AI Units Governance: Consumption Estimates for Every Generative AI Feature

What changed

Oracle now publishes estimated AI Unit consumption rates for each generative AI feature. AI Units are provisioned at the account level and consumed on every invocation. Usage is visible at Setup > Company > View Billing Information. These are estimates — actual consumption varies with request complexity.

Consumption estimates by feature

• Ask Oracle — 10 units (simple questions), 10–50 units (analysis queries), 50–200 units (research/aggregation queries like top-N A/R balances).
• Case Summary — 5–75 units per case summarization.
• Narrative Insights — 5–200 units. Triggered via the Generate Insight button on supported records (e.g., Customer). Summarizes account status, open items, transactions, and billing issues.
• Prompt Studio — 5–10 units per prompt-action test. Prompt Studio lets admins author and test reusable prompt templates for common writing tasks.
• N/llm Module (SuiteScript 2.1) — 5–25 units per llm.generateText()-style call. This is the programmatic entry point; every SuiteScript invocation that sends a prompt to an LLM and receives a response consumes units. See llm.ModelFamily for supported model enumerations.
• Text Enhance — 5–10 units per rewrite, translation, or summarization request.
• Transaction Matching Assistant — 5–10 units per match recommendation. Available only to eligible accounts with supported AI features enabled. Uses rule-based candidate generation plus LLM-based ranking and explanation.

What to do

1. Audit N/llm usage in custom scripts. Each N/llm call costs 5–25 AI Units. If you have scheduled scripts, map/reduce jobs, or user-event scripts that invoke the module at scale, estimate monthly unit burn: (invocations/month) × (estimated units per call). This is a real cost lever — treat it like governance units.
2. Check your allocation. Navigate to Setup > Company > View Billing Information to see current consumption and remaining AI Units. There is no documented API to query this programmatically yet.
3. Gate expensive features. Ask Oracle research queries (50–200 units) and Narrative Insights (up to 200 units) are the most expensive. Consider whether role-level access restrictions are appropriate to prevent accidental overconsumption.
4. Design prompts efficiently. For N/llm Module usage, shorter and more targeted prompts will consume fewer units. Avoid sending entire record bodies when only a few fields are needed.
5. Monitor for billing surprises. Oracle states these are estimates that may vary. Until consumption analytics are more granular, periodically review the Billing Information page, especially after deploying new AI-powered SuiteScripts or enabling features for additional roles.

What Oracle left vague

• No hard unit caps or throttling behavior is documented. It is unclear what happens when AI Units are exhausted — whether calls fail, queue, or incur overage charges.
• No per-script or per-feature usage breakdown is described beyond the aggregate Billing Information page.
• The llm.ModelFamily enum and whether different model families consume units at different rates is not addressed here.
• Transaction Matching Assistant eligibility criteria ("eligible customers with access to supported AI features") are not specified.