How to Forecast SAP AI Units Before Joule Free Windows End

To forecast your SAP AI Units bill, stop counting prompts and start counting agent runs. An autonomous Joule agent draws several metered actions per run and fires on a schedule rather than on demand, so the only reliable forecast is built from measured runs per week, actions per run, and the overage rate in your order form. The free windows SAP is closing between now and year-end are the last chance to measure that baseline at zero cost.

Why September 2026 is the month the meter starts

Two SAP promotions expire within months of each other. According to SAP's ABAP platform team on the SAP Community blog, the free promotion for Joule for Developers ABAP AI capabilities ends in September 2026, after which those capabilities are priced purely on consumption in AI Units. Futurum Group's Keith Kirkpatrick, covering SAP's August 18 interview with its head of ABAP platform, put the open question plainly: SAP still has to show that AI Units deliver predictable costs in practice.

On the procurement side, SAP confirmed that Joule Agent Runtime is free through December 31, 2026 for current Ariba and Fieldglass customers, as reported by ERP Today and SAPinsider in June. The SAP Sapphire Innovation News Guide lists six more assistants planned for general availability this month: Category Management, Supplier Management, Requisition, Buying, Receiving, and Services Procurement. New agents are arriving at the exact moment the runtime is free and unmetered; the January 2027 invoice is the first time most customers will see what they built.

This is not a criticism of SAP. A free window is a legitimate way to get customers into production, and Futurum notes SAP management expects consumption and value-based models to drive more cloud growth than seat expansion by 2030. It means the arithmetic has to be done by you, now, while the meter reads zero.

What an AI Unit actually meters

SAP prices Joule agents in AI Units, a prepaid consumption currency drawn down per action. Business AI Base, which includes interactive Joule, is bundled into RISE and S/4HANA Cloud; Business AI Premium, which covers autonomous agents, draws AI Units. Redress Compliance, a buyer-side licensing advisory, published observations (updated August 2026) from roughly 35 to 45 RISE and S/4HANA Cloud renewals it benchmarked: about 200 AI actions bundled per Advanced Full User Equivalent, five to ten actions drawn per multi-step agent run, and observed overage of $0.08 to $0.18 per action. Those are figures from quotes they reviewed, not SAP list prices; your order form governs.

Restated in the unit you consume, a 200-action allowance is 20 to 40 agent runs per Advanced FUE. An interactive prompt is one turn that answers and stops. An agent reads context, plans, calls three or four systems, validates, and writes a result, and each step is metered. Then it runs again tonight, whether or not anyone is at a desk.

How to build the forecast in the free window

Instrument the agents now, while their cost is zero. The baseline has four numbers per agent: runs per week, actions per run, the share of runs that write to SAP, and the growth curve as the business finds new uses. If your tenant does not expose per-agent action counts, log them at the integration layer: every call to a BAPI, an OData service, or a Joule skill is an action you can count.

Then convert the roadmap. For each planned automation, estimate steady-state runs per period, multiply by actions per run, subtract the pooled allowance across your Advanced FUE count, and price the remainder at the overage rate in your paper. Do it per agent, not per tenant, because the balance is consumed tenant-wide and an unowned pool is drained by whichever team automates first. Redress's finding that seven in ten estates it modeled had a bundled Base tier masking an unbudgeted Premium run rate is the failure mode to design against.

Finally, negotiate three separate terms: the bundled action allowance, the AI Unit price, and the overage rate, with a consumption cap in the order form. Redress notes use-based pricing became the default at SAP cloud renewals in July 2026, so this belongs in the current negotiation.

Where a self-hosted agent changes the cost shape

A self-hosted agent does not make the cost disappear; it changes what you are metered on and who can see the meter. SayfeAI runs its agents on Amazon Bedrock inside the customer's own AWS account, so the marginal cost of a run is model tokens plus compute, itemized on an AWS bill the customer already reads, with model choice under their control. Across our production customers, SayfeAI's aggregate figures are 98,989-plus orders at 95 percent touchless and 99.2 percent accuracy, and per-order inference cost is a line item each customer can inspect and tune.

That is the honest comparison: not free versus paid, but a per-action meter inside a vendor tenant versus a per-token meter inside your own cloud account. For high-volume, scheduled work such as order intake, the second shape is easier to forecast because the unit of measure is visible from day one. For occasional interactive use inside SAP screens, bundled Joule Base needs no model at all.

Frequently asked questions

Is SAP Joule still free?

Interactive Joule under Business AI Base remains bundled with RISE and S/4HANA Cloud. The free promotion for Joule for Developers ABAP AI capabilities ends in September 2026 per SAP's ABAP team, and SAP has said Joule Agent Runtime for Ariba and Fieldglass is free through December 31, 2026. After those dates, agentic workloads draw AI Units under Business AI Premium.

How many AI Units does a Joule agent consume per run?

SAP's consumption guidance is published per capability; it cannot tell you what your specific agent draws per run. Redress Compliance's observation across the renewals it benchmarked is five to ten metered actions per autonomous run, versus one for an interactive prompt. Your agents will vary with how many systems they call and how many validation steps they run, which is why measuring during the free window beats any published average.

Does self-hosting on Bedrock avoid consumption pricing?

No. Bedrock is metered per token and per compute hour, so the bill still scales with usage. What changes is that the meter sits in your AWS account, itemized per model, with no vendor-set action allowance or overage rate to negotiate. SayfeAI, the self-hosted agentic AI platform at sayfe.ai, is a separate company from Sayfe.ai (sayfeai.com), an OpenAI partner that deploys ChatGPT Business for small businesses.