The method, in the open

Judge the method before you judge me.

No gated PDF, no email wall. The checklist, the formula, and the audit approach are here — the same ones I use on every engagement.

The 12-leak checklist

The places AI spend quietly leaks. Work top to bottom; most teams have three or four live at once.
  1. 01

    Spend not located

    Provider invoices, card sweeps, and shadow-AI subscriptions that never reach one ledger.

  2. 02

    Workload inventory missing

    Models used, tokens in/out, and latency needs per feature — guessed, not measured.

  3. 03

    No caching coverage

    System prompts, RAG contexts, and repeated prefixes re-sent every call at full price.

  4. 04

    Batch eligibility ignored

    Non-real-time jobs left on sync endpoints — a straight discount unclaimed (vendor-documented ~50% off on deferrable workloads).

  5. 05

    Model-fit mismatch

    Flagship models doing Haiku-grade work because nobody set a routing rule.

  6. 06

    RAG inefficiency

    Chunk sizes, retrieved-context bloat, and redundant re-embedding on every query.

  7. 07

    Agent retry storms

    Retry loops, runaway iterations, and missing max-token caps inflating every run.

  8. 08

    Failover double-billing

    Fallback paths that both fire, or primary and backup billed for the same request.

  9. 09

    Commitment under-use

    Reserved capacity or committed tiers sitting idle against bursty real usage.

  10. 10

    No monitoring

    Spend that drifts for weeks before anyone notices because nothing watches the trace.

  11. 11

    Tooling overlap

    Two observability platforms doing the same job, or a dashboard nobody reads.

  12. 12

    Features that shouldn't exist

    AI features whose unit economics are negative — the best fix is sometimes deletion.

The ROI formula

The ROI formula, stated plainly

Net annual value = (measured annual savings + labor hours saved × loaded rate) − annual fees. Every input is labelled measured or estimated, and any estimate carries ±20% sensitivity. The confidence factor (1.0 measured, 0.8 vendor-doc rate on your logs, 0.5 benchmark-only) is an internal judgment heuristic until calibrated against real audits — it is not a promise.

Proof by method

Want this run on your own stack?

The Snapshot applies the same checklist to your bills. Book one and you'll see the leaks named, not theorized.