Madrone / AI Cost Audit
Flagship models on routine tasks, uncached prompts, real-time rates on batch work, silent retry loops — if you run LLM APIs in production, your invoice hides recoverable waste. Our fixed-fee audit finds it, prices it, and hands you the roadmap.
No pitch · You talk directly to the person doing the work
Based on industry-typical waste rates of 30–50%. Your real number comes from the audit — and if we don't see a credible path past the fee on the free review, we'll tell you.
of AI product companies don't track their LLM API costs at all.
Mavvrik industry study, 2025price gap between flagship and right-sized models on routine tasks — at equal quality for those tasks.
Published provider price sheetstypical payback period on the audit fee, from implemented savings alone.
Engagement structure, not a promiseWhat we find
Flagship models doing routine work. Classification, extraction, and formatting that a model a tenth the price handles identically.
Static context re-billed every request. Caching cuts repeated-context cost by up to 90%. Often a same-day fix.
Overnight jobs at real-time rates. Batch endpoints price the same work at half.
Failed calls resent with full context. Spend that appears on the invoice and nowhere else.
Finding: 63% of requests are classification and extraction. Routed to a smaller tier, validated by A/B evaluation before rollout.
Finding: static context re-billed on each call. Prompt caching enabled — hours of work, ~90% off this line.
Finding: batch-eligible workload on real-time pricing. Same output, half price, done by morning.
Finding: unbounded retries. Capped, trimmed, alerting added — failures now page someone instead of billing someone.
How it works
Send two recent invoices. We come to the call with 2–3 concrete observations and a realistic savings estimate. No instrumentation, no obligation.
One URL change routes traffic through read-only instrumentation. You receive findings with dollar figures and a sequenced implementation roadmap.
Your engineers execute the roadmap, or we work alongside them. Every change ships behind a flag, validated against your quality benchmarks, with one-line rollback.
Model prices change monthly; optimizations decay. We watch continuously and flag issues before the invoice does.
Fixed audit fees, set by your current monthly AI spend. The audit typically pays for itself within the first 4–8 weeks of implemented savings.
We see metadata and token counts, not your customers' content. Engagement data is deleted at close-out.
Changes are recommendations until your team — or ours, with sign-off — ships them behind a flag with rollback.
No recommendation is "cheaper but worse." Every change is A/B evaluated against your benchmarks first.
Ongoing monitoring
Continuous cost monitoring with alerts, monthly breakdowns, and advisories when a cheaper equivalent model ships. You'll have dashboard access — most clients never open it. That's the point.
Common questions
Most findings come from traffic patterns, not source code: which models handle which request types, token volumes, cache hit rates, retry behavior, and timing. One URL change routes your API calls through read-only instrumentation that captures this metadata. Where code review helps — routing logic, prompt construction — it's optional and scoped with your team.
Not if it's done properly, and we don't ship anything that isn't. Every routing recommendation is A/B evaluated against your quality benchmarks on a slice of real traffic before full rollout. Tasks that genuinely need a flagship model stay on one — the waste is in the tasks that never did.
Yes — everything we do uses documented provider features, and the audit report shows your team exactly how. Most teams don't because nobody owns the problem: engineers are shipping features, finance sees one invoice line. If you'd rather build the muscle internally, the audit is a fast way to start; the roadmap is yours either way.
The free 30-minute review exists to prevent that. We look at your invoices first and give you a realistic estimate before you commit to anything. If we don't see a credible path to savings that clearly exceeds the audit fee, we'll tell you on that call and you keep your money.
OpenAI, Anthropic, Google, AWS Bedrock, and Azure OpenAI, plus multi-provider and agentic pipelines. The optimization levers — routing, caching, batching, spend controls — exist across all major platforms.
Instrumentation is read-only and covered by NDA. We work with metadata and token counts, not your customers' content. All engagement data is deleted at close-out, and we'll sign your security documentation where needed.
Tell us roughly what you spend, and we'll come to the call with observations — not a pitch.
No obligation · NDA on request · Canadian & US clients