Zero attribution
Spend exists, but nobody can assign it to an app, team, tenant, workflow, or customer.
The flagship platform
From Northwatch Systems.
Blackridge
The AI Runtime Economics platform.
Every model call your company makes, attributed, explained, and defensible — from a live spend spike to a finance-signed cause file.
Cost Attribution
Who owns every dollar.
Every request is tied to a workflow, team, tenant, app, and user at runtime — not reverse-engineered from a billing export at month-end. Unattributed spend is reported as a finding, not silently dropped.
Economic Forensics
Why the bill moved.
Ten waste classes, from retry amplification to context bloat. Every finding carries evidence rows, an owner, and a dollar value split into observed waste and modeled opportunity.
Retry amplification
A single user action fans out into repeated model calls during soft failures.
Duplicate inference
Similar prompts or semantic duplicates are paid for repeatedly.
Fallback tax
Failures silently route to more expensive models or providers.
Premium-model overuse
Expensive models become defaults for work cheaper models could handle.
Abandoned branch waste
Agent branches call models or tools but never affect the final answer.
Workflow Reconstruction
Replay how the money was spent.
Follow one user action through every agent step, branch, retry, and model call to the final answer — including the branches that spent tokens and never shaped the output. If a workflow can't be reconstructed, spend on it can't be trusted either.
Financial Evidence
Numbers two departments can sign.
Observed vs. modeled, never blended. Confidence and coverage on every claim. Pricing assumptions visible. Any engineer can re-run the investigation from the ledger; any analyst can tie the total to the invoice.
Observed vs. modeled
Observed waste is tied to real events. Modeled opportunity is a labeled estimate engineering should validate before treating it as savings.
Coverage and confidence
Attribution is only as strong as the available metadata. Blackridge reports gaps instead of hiding them behind a chart.
Reproducible
Every finding traces back to evidence rows in the request ledger — any engineer can re-run the investigation.
Architecture
Exact because it's in the path.
Blackridge runs as an inline, pass-through collector: fail-open, metadata-only by default, single-digit-ms overhead, deployable on a single workflow first. Inline position is why the numbers are counted, not sampled — and why policy enforcement is possible later without re-architecture.
Next step
See Blackridge on your data.
Get your AI bill explained — a founding engineer replies within 2 business days.