The token bill moved to finance. Most SAP budgets still book it under IT.

SAP's chief controlling officer does not usually write for the News Center. On 8 September, Lukas Deutsch did, with David Imbert of SAP Financial Management, about a cost category SAP's own finance team was not built to handle: AI tokens. There is no product in the piece and nothing to switch on. It is still worth a CFO's twenty minutes, because it describes an operating-model change that lands in the 2027 budget round of every company running SAP.

The short version: AI consumption stops being a platform fee somebody in IT signed and becomes a resource finance forecasts, allocates, and scores. The longer version, for an SAP landscape, is that the meter is not in one place. It is in three, and only one of them says "token".

What shipped

A practice piece, not a release. Deutsch and Imbert describe what SAP learned managing its own generative AI bill as consumption moved from chat into code assistants and agents. Four lessons carry the article.

First, a single line item hides everything useful. Total spend does not show which teams, workloads, or usage patterns drive it, or whether the consumption produces anything. SAP built that visibility across commercial, engineering, finance, and product, and the forecast improved with each cycle as operational detail was layered in.

Second, consumption needs an owner. A central AI budget makes experimentation easy and disconnects the people spending tokens from any accountability for them. SAP now allocates token cost to business areas. The question moved from "how much are we spending" to "what are we getting for this".

Third, cost per token is the wrong scorecard. After rolling out AI developer tools, SAP recorded a mid-double-digit percentage increase in pull-request merge rates. That consumption was worth paying for. Finance needs cost and value side by side, or governance optimises the wrong thing.

Fourth, guardrails should hit waste, not adoption. Three root causes of disproportionate spend showed up in SAP's data: concentration in power users and automated agents, model misalignment, and tool proliferation. Three levers answered them: token capping, model routing, and tool rationalisation. The article says this contained a triple-digit-million-dollar financial risk while adoption kept moving.

What the article does not do: name an SAP meter, a dashboard, or a product. Nothing here is generally available because nothing here is software. Read it as SAP the customer, not SAP the vendor.

Background

For most of the SaaS decade, AI cost had no line of its own. It was inside a subscription, or a seat, or an innovation budget under the CIO. Usage was bounded by how many people opened a chat window.

Two things broke that. Coding assistants made engineers the heaviest consumers in the company. Then agents arrived, and an agent's loop does not wait for a human to type. Consumption now scales with process volume, not headcount, which is the one thing a seat-based budget cannot model.

SAP's CFO had already said the loud part. In a CNBC interview after Q2 results in July, Dominik Asam called enterprise token spending "going through the roof", noted that most of it still goes to chatbots and coding tools, and said SAP runs 58 large language models on its platform so it can play performance against cost. His advice: pick the cheapest model that is reliable enough, and where the job is deterministic, use no model at all. The September article is the same finance organisation explaining how it operationalised that advice.

It matters more for SAP customers than for a generic enterprise for one reason: SAP's own commercial model already prices AI by consumption. Deutsch and Imbert are describing, from the inside, the discipline SAP asks its customers to build on the outside.

How it works

Where does a token turn into an invoice for a company running SAP? Three places, with three different units.

**Prebuilt SAP AI: AI Units.** SAP sells Premium AI through AI Units, a virtual currency bought up front on SKU 8019164 and expiring after twelve months if unused. Base AI is included in the cloud subscription; SAP Cloud ERP 2608 put natural-language search, form filling, and summaries in the standard license. Premium AI is metered: on the pricing page, Document Grounding costs 0.005 AI Units per record, and SAP's training material prices every prebuilt agent in 2026 at a single 0.02 AI Units per action. A customer never sees a token here. The forecasting unit is agent actions, which means process volume: invoices, orders, tickets. That same training material dated a move of most non-agent generative AI into Base AI for Q3 2026. Whether that has landed in a given contract is a check, not an assumption.

**Customer-built agents: Joule Studio runtime.** Design time in Joule Studio is free under fair use. Production runtime for customer-built agents, apps, and workflows is paid on resource consumption, and the training material lists the meters as compute, storage, transactions, and AI tokens, settled through the SAP BTP Enterprise Agreement, CPEA, or pay-as-you-go. This is the one place in the SAP model where "token" is literal. SAP has described custom-agent runtime as free through 2026 as a promotion. If that holds, 2027 is the first year a Joule Studio agent shows a real runtime cost, and the first year finance needs a forecast for it.

**Outside SAP: raw tokens.** Developer tools, hyperscaler model endpoints, and third-party agents that reach into SAP data through MCP-style connectors bill in tokens directly. Here architecture is a cost lever finance cannot read off the invoice. SAP's Anirban Majumdar gave the illustration in VentureBeat in May: a query for an employee's manager and peers in SAP SuccessFactors consumed 565,000 tokens under a standard MCP implementation and 80,000 under a context-aware one. Roughly $1.70 against $0.24, per operation, repeated across thousands of daily calls. Same question, seven times the bill, decided by an integration pattern.

Deutsch and Imbert's three levers map onto those meters. Model routing is what a Generative AI Hub with 58 models exists to do. Tool rationalisation is the job SAP describes for SAP AI Agent Hub. Token capping is a finance control on runtime and external endpoints.

Why it matters for SAP leaders

Three meters, three owners. AI Units, Joule Studio runtime, and external token bills sit in different contracts and often different cost centres. Decide who forecasts each before the 2027 planning cycle, or the CIO forecasts all three by default and owns none of the value.
Forecast prebuilt AI on process volume. An agent priced per action scales with the number of invoices it clears, not the number of people who can see it. That is a controlling model finance already knows how to run.
Treat 2026 as the free year it is. Runtime promotions end. A pilot that costs nothing to run today has no baseline for next year. Build the baseline now, while the meter reads zero.
Put architecture on the finance agenda. A sevenfold difference in tokens for one query is not a technical footnote. Integration pattern is a cost decision, and an SAP API Policy question as well.
Do not let finance become the brake. SAP's own numbers show the point: the developer-tool spend that looked alarming on the invoice paid for itself in merge rate. Governance that only sees the cost column kills the wrong things.

Near term

This quarter, the question is ownership. Ask which cost centre carries AI Units today, whether AI Unit consumption is reported by business area or as one platform line, and who signs for Joule Studio runtime once the promotional period ends. If all three answers are "IT", the operating model Deutsch and Imbert describe has not started.

Longer term

The article ends with the work still to do at SAP: embed token practice in planning and reporting, sharpen allocation, and forecast cost before it arrives. For customers, the equivalent is a controlling model in which AI actions sit next to the process they serve, the way freight sits next to deliveries. As Autonomous Domain Blueprints and value-based Industry AI pricing arrive, the unit of purchase keeps moving toward outcomes. Finance teams that can already tie consumption to a process negotiate those contracts from knowledge.

Open questions

Which finance owner should forecast this, and against which SAP meter, is still an open question. The article does not say whether SAP's internal allocation framework will surface as anything a customer can use, such as AI Unit consumption by business area inside SAP for Me or the BTP cockpit. Whether the Q3 2026 Base AI move has reached contracts, and whether the runtime promotion ends on 31 December, are dates to confirm with the account team rather than infer from training slides.

What to do this week

Ask the controlling lead one question: if a Joule agent cleared ten thousand more invoices next quarter, which line in the plan would move, and who would have predicted it?

If this kind of decoding is useful, follow along.