Now accepting guest editorial from SAP, Oracle, Workday and ServiceNow practitioners.Pitch a guide
The Sign-Off Table

Lesson 6 of 6

Estimating AI Unit consumption

8 min

Listen to this lesson

0:00 / 6:32

Since the free promotion ended on 30 September 2026, every Joule for Developers request that uses AI has a cost. Finance will ask how many AI Units a team needs before anyone signs a purchase order. SAP's sources don't agree on the metric: SAP's 23 September 2026 post calls the license per user per month (PUPM), while the Discovery Center AI Estimator meters it at 0.007 AI Units per AI request. Ask your account team which metric your contract uses. Either way, the honest approach is a measured estimate: pilot, measure, extrapolate, and review monthly.

How AI Units work

  • AI Units are a virtual currency for premium AI capabilities across SAP cloud services, drawn from a central pool rather than tied to one product.
  • They are bought annually and expire after 12 months if unused.
  • Usage and remaining balance are visible in real time in SAP for Me.
  • SAP says it provides guidance and estimation tools, and that your account executive or partner can help model usage scenarios.
  • J4D requests are executed and metered in the ABAP AI Hub service on SAP's cloud.

What draws down AI Units

Not every action in the IDE consumes AI. SAP's MCP tool reference marks which ADT MCP tools need Joule tokens: object explanation, AI-based ATC fixes and their results, and the ABAP documentation search are marked "Required"; activation, search, object checks, ABAP Unit runs, ATC runs, deterministic quick fixes and transport tools are not. In the IDE, the AI capabilities (chat, Explain, predictive completion, test generation, the wizards) are the consumers.

Rare, largeOData UI Service from Scratch, star schema generation: few runs, big outputs
Frequent, largeAgent sessions with many AI fixes or explanations: watch these closely
Rare, smallOccasional Explain on a legacy object
Frequent, smallPredictive code completion while typing: many small requests all day
Consumption comes from two directions: many small requests and a few large ones. Measure both.
The Sign-Off Table · thesignofftable.com/academy/sap-aid-201

A four-step estimate

Estimate from your own measured usage, not from a guess about how much people will use AI.
The Sign-Off Table · thesignofftable.com/academy/sap-aid-201
  1. Pilot with a mix of roles. Contoso's pilot uses six developers: two RAP builders, two maintaining converted classic code, one on custom code findings, one test-focused.
  2. Measure weekly in SAP for Me. Record units used per developer and what they were doing that week.
  3. Normalize. Compute units per developer per week for each work pattern; note spikes (a wizard run, an agent session) separately from the baseline.
  4. Extrapolate with a buffer. Multiply by headcount per pattern and weeks per year, add a buffer for growth and agents, and remember unused units expire after 12 months.
Work pattern (Contoso)DevelopersMain J4D useWhat to watch
RAP builders4Completion, wizard, business logic predictionWizard runs during new app starts
Classic maintenance8Explain, completionExplain on large programs
Custom code findings2ATC explanations and AI fixesAgent sessions applying AI fixes in bulk
Test-focused2ABAP Unit and CDS test generationRepeated regeneration of the same tests

Illustrative team split for the estimate; headcounts are Contoso planning figures, not SAP data.

Keeping consumption in check

  • Assign J4D only to named developers who need it; the Identity Services group is your on/off switch.
  • Teach "start small": generating tests for one to three methods uses less than regenerating a whole class repeatedly.
  • Watch agent sessions. An agent that loops on AI fixes or explanations can use far more than a person typing.
  • Review usage monthly with the development lead, and compare it with the value delivered: tests written, findings closed, apps built.
  • Plan the next order from the trend, remembering that unused units expire after 12 months.

What would you do?

After the pilot, Contoso's usage shows one developer used five times the units of anyone else in week 3. Grace's PMO asks Kenji whether to cap that developer. What should Kenji do first?

Key points
  • AI Units are bought annually, expire after 12 months and are tracked in SAP for Me.
  • AI capabilities and some MCP tools (explain, AI fixes, docs search) consume units; activation, ABAP Unit and ATC runs don't.
  • Estimate from a measured pilot across work patterns, with a buffer.
  • Investigate spikes before acting on them.
  • Get unit prices and the metric (PUPM or per AI request) from your SAP contract, not from public sources.

Check your understanding · not graded

Lesson check

  1. 1. Which ADT MCP tool does SAP mark as requiring Joule tokens?

  2. 2. Where do customers track remaining AI Units?

  3. 3. What is the most reliable basis for an annual AI Unit estimate?