Oracle's $0 Fusion AI Agents vs SAP and Workday Meters: What Startups Face

Oracle's Fusion AI agents cost $0 extra. Oracle says they come natively integrated within Fusion Applications at no additional cost to customers. SAP and Workday also give an entry tier of agents away, then meter the rest. For agent startups selling into those accounts, the free price is the smaller problem. Access rules and acquisitions decide more.

Short answer

Does free bundling starve agent startups? Only the ones that duplicate a single incumbent module. Oracle's no-cost agents erase the price case for a one-module agent inside Fusion accounts. SAP restricts how third-party autonomous agents reach its APIs. Workday meters actions on its platform and has bought partners. Startups that work across systems keep a market.

How Oracle, SAP and Workday price their ERP and HCM agents

Oracle sells finance, supply chain, HR and customer software under the Fusion name. SAP sells ERP and SuccessFactors. Workday sells HR and finance. Each now ships agents inside the suite: software that completes a multistep task, such as paying an invoice, without a person driving every step.

All three give something away. They disagree about where the meter starts and what a third-party agent may do on their platform.

Oracle: no meter on embedded agents

At AI World in October 2025, Oracle said its agents are "prebuilt with advanced security and natively integrated within Oracle Fusion Applications at no additional cost," according to the release distributed by PR Newswire. The release named agents for payables, ledger, planning, payments, HR, supply chain and customer experience.

Forrester, writing a week later, counted over 600 new AI agents embedded at no additional license fee. It quoted executive vice president Steve Miranda: "all the new AI features and agents are provided at no extra cost." Its verdict carried a qualifier: the bundle swaps metered experimentation for predictable cost, "at least for now".

The earlier analysis of what SAP and Oracle customers actually switched on covers adoption. This post asks who loses a sale because of the packaging.

SAP: free Base AI, metered premium agents

SAP's Business AI pricing page says Base AI "is available with every SAP cloud subscription at no additional cost," with no AI Units required. Anything premium runs on AI Units, which SAP calls its "single virtual currency" for premium AI. Customers buy them annually into one shared pool.

The page now publishes a per-action rate. Standard agents cost 0.02 AI Units per action and advanced agents 0.05, so an advanced action burns 2.5 times the units of a standard one. Unused units expire after 12 months. The rates are in AI Units, not currency, so turning an agent action into money needs a quote.

Workday: complimentary credits, then the meter

Workday's Flex Credits page calls the model usage-based pricing for AI agents. Credits are bought annually and drawn down when an agent completes "a defined task with known value". Metering happens "almost exclusively in production". Workday also gives "Complimentary Flex Credits based on your company size" and lets customers test most agents free outside production.

So Workday's entry price is also zero, bounded by an allowance. The credit pool's mechanics are in the analysis of Workday Flex Credits as a seat hedge.

Every ERP incumbent sets the entry price of an AI agent at $0

Side by side, the "Oracle is free, the others charge" framing breaks down. None of the three charges a customer to start using its own agents. What differs is how far a customer gets before a meter turns on.

Workday's numbers show the gap a free entry tier opens between use and payment. Its second-quarter fiscal 2027 release says more than 5,500 customers now use one or more of its organic agents, up more than 35% in a quarter. On the same day's call, president and chief commercial officer Rob Enslin said of Flex Credits: "We have already signed 200 customers this quarter".

Divide 200 by 5,500 and you get roughly 3.6%. One number is a quarter's signings and the other a running total of users, so the ratio is directional. It still shows that most Workday customers touching its agents have not signed for the meter.

When a customer already runs an incumbent agent on complimentary capacity, a startup is not competing with a price. It is competing with something already installed, approved by security and in the user's workflow.

What Oracle's free Fusion AI agents do to startups in the same account

Oracle's packaging is the purest bundling of the three. Two things happen at once. The price case for duplicate products collapses, and the official third-party channel offers distribution without revenue.

The single-module duplicate loses the price argument

Take the most exposed startup: one selling a single agent for a job Oracle now ships. Oracle's October 2025 release named a Payables agent. A payables automation startup selling into a Fusion customer must now justify a new line item against a capability already in the subscription.

The startup's agent may well be better. It still has to clear procurement, security review and an integration project, and the free alternative needs none of them. Free does not have to win on quality. It only has to make the buyer ask why they would pay twice.

This is the absorption pattern in the analysis of point SaaS absorbed into agent ecosystems. The startup kill zone test sets out who sits inside the danger zone. Oracle's pricing pushes that zone outward for anything touching Fusion data.

A marketplace that pays its contributors nothing

Oracle's answer for third parties is the Fusion Applications AI Agent Marketplace. Oracle's release describes it as a way to "find and deploy validated, partner-built AI agents" inside the customer's environment. It says nothing about how partners get paid.

Forrester fills that gap, and for a startup weighing Oracle as a channel it matters most. The marketplace is "free yet unincentivized for contributors" and "currently nontransactional," with partners providing their intellectual property for free. Forrester counted about 40 launch partners, each contributing at least five agents, and named Deloitte among them.

A systems integrator can give an agent away because it sells the implementation hours around it. A venture-backed startup that gives its agent away has given away its product. I read Oracle's marketplace as a services channel and, until it transacts, a showroom at best. Which marketplaces pass money to sellers is covered in the analysis of agent marketplace distribution value.

SAP's squeeze on agent startups is access, not price

SAP charges for premium agents, which leaves a third party more room to compete on price than Oracle does. SAP's constraint sits elsewhere: in who may call its interfaces, and how.

In April 2026 SAP published a revised API policy. DSAG, the German-speaking SAP user group, summarised one provision in its statement of 29 April 2026. Use of SAP APIs "in conjunction with (semi-)autonomous or generative AI systems is also restricted, unless this takes place explicitly within architectures or services provided by SAP". DSAG added that, on the information available to it, existing customer integrations and authorised partner solutions are not affected.

DSAG objected. It said changes to API status, usage rights or supported scenarios "must not be made unilaterally or retroactively," and asked SAP to give customers more time for the transition.

SAP chief executive Christian Klein framed the policy differently on an earnings call quoted by Constellation Research on 8 May 2026. He separated customer data, which SAP has "no plans to monetize at all," from "the IP, the domain know-how sitting in our ERP," which SAP will offer on its platform and protect. He also said "No partner needs to worry," and that SAP would throttle APIs facing mass data egress.

For a startup, an autonomous agent working on SAP data should expect to run through SAP's own routes, not call the interfaces directly. That makes SAP a gatekeeper on access even where it is not a price competitor. I think this is a heavier constraint than Oracle's free agents. A better product can beat a price, but not an access rule. The same pattern across vendors is documented in the review of SaaS API terms and AI walled gardens.

SAP has also bought where it wanted to own a category outright. It completed its acquisition of SmartRecruiters in September 2025, per SAP's announcement.

Workday's agent model: meter the rails, then buy the partners

Workday charges for agent work, recruits third-party agents, and has bought several companies that build them. That makes it a startup's most useful partner, and its most dangerous.

Paradox went from launch partner to acquisition in about two months

On 3 June 2025 Workday launched its Agent Partner Network and named 15 first partners, including Paradox, Auditoria.AI and Glean, per its DevCon release.

By August 2025 Workday had announced deals to buy Paradox and Flowise, an agent-building platform, according to Fortune. In September it agreed to buy Sana, an enterprise knowledge and agent startup, for around $1.1 billion. The same month Workday said the partner network had grown more than fourfold to over 50 partners, including more than 15 Workday Ventures portfolio companies.

That makes 1 of the 15 named launch partners, about 7%, the subject of an announced acquisition within roughly 2 months of joining. It is one data point, not a rate. It still shows what a partner network can also be: a place where the platform learns which agents its customers adopt.

A meter on every action, whoever built the agent

The key line for third parties came from Workday's chief technology officer on the August 2026 call. "Because all of it runs through flex credits, we monetize every action happening on our rails regardless of the path," he said, per the Motley Fool transcript. Chief executive Aneel Bhusri said that whether customers run Workday's agents or build their own on Workday, "we win either way".

Workday's quarterly release also describes Agent Passport, which "tests and verifies every AI agent, Workday-built or third-party, before it goes into production". Together, the two statements make Workday both certifier and meter for agents inside its tenant. A startup agent that reads or writes Workday data does work Workday can count.

Glean shows the tension. It was named a launch partner in June 2025. Three months later Workday agreed to pay about $1.1 billion for Sana, whose knowledge search now sits inside the Flex Credits pool. A partner and an owned competitor now share one marketplace, and only one draws on the customer's prepaid credits.

Where agent startups can still win against bundled AI agents

Free bundling is aimed at a specific shape of product: agents working inside one suite, on its data, for its modules. That leaves positions where a startup's economics survive.

Work that crosses suites

An agent that reconciles a Workday headcount plan against an Oracle general ledger and a CRM pipeline cannot be bundled by any one of those vendors. No incumbent wants a rival's data first-class inside its own agents. That is the opening, the exact reverse of the single-module duplicate.

Raising money today, I would not pitch a single-module agent sold mainly into Fusion accounts. I would pitch one whose hardest work lives between two suites, the part no bundle reaches.

Depth the bundled agent will not reach, or a contract it cannot offer

Bundled agents are built for a suite's median customer. A startup that goes deep into one regulated vertical or one ugly workflow can be the specialist the generic agent is not. The test is whether a buyer would still pay if the incumbent's version improved by half.

The contract is the other opening. Oracle cannot charge for embedded agents without breaking its own message, and SAP and Workday price per action. A startup pricing per resolved outcome offers a different contract, not a cheaper copy. The tradeoffs are covered in the analysis of AI bundling versus metering.

The table below turns that map into what a founder says in a first meeting. The weak versions are exactly the claims a free bundle is built to answer.

Where the free bundling argument is weakest

The case above is built on packaging, and packaging is not outcomes. I found no public evidence of a named startup losing a named deal to a bundled ERP or HCM agent. The squeeze is inferred from price lists, policies and deal announcements, the honest limit of this method.

Free may also be temporary. Forrester's "at least for now" is doing real work in its sentence about Oracle. Oracle's appetite for giving away compute depends on cash flow under visible strain, as the breakdown of Oracle's AI cash flow math shows. If Oracle introduces a meter at renewal, the single-module duplicate gets its price argument back.

The bundled agents may simply be weak. A free agent nobody switches on is not competition, and the adoption evidence for ERP agents is thin. Workday's 5,500 users are a counter-example on the HCM side, but "using at least one agent" is a low bar that says nothing about depth of use.

The incumbents may also need startups more than this post implies. Workday's network quadrupled in three months, and Klein said SAP will run "an open platform" with agents "not coming from SAP". The strongest counter-case, in my view, is that partner networks are a cheaper source of breadth than in-house teams, so acquisitions stay the exception. One acquisition among 15 launch partners supports that reading as easily as it supports mine.

The source base has gaps of its own

The SAP API policy is described here through DSAG's summary, because the policy document itself could not be read in full for this post. The Workday call quotes come from an automated transcript with visible errors. Forrester's 600-agent count is not in Oracle's own release, which names 13 agents and gives no total.

Frequently asked questions

Where to start

If you run an agent startup, list your 20 largest target accounts this week and mark which suite each runs for finance and for HR. For every Fusion account, name the Oracle agent that overlaps your product. If you cannot explain in one sentence why that customer should pay twice, move the account off your target list.

Then read the API policy of every suite your agent touches, not the partner page. The policy says what a vendor will enforce, and your buyer's security team reads it first.

Related reading