The number, and how it was measured.

Because permission to describe the work and permission to name the client are different conversations, and we only publish what we hold. The engagements are real and the numbers come from the delivery record. Where a client agrees to be named we will name them, and until then a description is the honest form.

Because we publish the ones we can stand behind. A metric needs a baseline, a method and a period before it means anything, and assembling that properly for an engagement takes work that many firms skip. We would rather show fewer figures that survive a question than a page of percentages that do not.

By measuring the current process before anything changes, which is the step most often skipped and the reason so many projects cannot say afterwards whether they worked. Where no baseline was taken we say so instead of reconstructing one, because a baseline invented after the fact is not evidence.

Real users, real data, and somebody accountable when it breaks. A pilot with a friendly group of testers is a useful stage and a different thing, so we report them separately. Pilot numbers flatter reliably enough that treating them as results is the most common way AI work is oversold.

Usually, once a conversation is serious enough to justify asking a client for their time. Reference calls are arranged with the client's agreement for the engagement closest to your situation, which is more useful than a general one. We ask rather than assume, which is why it comes later in the process.

Then it appears here with what went wrong, because a portfolio of unbroken successes is not credible and does not help you judge us. Every case names something we would do differently. The useful question in a first conversation is what we have got wrong before, and we answer it.