Leak 01
Repeated context
The same system prompt, retrieved context, and few-shot examples are re-sent on every call — billed at full price, every time.
The AI Cost Auditor
Your AI bill is bigger than it needs to be. I find the waste in two weeks.
Fixed-fee AI cost audits for teams spending ~$5k+/month on models and APIs. You get a prioritized savings list with the ROI math — then help implementing it.
The method
No dashboards-for-dashboards. No vendor kickbacks.
The method follows the money through your stack: prompts and context, model choices, RAG retrieval, agent call loops, batching, caching, and duplicated workflows. I read what your bill is actually doing — not what a dashboard assumes.
Three leaks most teams never see
Every audit starts by tracing real spend. These are the patterns that show up first — named plainly, not dressed up as a feature list.
Leak 02
Wrong model for the job
A flagship model handles Haiku-grade work because nobody set a routing rule. The capability is paid for, the task doesn't need it.
Leak 03
Deferrable on real-time
Nightly reports and bulk jobs sit on synchronous endpoints. Batch APIs are a straight discount left unclaimed on work that was never urgent.
Representative example
A representative spend trace
A 45-minute look at one stack. Four of five entries carry avoidable cost — the audit names each one, ranks it, and shows the fix.
tokenledger.app/spend-trace
spend-trace.csv · 5 entries4 of 5 avoidable
| Feature | What it's doing | Spend | Flag | Fix |
|---|---|---|---|---|
| Support copilot | Flagship (over-capable) | $4,200/mo | risk | Route Haiku-grade work to a small model. |
| Nightly report batch | Sync realtime endpoint | $1,150/mo | risk | Move to batch API (vendor-documented ~50% off). |
| RAG retrieval | Re-embed every query | $2,800/mo | risk | Cache embeddings; shrink retrieved-context bloat. |
| Onboarding email gen | Small model (fit-for-job) | $320/mo | ok | Correctly scoped — leave as is. |
| Agent eval loop | Retry storms | $1,640/mo | risk | Cap max-tokens and max iterations per run. |
Illustrative only. Figures are representative of patterns seen across stacks, not a specific client's bill.
How the audit works
Two weeks, read-only access, no production credentials. You export; I analyze.
- Week 1
Discovery & spend map
You export 60–90 days of provider invoices and API logs. I map every dollar to a feature, a model, and a workflow — no live access required.
- Week 1–2
Leak analysis
Each line is ranked by annualized impact × effort: caching coverage, batch eligibility, model-fit, RAG waste, agent retry storms.
- Week 2
Findings & roadmap
A 12–15 page report with a prioritized savings list, the ROI math, and a 90-day implementation roadmap you can hand to an engineer.
- After
Implementation support
Optional: I help wire up the top fixes — caching, batch routing, model downgrade — and stand up monitoring so the gains hold.
One path, four steps
Start where the math works. Most teams should not jump straight to an audit — the Snapshot is free, and it tells us if an audit will pay back.
Free
AI Spend Snapshot
A 45-minute call and a one-page spend map naming your three biggest leaks with rough annualized costs. No fix promised — just clarity.
Learn moreFixed fee
AI Cost Audit
The two-week engagement. A ranked, ROI-math-backed savings list with the implementation roadmap.
Learn moreScoped
Implementation Sprint
I build the top three fixes plus one automation workflow, and wire up monitoring so the savings stick.
Learn moreRecurring
FinOps Retainer
Monthly spend review, anomaly alerts, price-change briefings, and one new automation a month after your first audit.
Learn moreHonest fit
Built for spend that can pay an audit back
The work is honest about fit. If you spend under about $5k/month on AI, an audit can't pay back — take the free checklist instead. If you're an enterprise with procurement and a platform team, you already have this covered.
Questions, before you book
No. The posture is “you export, I analyze.” You run the exports and dashboards; I interpret them. No API keys or production access are ever handed over.
We measure against price-held baselines: what the same workload would cost today without the changes. Falling prices make disciplined architecture more valuable, not less.
For visibility, yes. Dashboards answer “what did we spend.” They don't answer why a trace is expensive or what to change on Monday. That interpretation and implementation is what you're buying.
If you spend under about $5k/month on AI, take the free Snapshot and the checklist instead — an audit can't pay back at that level. If you're a large enterprise with procurement and a platform team, you likely have this covered internally.
Start free
See your three biggest leaks in 45 minutes
Book a free AI Spend Snapshot. No findings-fix promise, no sales theater — just a one-page map of where your AI money goes.