Budget policies

NuBench treats AI spend as a governed resource. Budgets are defined in tokens, scoped to the organisation, a matter (cost code), project, or an individual user, and evaluated by the governance pipeline before a request reaches any model.

Hard-block enforcement is live: when a request would exceed a configured hard cap at any applicable scope, the pipeline refuses it before dispatch. Warning and critical alert thresholds remain visible in the consoles without blocking.

The budget model

Each budget allocation has four states, evaluated on every request:

| State | Meaning | Behaviour at target | | --- | --- | --- | | Allow | Comfortably within budget | Request proceeds normally | | Advisory | Approaching the configured threshold | Request proceeds; users see a notice | | Warning | Close to the limit | Request proceeds; users and admins see a prominent warning | | Blocked | Hard cap would be exceeded | Request is refused before the model is called |

The design goal is overage is impossible, not just visible: the pre-check and deduction happen in-line as part of the pipeline, not in a nightly reconciliation job. If the budget counter store is unavailable, requests are refused (fail-closed) rather than allowed through.

Scoping

• Organisation hard cap — token ceiling for the org period (Billing → Hard cap). Warning/Critical percentages are measured against this limit.
• Matter (cost code) hard caps — optional monthly and/or daily token caps on a cost code. Empty means no matter-level cap; the org cap still applies.
• Project budgets — optional project sub-allocations enforced alongside the org envelope.
• User budgets — optional per-person allocations when configured in the allocation store (no dedicated admin UI for user caps in this release).

Client-level and standalone team-level hard caps are not available yet; matter spend is attributed via cost codes (filterable by client in Cost Allocation).

The narrowest applicable blocking scope wins. A request under the org cap but over a cost-code monthly cap is refused with a cost-code scope error.

Configuring budgets

From the admin console:

1. Billing — set the organisation hard cap (tokens) and Warning/Critical alert percentages.
2. Cost Codes — optionally set monthly and/or daily hard caps per matter.
3. Optionally configure project token budgets when creating or editing a project.

Budget positions and burn-down are visible in Billing and Usage. Every hard block is recorded in the audit trail as gateway.budget.blocked.