How it works
From read access to a
decision, in one week
Nothing gets installed in your tenant or your workspaces. No agents, no deployed resources, no write permissions at any point in the engagement.
Step 01
Grant read access
Four grants, all read-only, all revocable by you. Azure covers the infrastructure and Foundry token billing; Databricks and Fabric each need one additional read scope because their consumption data does not surface in Azure Cost Management at useful granularity.
| Platform | Grant | Why it’s needed |
|---|---|---|
| Azure | Reader | Resource inventory, configuration and metrics across the subscriptions in scopeNo write, no delete, no action operations |
| Azure | Cost Management Reader | Usage, rates and commitment detail, including Foundry token and PTU billingRequired to price findings at your contracted rates rather than list price |
| Databricks | system.billing / system.compute | DBU consumption by job, cluster, warehouse and user via Unity Catalog system tablesRead grant only. We do not read your data tables, notebooks or model artifacts |
| Fabric | Capacity metrics read | CU utilization by workload, throttling events and smoothing behaviour per capacityRead via the Capacity Metrics app or admin API |
Access is revocable at any moment from your side, and we ask you to revoke it once the walkthrough call is done unless you are continuing on a subscription.
Step 02
The audit runs
A 90-day window of inventory, token telemetry, DBU consumption, CU utilization and billing detail. Findings that survive are resolved down to named objects and priced against the rates you actually pay.
| Phase | What happens | Typical |
|---|---|---|
| Collection | Azure inventory and usage, Foundry token and PTU telemetry, Databricks system tables, Fabric capacity metrics | 6–16 hrs |
| Analysis | All 118 checks run at the current version; findings deduplicated and scored for riskA deployment flagged by three checks is one decision, not three | 1–2 days |
| Review | A practitioner reads the output before you do and pulls anything the model got wrongThis is the step tools skip. It is why the list survives contact with your engineers | 1 day |
| Assembly | Workbook and executive summary produced, quality-checked and stamped with the check-set version | 1 day |
Step 03
You work the list
Two artifacts land at the same time, aimed at two different rooms.
The audit workbook
A multi-tab spreadsheet, one tab per finding domain. Every row is a named object — a Foundry deployment, a Databricks job, a Fabric capacity, an Azure resource — with the check that flagged it, the evidence, the recommended change and its annualized value.
The executive summary
Three pages. Where the money is, what it costs to go get it, the risk of each move, and a read on the practices that let the spend accumulate unattributed in the first place.
Then a call. An hour with your engineers to walk the list, hear why four of the findings are wrong, and mark what is actually going to get done.
After the first run
The second audit is where it compounds
AI spend regenerates faster than anything else on your bill — a new model, a new agent, a prompt change, and last quarter’s answer is stale. Because the check set is versioned, a later run tells you what got fixed, what came back, what is new in your estate, and what is new because Microsoft or Databricks changed the rules underneath you.
Start with one workspace
You do not have to open the whole estate to find out whether this is worth doing.