AI cost controls

What it is

A two-bucket accounting model + per-run guardrails that make AI costs predictable for finance and reviewable for security ops:

• Bundled monthly allowance: your plan includes $X of AI usage every month; drains first.
• Top-up balance: pre-paid credit, used only after the bundled bucket is exhausted. Never auto-charges; you decide when to add.
• Per-run cost estimate: before kicking off, the engine shows a $-figure based on the planned tool calls / token budget.
• Hard ceiling: every run is bounded by a max-spend cap; exceed it and the engine stops and finalises with what it has.

What it produces

• A workspace ledger: bundled remaining, top-up remaining, per-run cost history
• Audit log entries for every AI run with the actual cost charged
• Operator alerts when bundled drops below 25% / 10%

What to use it for

• Setting a sustainable cadence: “validate every critical CVE on every new finding” needs predictable cost
• Budget reviews: hand finance one number (total AI spend last month), pulled from the ledger
• Per-team accountability: each run’s created_by is the user who fired it, so a noisy team is visible in the cost log

Why it matters

The pitch of “AI security tooling” usually ends at the first surprise $5,000 OpenAI bill. We assume you’re going to push usage hard if it works, so the cost surface is first-class from day one, not an afterthought via a billing-portal export.

Why it’s valuable

• Predictable spend. Plan-bundled allowance covers most workloads; top-ups close the gap when you push harder.
• No surprise overruns. Hard per-run ceiling means a runaway loop stops at $X, not at $5,000.
• Cost is shown BEFORE the run. The engine estimates and shows it; you confirm or cancel.

How to use it

• Settings → Usage shows bundled remaining, top-up remaining, per-run cost history.
• Settings → Billing → Add credit to top up.
• Per-run, the “Run AI exploit” dialog shows the estimate; you confirm before any tokens are spent.

For platform admins (Umbra staff): cross-org cost rollup is at /api/admin/costs, showing every org’s bundled vs top-up vs total spend.