No public case studies yet.

Last updated: 24 April 2026

We're early-stage. Every active engagement is under NDA. We won't publish a customer logo or a savings number without explicit written approval ; and so far, none of our customers have asked us to.

What we can share under NDA

If you're evaluating us for an engagement, ask. We can typically share:

• Anonymized baseline → after for an engagement matching your provider mix and workload shape (e.g. "RAG-heavy SaaS on Anthropic + OpenAI, 7-figure annual spend").
• Reference call with a current or former customer, when they've consented. We don't keep a "reference list" ; we ask per-conversation.
• Sample audit deliverable with customer name and identifying details redacted. Shows the format and depth, not the numbers.

What we won't do

• Publish unverifiable savings claims. The headline numbers in our pricing example on the homepage are explicitly labeled illustrative. Real customer numbers aren't on this site.
• Use a customer logo without written permission, even if their procurement team mentioned us in an RFP.

When this page changes

When a customer offers to be a public reference, we'll publish here ; with their words, their numbers, and their approval on file. Until then this page stays honest.

Get in touch

Want anonymized references for your evaluation? hello@llmcfo.com.

Why savings percentages are the wrong thing to publish

A cost case study can be made to say almost anything depending on where the baseline is drawn. A team that had never enabled prompt caching will show a dramatic number for switching it on — that number describes their starting point, not the work. Percentages also hide the denominator: cutting 60% of a bill that was inflated by a runaway retry loop is a bug fix, and quoting it next to a structural optimization on a well-run system is comparing two unrelated things.

The number worth publishing is the verified difference between two invoices, with the traffic mix and the quality measurement stated alongside it. That takes a customer willing to have their spend discussed in public, which is a real thing to ask for and not something to buy with a discount.

What a published case study here will contain

• The billing period the baseline came from, so the comparison is anchored to something checkable.
• The workload shape — provider mix, traffic pattern, and what the model was actually doing.
• The levers, in the order they were applied, with the measured effect of each rather than one combined headline.
• The quality measurement used to confirm that nothing regressed, and what its limits were.
• What did not work. A case study with no failed lever in it has been edited.

Until then

Method is fully shareable even when outcomes are not. We will walk through how baselines are built, how routing tests hold quality constant, and how savings get verified against a later invoice rather than a projection — in detail, before you sign anything. That is a slower answer than a percentage on a landing page and a considerably more useful one.