AI Allowance
What It Models
An organization’s AI allowance combines monthly accounting with staff-controlled limits and temporary capacity. Requests hold credits during execution and replace those holds with charges. Corrections and threshold notices retain separate records so staff can explain subsequent changes.
Identity and Attributes
The organization and local month identify a period. Configured limits have one versioned row per organization; absence selects the defaults. Grants and corrections have generated identifiers. A crossing is unique for its month and threshold and retains the figures used to decide it. Grant state and remaining capacity are calculated when read.
Relationships
Requests reserve and settle against a period. Limits and exemptions refer to catalogue operations. Usage reads allowance data to present current status. Security retains the audit events and resolves Administrator role holders for notices.
Invariants
Admission checks capacity and the running-request cap in one conditional update. Settlements floor remaining holds and running counts at zero. Grant validation checks window length, kind and overlapping capacity under the platform ceiling. Changed limits require the inspected version, while identical existing values return unchanged. Only charged usage creates threshold crossings.
How It Is Stored and Shared
The LLM service owns periods, configuration, grants, corrections and crossings. Period accounting follows organization row policy. Limits and grants use explicit organization predicates under documented row-policy exemptions. Customer summaries expose credits, percentages and reset dates; staff views also expose control and monetary details.
Open Design Questions
Admission may refuse an estimate before usage reaches the dashboard’s exhaustion threshold. Corrections and limit reductions do not trigger notices. Repeated grants or corrections create additional records, and refunds against a missing current period bypass the existing-balance check. The interim Administrator audience still needs its planned responsibility-based replacement.