Walk the four stages. No form, no demo call.

Six domains, one schema, one query surface. Invoice truth and execution truth are joined, so the Databricks line shows credits by cluster and by job, priced.

No proprietary schema in the path. What you query is what you can export.

Allocation runs as an ordered pipeline. Every number walks backwards to its rule, and every allocated dollar carries a grade, so a disputed chargeback is settled by evidence rather than seniority.

Pipeline: invoice → environment → project → team. Each stage logs rule, inputs, outputs.

Each finding carries severity, saving, and evidence. Accept one and an AI agent raises a PR against the repository that owns the resource. Your engineers review it, your Terraform applies it. Nothing changes silently.

• Non-prod EKS node groups: apply schedule
  14 node groups · idle 71% of hours -$47K/mo ◔ PR open
• Frontier model on low-complexity refactors
  routing rule · 2,140 runs last month -$31K/mo ◑ Review
• Orphaned persistent volumes: 312 GB
  unattached >30 days across 4 clusters -$8.4K/mo ✓ Merged
• Batch-eligible inference not batched
  50% discount unclaimed · 3 pipelines -$19K/mo ◑ Review

The pipeline never holds write credentials. It opens a PR; your approval applies it.

The board-facing view. Each domain reports its signature metric, and because everything shares one schema they compose into AI-inclusive cost per customer: one formula, one auditable answer.

Most tools report the first bar and let you assume the third. The gap is where credibility is lost.

Every dataset is exposed over the Model Context Protocol. Claude, Copilot, Gemini, or an in-house LLM can query cost and value data conversationally, inside your environment, under the same RBAC as every dashboard.

The assistant inherits the permissions of the person asking, never more.