Skip to main content
Superhawk uses AI to read evidence and make judgements. It is genuinely good at some things and genuinely unreliable at others, and knowing which is which is most of using it well.

Where it is strong

Recall across everything. Superhawk remembers every call, ticket, and email. It will find the sentence from March that nobody else remembers, and this is where it beats any human reading the same material. Noticing change. Usage falling, a stakeholder going quiet, a theme repeating across quarters. Change against a baseline is what it is built to see. Grounded extraction. Pulling what was actually said out of a transcript, with attribution, is reliable. Coverage. It reads every account every cycle. Nothing is skipped because the week got busy.

Where it is weak

Thin evidence. With one connected source and no call history, judgements are guesses dressed as conclusions. This is what the insufficient data band and the low confidence marker exist to tell you, and they are worth believing. Context it cannot see. A customer who told you something on a phone call that was not recorded, an internal reorg, a commercial negotiation happening in a channel Superhawk is not connected to. It reasons from what it has. Entity matching against external sources. Signals sourced from public company data depend on matching your customer to a public record. When the match is wrong, the signal is confidently about a different company. Always check the company name in the evidence. Anything genuinely novel. Patterns require history. A first-of-its-kind situation gets read against the nearest thing it has seen, which may not be close.

How to tell which you are looking at

Working with it well

Follow the evidence before you act. Especially before repeating something to a customer. The citations exist for this, and checking one takes thirty seconds. Treat confidence as part of the answer. A red band at low confidence means “go and look”, not “this account is failing”. Dismiss what is wrong. A dismissed signal keeps the list trustworthy and feeds back into tuning. Leaving noise in place trains your team to ignore the whole surface. Connect more sources. Almost every accuracy problem in practice is a coverage problem. The single highest-value thing you can do is connect your call recorder.
Superhawk can be wrong. Verify anything you are going to act on or repeat externally. This is not a disclaimer for its own sake — a confidently-worded judgement built on two data points reads exactly like one built on two hundred, and the confidence marker is how you tell them apart.