The third is the only one that tells you whether the product works. Across four quarters of SAP and Oracle disclosures read for this post, the third bucket is almost empty.

That is the core finding, and I think it matters more than any individual agent count. A vendor that knows its activation rate and declines to publish it is telling you something. A vendor that does not know it is telling you something worse.

What SAP has disclosed about Joule agents

SAP's AI story runs through Joule, its assistant, and a growing set of agents that sit behind it. The disclosures are frequent and specific about plans. They are thin on what customers have running today.

In its Q4 2025 results, SAP chief executive Christian Klein said Business AI was included in two thirds of Q4 cloud order entry. Total cloud backlog stood at about €77.3 billion at the end of 2025, per Verdict's report on the results.

Read that sentence carefully. "Included in" means AI was part of the contract. It does not mean the customer has deployed it, and order entry is measured by value, so a handful of very large deals can carry the ratio.

On the Q2 2026 call on 23 July, Klein said SAP would release close to 50 assistants by the end of Q3, underpinned by more than 400 agents by year end. He described Joule Work, a single interface across SAP's portfolio, as launching that quarter, with hundreds of customers in beta.

Two other lines from the same call matter more for adoption. Klein said over 3,000 SAP consultants were driving AI adoption directly with more than 2,000 customers. He also described a new consulting offer carrying a commitment to activate and adopt AI assistants and agents within the first year, per the Benzinga transcript.

A vendor does not put 3,000 of its own consultants on adoption when adoption is happening by itself. I read that line as the most honest disclosure SAP has made on the subject.

The financial backdrop is strong. Current cloud backlog reached almost €23 billion, up 26%, and cloud ERP suite revenue grew 27% to 88% of total cloud revenue. None of those figures isolates AI.

What Oracle has disclosed about Fusion agents

Oracle's disclosures follow a different shape. It talks about agents in volume and almost never in its formal filings.

Oracle said it had delivered more than 1,000 AI agents across its application suite over the past year, according to CXM's coverage of the Q4 fiscal 2026 results. Forrester heard the same figure at Oracle's April analyst summit, alongside hundreds of partner agents and 73,000 certified developers.

The earnings release Oracle filed with the SEC on 10 June 2026 does not mention that count. It reports cloud applications revenue of $4.1 billion for the quarter, up 10%, and $15.9 billion for the fiscal year, up 11%. The September release for Q1 fiscal 2027 showed the same 10% growth on $4.2 billion, with no agent figures either.

That gap is worth noticing. Statements on calls and at analyst events carry less legal weight than numbers in a filed release. Oracle chose to keep the agent count out of the document its lawyers sign.

Go-lives and the 33-customer number

Oracle said more than 300 Fusion customers went live during the fourth quarter. That is a count of application go-lives, not agent activations. The more revealing figure is smaller: a limited token bundle rollout reached 33 customers in Q4, including Aon and Liberty Energy.

Oracle's clearest adoption claim was qualitative. Mike Sicilia told analysts that customers "have moved past the experiment stage with AI." That may be true. It is not a number, and the 33 is.

The ERP AI agent inventory: GA, preview and promised

Status labels matter more than counts. A generally available agent is supported in production. A beta or preview is not, and an agent promised for year end does not exist yet. The table sorts what was disclosed by that test.

Status is the vendor's own statement or a named analyst's first-hand observation, never an inference. Oracle's 29 agents were reported by Cloud Wars as 13 for supply chain and 16 for customer experience. Whether SAP's Q3 launches shipped on schedule will be confirmed or not in its Q3 2026 results.

Why the live list is all finance back office

The agents with documented production use are narrow. Oracle's ledger agent works on accounting entries. On SAP's Q2 call, Klein cited an agent built with Amadeus that reconciles unstructured payment data and was clearing around 40,000 incorrect transactions, and a customer reporting over 90% touchless processing.

These tasks share three properties. The input is structured, the correct answer is checkable, and a wrong answer gets caught at month end. That is exactly where you should expect agents to work first, and where I would start if I ran an ERP programme today.

Notice what is missing. No vendor has named a customer running an agent that approves spend, changes a supplier master record or posts to the general ledger without a human in the loop. The write-access problem is real, and the permissions question it raises is covered in the analysis of giving agents write access to legacy systems.

SAP AI pricing versus Oracle's no-charge agents

The two vendors have chosen opposite pricing models, and both choices tell you about the business underneath.

SAP meters through AI Units

SAP's pricing page says Joule Base is included at no additional cost in SAP cloud subscriptions. Premium capabilities run on AI Units, a prepaid currency bought annually that expires after 12 months. Joule Premium packages for finance, supply chain and other functions consume units per user per month on a sliding scale, and SAP lists them with a request-a-quote button rather than a price.

On the July call Klein went further, saying SAP is pricing agents on value and sees "a unique chance" to price on outcomes and "completely reset the price level". No metric or rate was given. When a vendor says it intends to reset prices on a product most of its customers have not deployed, the buyer should hear that the meter is not settled yet.

Oracle gives agents away, for now

Oracle's agents ship with the standard quarterly Fusion update at no extra charge. Forrester reported that Oracle reiterated no plan to charge for embedded agents and also that it did not commit to a model for renewal-time pricing once agent workloads scale.

Oracle is simultaneously rolling out outcome-based pricing across its applications, with interview agents priced per candidate screened and hospitality upsell agents priced on conversion. So "free" describes the embedded ERP agents today. It does not describe Oracle's direction.

Oracle can afford free because Fusion agents sell a migration. Rondy Ng told analysts the agents are not coming to E-Business Suite or on-premises products: "To enable this, you have to get to SaaS." Free agents are a reason to move to Fusion, and the Fusion subscription is the product. The cash pressure behind that strategy is covered in the breakdown of Oracle's AI cash flow math.

Prepaid credits that expire are the same structure appearing across enterprise software, and the confusion they create for finance teams is examined in the piece on AI billing credits and tokens. Workday made a similar move with its own credit scheme, and its Flex Credits split between users and payers is the closest public comparison.

ERP AI adoption: what customers say runs in production

Vendors will not publish activation rates, so the best evidence comes from customers and analysts. There is not much of it, and what exists points the same way.

The DSAG survey: 3% on SAP, 77% elsewhere

DSAG, the German-speaking SAP user group, surveyed one decision maker at each of 198 member companies between December 2025 and January 2026. Of the respondents, 43% had implemented AI use cases. Among those, 77% ran them in production with non-SAP solutions and 3% with SAP. In testing, the split was 65% to 8%. At proof of concept, 62% to 26%.

The funnel is the story. SAP's share is 26% at proof of concept, 8% in testing and 3% in production. Something stops SAP AI use cases between pilot and live, and DSAG names the likely causes: complex licence models and heterogeneous system landscapes. 70% of respondents cited SAP licensing and contract terms as a challenge for investment decisions.

What analysts saw at the vendor events

Forrester's verdict on Oracle's April summit was blunt: agent supply is strong, adoption proof is not. Its analysts listed what was missing as production deployment counts, activation rates and business outcome metrics, and called that the key gap in Oracle's AI narrative.

SAP was unusually candid about its own builder. At Sapphire 2026, chief product officer Manoj Swaminathan said Joule Studio adoption "has been minimal compared to what we'd like", per CIO.com's report. The same piece found that tools promised at Sapphire 2025 were technically available but lagging in adoption.

The pattern across both vendors fits what tends to happen when an agent pilot meets real process ownership, a pattern laid out in the review of why agent pilots fail. The technology ships on time. The controls, data and process owners do not.

Where this argument is weakest

The case above rests on thin independent data, and it should be read with that in mind.

The 3% is narrower than it sounds

DSAG surveyed 198 firms, mostly in Germany, Austria and Switzerland, and DSAG itself sent the survey this year, which meant fewer large companies took part than in 2025. The 3% is a share of firms already running AI use cases, not of all SAP customers. Board chair Jens Hungershausen cautioned that SAP use cases are more complex than standard large language model tools, which partly explains slower movement. Secondary articles have repeated the figure as "3% of SAP customers use Joule", which DSAG did not say.

The survey also closed in January 2026. SAP has shipped a great deal since, and a survey taken today could look different. I would not use this number to argue SAP AI has failed. I would use it to argue the burden of proof sits with SAP.

Embedded AI may be used without being counted

Many ERP AI features switch on inside an update and run in the background. A user who accepts a suggested invoice match may not know an agent proposed it, and would not report it in a survey. Oracle's no-charge model makes this more likely. Real usage could be higher than any survey captures.

The angle I set out to cover also included what ERP implementers say goes live. Systems integrators do not publish that data, and none of the large firms has released a count of agent go-lives. SAP's own 3,000-consultant programme is the closest thing, and it is the vendor's data. This post compares vendor disclosures with user-group and analyst evidence because that is what exists.

What to ask before your ERP AI renewal

If you run an SAP or Oracle programme, the disclosure gap becomes your negotiation position. Vendors who cannot show activation across their base can still show it inside your tenant, and you should make them.

Ask for your own usage data first. Both vendors can report which AI features have been invoked in your environment, by whom and how often. If they cannot produce it, you have learned the feature is not instrumented, which is itself a reason not to pay for it.

Then pin the meter. On SAP, get the AI Unit consumption rate for each premium capability in writing, along with what happens to unused units at month 12. On Oracle, get a written statement on whether embedded agents stay at no charge through your next renewal term, because Forrester found Oracle had not committed to that. The clauses worth asking for are in the guide to outcome pricing contract clauses.

Finally, ask for one named reference running the specific agent you are buying in production, at your scale, in your industry. The ledger agent and payment reconciliation agents will probably pass this test. A year-end target agent will not, and that tells you what to put in the contract and what to leave out.

Frequently asked questions

Where to start

Pull a usage report for every AI feature already switched on in your SAP or Oracle tenant this week. You are probably paying for more ERP AI than you use, or using more than you know, and either finding changes your next renewal conversation.

Then pick one back-office task with a checkable answer, such as payment matching or ledger review, and run the vendor's included agent against last month's data. Measure the error rate yourself. That single number will be more useful than every agent count in this post, and the same discipline is how the platform-versus-feature debate in CRM will eventually be settled too.

Related reading