Behind every thirty-second agent demo sit three separate meters — model tokens, SAP AI Units, and the human who has to approve what the agent posts. The one nobody prices is the one that decides the business case.
Every SAP AI agent demo ends the same way: a document gets read, a decision gets proposed, a posting happens, applause. What the demo never shows is the invoice. Behind that clip sit three independently metered cost pools — LLM tokens, SAP’s AI Unit licence layer, and the human reviewer who signs off before anything hits the ledger — and every vendor, SAP included, has an incentive to let you evaluate only the first.
The wrong question
The question sponsors keep asking is “what does the agent cost per token?” That is the wrong question, because tokens are the smallest and most stable of the three meters. In July 2025 SAP restructured its Business AI commercial model, replacing fragmented BTP credits and feature entitlements with AI Units as a universal consumption currency across Joule, AI Foundation, and embedded AI. That moved AI spend from a fixed line inside your RISE or GROW subscription to a variable, SAP-controlled utility bill — and it is the second meter, not the first, that most finance teams have never modelled.
Meter one: tokens, priced asymmetrically
SAP AI Core and the Generative AI Hub meter LLM usage in tokens, converted into capacity units by model-specific factors SAP publishes per model. Input and output tokens are priced differently — output costs more, because generation is heavier than ingestion. The asymmetry matters in practice: a summarisation task is input-heavy (lots of context, short answer), while a drafting task is output-heavy (short prompt, long generation), so two things that look like “one task” in the demo can cost very different amounts. This is the meter everyone models in the pilot, because it is the only one with a transparent formula. In dollars, it is usually the smallest of the three.
Meter two: the AI Unit licence layer
AI Units are sold in blocks of 100 per year, purchased annually, and expire at the end of the contract year with no rollover by default. List pricing sits around $0.40 per unit pay-as-you-go, with committed-volume tiers pulling the effective rate toward roughly $0.24 — but SAP publishes no standard price list, so per-unit cost, allocation size, and overage terms are negotiated contract by contract. That opacity is structural. When you run short, SAP bills the excess under an “Excess Use” clause priced at roughly list minus a 30% discount — a meaningfully above your committed-tier rate, which is why licensing analysts describe overage as costing several times the contracted unit price. A cap is achievable, but only if you ask for it before signing.
Two mechanics inside this meter matter more than the headline rate:
- Document Grounding is metered per record, where one record is up to 50 MB of uploaded content and consumes 0.005 AI Units regardless of how many tokens later queries burn. Ingesting a mid-size policy or contract library can drain units before the agent answers a single question.
- Agents are priced by step, not by request. A step is any discrete unit of work — a planning move, a tool call, a reasoning cycle — and the per-step cost rises with the agent’s category (Basic, Standard, Advanced). This is exactly why a “simple query” and an “agentic task” that look identical on stage differ by an order of magnitude in consumption: the query is one call; the agent is a chain of billed steps, each drawing from the same finite, expiring pool.
Meter three: the one nobody prices
Here is IOTEK’s position: none of the above decides whether an agent is actually cheaper than the job it replaces. The deciding meter is human-in-the-loop review — the labour of a controller, AP clerk, or compliance officer checking what the agent proposes before it posts to the ledger, the vendor master, or payroll.
The math is unforgiving and linear in a way token pricing is not. A simple model makes the point: review a quarter of an agent’s runs for three minutes each at a loaded reviewer cost of $65/hour, and you add about $0.81 to every run — turning a $0.25 pre-review workflow into $1.06 all-in, a 4x markup owed entirely to the human gate. Set that against a manual invoice-review baseline of roughly $8–$13 per document and an agent can still win — but the margin compresses fast as the review rate climbs, and it flips negative when governance (reasonably) demands 100% sign-off on anything touching GL postings, vendor bank details, or payroll. Practitioners building these business cases routinely understate true cost, almost always because the review-labour line is missing from the model entirely.
Why the demo hides all three
| Meter | What it prices | Why it’s easy to miss |
|---|---|---|
| LLM tokens | Input/output tokens ₒ capacity units | Transparent formula, smallest dollar impact — easy to over-focus on |
| AI Units / licence | Per-step agent cost by category; Document Grounding per record; annual expiry | No public price list; opaque per-operation rates inside bundled tiers |
| Human review | Reviewer time on every output requiring sign-off | Not billed by SAP at all — it lives in your headcount budget, invisible to the AI cost dashboard |
A demo is, by construction, a zero-review, single-request environment: one query, one model call, one clean answer, no governance gate. Production is the opposite — chained steps, grounded documents, and a compliance function that will not let an agent post a journal entry unsupervised. RISE and GROW customers who signed before July 2025 are meeting this at renewal, when previously bundled AI access reappears as a metered AI Unit add-on they never separately negotiated.
Practitioner takeaway: how to price a pilot honestly
Don’t price a pilot on token cost, and don’t treat SAP’s bundled AI Unit allocation as a proxy for production cost. Instead:
- Model steps, not requests. Get the Basic/Standard/Advanced step-cost breakdown for the specific agent before you commit budget — category alone can multiply consumption several times over.
- Price Document Grounding as its own line — every ~50 MB of grounded content at 0.005 AI Units per record, not an afterthought inside the token estimate.
- Build the review-labour line first, not last. Use your organisation’s actual required sign-off rate for the task (not the vendor’s assumed rate), apply your loaded reviewer cost, and add it before comparing against the fully manual baseline.
- Negotiate AI Unit terms before you scale — allocation sized to a real 12-month consumption model plus buffer, an overage cap, and a rollover provision; none are offered by default.
- Only call it a win if the fully loaded number — tokens plus AI Units plus review labour — beats the job it replaces. If it takes zero-review assumptions to clear the bar, the pilot hasn’t proven anything yet.
The demo will always look free. The renewal invoice — and the review queue behind it — are where the real P&L for agentic SAP AI gets written.
### Sources – SAP — **Software Packages and Pricing | SAP Business AI: https://www.sap.com/products/artificial-intelligence/pricing.html – SAP Licensing Experts — SAP AI Units Explained: https://saplicensingexperts.com/blog/sap-ai-units-explained – Redress Compliance — SAP Joule and AI Units: 2026 Pricing and Levers: https://redresscompliance.com/sap-joule-ai-units-licensing-pillar-2026 – SAP Help — Metering and Pricing for Generative AI (SAP AI Core): https://help.sap.com/docs/sap-ai-core/sap-ai-core-service-guide/metering-and-pricing-for-generative-ai – SAP — Document Grounding: https://www.sap.com/products/artificial-intelligence/document-grounding.html – SAP Community — Document Grounding: a (hidden) gem in SAP Business AI (per-record / 50 MB): https://community.sap.com/t5/technology-blog-posts-by-sap/document-grounding-a-hidden-gem-in-sap-business-ai-s-portfolio-for-smaller/ba-p/14232864 – JNC — SAP Business AI 2026: Usage, Units, and Reality*: https://jncuk.com/sap-business-ai-2026-usage-units-reality/