Put a ceiling on AI spend — per user, per agent, per tool.

Put a ceiling on AI spend per user, per agent, or per tool — then run the same Audit → Assist → Enforce ladder as Policy & Enforcement. Today we only observe; this names exactly who has no ceiling and who is already past one.

Bottom line

Cap coverage is 78% (14/18). 3 entities are already over cap (svc-data@qpoint.io, support-copilot, gpt-4o), 2 near it. $218.50 flows through 4 uncapped entities — starting with data-pipeline-agent (the LLM10 unbounded finding). We're in Audit: nothing is blocked yet. Set ceilings on the uncapped first, then stage Assist before Enforce.

Do nextSet ceilings on 4 uncapped entitiesEnforce caps in Policy & Enforcement →See full spend in Cost Center →

How a breached cap is handled

Audit live · Assist/Enforce projected

Audit

Observe consumption against each cap. Surface who has no ceiling and who is over — no intervention.

live now

Assist

Alert the owner at 80% of cap and require approval to keep spending past it.

projected

Enforce

Hard-stop the run when the cap is hit. The spend simply cannot exceed the ceiling.

projected

78%

cap coverage

14 of 18 entities have a ceiling

over cap

already past a ceiling this period

near cap

≥ 80% — Assist would alert

$218.50

uncapped spend

4 entities with no ceiling at all

Limits by User

A monthly ceiling per human or service identity — who is spending.

UserSpend vs. capUsedCapStatus

dev-7f3@duck.com est

no ceiling — runs unbounded

$78.00noneuncapped

(proxy-only, no identity)

no ceiling — runs unbounded

$44.10noneuncapped

svc-data@qpoint.io

$52.30$40.00131% over

mark@qpoint.io

$245.50$300.0082%

svc-support@qpoint.io

$168.40$250.0067%

jane@qpoint.io

$96.20$200.0048%

Bar fills toward the dashed cap line (100%). Rows past it are over; rows with no ceiling are uncapped — an enforcing policy can't stop what has no limit. At the recent ~$115.57/day fleet run-rate, the near-cap rows breach within 2 days.

No ceiling — set a cap first

2 uncapped · $218.50

These actors spend with no limit. data-pipeline-agent is the LLM10 unbounded-consumption finding by name — it burned 9.2M tokens with no per-agent ceiling. Capping here is the highest-leverage move.

dev-7f3@duck.com

$78.00 spent · proposed cap $100.00/mo

(proxy-only, no identity)

$44.10 spent · proposed cap $75.00/mo

Enforce caps in Policy & Enforcement

projected

A spend ceiling is one policy on the same Audit → Assist → Enforce ladder. Flip to Enforce there to hard-stop a run when its cap is hit — paired with the tool allow-list and approval gates.

Fields: usageCaps (policy config) joined to costs.by_user / by_agent / by_model · usageLimits() · usageLimitSummary() — from app/data/c16-fleet.ts. Caps are projected policy; consumption is live. Breach dates computed from the daily run-rate.