You can see the AI bill.
You cannot see what is causing it.
An independent audit of what you actually spend on AI and LLM APIs. You get a measured breakdown, a waste list with dollar figures, and a roadmap ranked by payback. No tool to buy. No savings claim we cannot show you in your own invoice.
The diagnosis is free, in writing, and stays free if you walk away.
The audit fee is fixed and quoted before we start.
Every AI cost tool promises big savings.
None of them start by measuring.
Guessing is how teams end up paying for a tool that "saves" money they were never spending. We start at the other end: your invoices, your traffic, your actual numbers — then we tell you what is worth changing, and what is not.
You cannot tell which endpoint or feature is eating your budget.
Your team is guessing at the biggest win.
Every vendor's number comes from their own dashboard.
Identified is not the same as realized
Almost every "we cut AI costs by X%" number you will be shown is identified savings: money an analysis says is recoverable if you make the change. That is not the same as realized savings: money that has actually left your bill, measured on a later invoice. Reports that only count the first kind overstate the result, and you only find out months later.
What you get
A report of 8–14 pages (scope dependent), in six sections:
Spend map
Cost split by provider, model, endpoint and feature.
Unit economics
Cost per request, per user, per feature — the numbers that tell you what to fix first. (This section needs request-level logs; see the two levels below.)
Waste list
5–10 specific findings, each with a dollar estimate and the evidence behind it.
Optimization roadmap
Ranked by savings divided by effort, so you know what is a five-minute config change and what is a project.
What we do not recommend
Including things a vendor would happily sell you. If a change is not worth it, it goes here.
Method appendix
Our calculations and assumptions, so you can check our arithmetic instead of trusting it.
Two levels of access — you decide how deep we go
If you only have the invoice, we say so up front — and the report stays honest about what cannot be concluded from it.
Open the worked example →A 14-page sample report on synthetic data, with the arithmetic shown.
Who this is for
Teams spending $1,000 – $50,000 per month on LLM APIs who suspect the bill grew faster than the product did.
Pricing
The diagnosis is free and written — and if the honest answer is "do not buy an audit yet", that is what it says. The audit fee is fixed and quoted before we start: no hourly billing, no scope creep. It is not contingent on results. The performance fee is the only part that depends on outcomes: 15% of savings that actually show up in your invoice, applied only above $5,000/month of LLM spend, and only on the realized part. If we cannot find anything worth changing, the report says so — a valid, useful result, and no performance fee.
Why trust us
We publish our method. The method appendix of every report shows the exact calculations and assumptions we used. You can verify our arithmetic instead of taking our word for it.
We show the baseline first. A savings percentage means nothing without a measured starting point. Every number in the report traces back to a line in your own data — and we separate what is identified from what is realized.
We tell you what not to do. Most cost advice comes from companies selling the fix. We do not sell a tool, and we take no vendor commissions or referral fees, so "change nothing yet" is a recommendation we are allowed to make.
Read-only by default. Exported invoices and usage CSVs only. No credentials, no production access, no code, and no prompt or response content. A mutual NDA is available on request.
Operator-led, not sales-led. This is a small independent practice, run by Lapisli: you work directly with the person doing the analysis, not an account manager. Ask for a redacted sample report — a worked example on synthetic data — before you spend anything, so you can judge the format and the depth first.
Questions
What if you find nothing?
Then the report says so, and it tells you which of your current costs are actually justified. That is a real answer, you still get the breakdown, and no performance fee applies.
How is the performance fee calculated?
On realized savings only: the same line items in your invoice, before and after the change, across 1–3 billing cycles. You see the arithmetic. Savings we identified but cannot yet confirm in your bill are not charged for.
Do you implement the changes?
Not inside the audit. The audit tells you exactly what to change and what it is worth. Shipping the changes and re-measuring them monthly is the implementation retainer.
How fast?
Free diagnosis: 3 business days from your invoice. Audit: 7–10 business days from the moment we have your data. The clock starts when the export arrives, not when you pay.
Do you need access to my code or cloud account?
No. Exported invoices and usage CSVs — plus request-level logs only if you want per-request attribution. We never ask for credentials.
How do I pay?
In USD by bank transfer; a card or Payoneer invoice is available on request. The audit fee is due on delivery of the report, and the performance fee is invoiced only after savings are confirmed in your bill.
Start with last month's invoice
For the free diagnosis, send one line on what the product does and your most recent API invoice or usage export. You get a written read and, if it is worth it, a fixed quote and scope — before any payment.
No newsletter, no drip sequence. One reply, from the person who will do the work.