Case study
A screener that has to show its work.
Every AI screening tool on the market will read a resume and tell you whether the person is a fit. They are good at it, and they are also entirely willing to invent the sentence that justifies the answer.
Which is a problem twice over.
The first cost is your time. A confident yes built on a quote that was never in the document sends you into an interview you did not need to take, and a confident no loses you someone you should have met.
The second is worse. A rejection with a fabricated reason attached to a real person's name is the kind of thing that ends up in front of a lawyer, and “the AI said so” has never once been a defence.
The fix
Three rules, written into the code.
Not asked for in a prompt, where the model is free to ignore them. Enforced afterwards, on every verdict, before anything reaches the report.
- 01
No quote, no claim
A verdict only counts if it is tied to a line copied out of the resume. The model can propose “met”, but the report never sees it unless that exact text is found in the document.
- 02
Silence is not a no
A resume that never mentions a skill is not evidence the person lacks it. “Not met” needs something in the document that actually contradicts the requirement. Everything else comes back “unclear”.
- 03
Nothing about the person
Age, gender, nationality, race, religion, marital status, health and photographs are stripped out before the text is sent anywhere. The system does not read them, store them or report them.
What happened when it ran.
Against a real job description with eight requirements, the model returned a confident “met” for one candidate and cited a line about running Kubernetes clusters at scale for six years. That sentence appears nowhere in the resume. The check did not find it, the quote was thrown away, and the verdict dropped to “unclear”. Genuine quotes in the same batch passed through untouched.
Another candidate came back 0 of 8. That was checked by hand and it was right: an IT technician against an engineering role. The tool returned nothing rather than finding something.
Expect “unclear” to be common. That is the design working, not the tool failing. A screener that always has an answer has stopped checking.
Sample resumes are synthetic and contain no real personal data.
Start here
Have something you keep doing by hand? Let us automate it.
Tell me what the task is and how often it eats your week. I will give you a straight answer on whether it is worth automating, and roughly what it would take, before you commit to anything.