No public case studies yet.
Customer outcomes should be published only when the customer approves the wording, numbers, and context. Until then, this page stays conservative.
For evaluations, we can discuss anonymized workload patterns and sample deliverable structure under NDA.
Why this page is empty on purpose
Cost case studies are the easiest marketing artifact to write dishonestly. The savings percentage depends entirely on where the baseline was drawn, and a baseline chosen after the fact can make almost any change look transformative. A team that had never enabled prompt caching will show a large number for switching it on; that number says something about their starting point, not about the work. Publishing it without that context would be technically true and practically misleading.
What we can share before a case study exists
- Anonymized workload patterns — the shapes that produce runaway spend, such as agent runs with unbounded retries, retrieval that grows context faster than it improves answers, and background jobs nobody has attributed.
- The deliverable structure — what a baseline report contains, what the attribution model looks like, and which decisions it is meant to support.
- Method, in full — how invoices are reconciled, how quality is held constant during a routing test, and how savings are verified against a later invoice rather than a projection.
What a published case study will contain
When customers approve one, it will state the billing period the baseline came from, the traffic mix, which levers were applied in what order, the quality measurement used, and the verified invoice difference — including anything that did not work. A case study without a failed lever in it has usually been edited.
Until then, references and method discussions are available under NDA, which is a slower answer than a number on a page and a more accurate one.