Why we won’t tell you our accuracy rate (yet)

It would be easy to put a number on our homepage. “95% fraud detection accuracy.” It would look good. It would also not be true — not because the system doesn’t work, but because we haven’t earned that number yet, and saying it anyway would be a lie dressed up as marketing.

What “unsupervised” actually means

Our machine-learning layer is unsupervised. It finds statistical outliers relative to a fitted distribution of normal invoices — invoices that look unusual compared to what’s typical for that supplier, that amount, that pattern of activity. It does not learn from a labelled dataset of “this was fraud, this wasn’t,” because no such dataset exists for us yet. There’s a real difference between finding the unusual and confirming the fraudulent. We let a human reviewer do the second part.

Where the real numbers come from

Every decision a pilot reviewer makes — approve, reject, escalate, mark as a false positive — becomes a labelled example. That’s the foundation of a genuine evaluation set: precision, recall, and a properly chosen threshold, measured against real outcomes rather than asserted from a slide. We’re building that set now, pilot by pilot, and we’ll publish the numbers once we trust them ourselves.

Until then, if you ask us for an accuracy figure, the honest answer is: we don’t have one we’d stand behind yet. We think that answer, consistently given, is worth more than a number that sounds better.