Why an AI should never reject anyone by itself

No. A CV score is evidence, not a decision, and the difference matters more than it sounds. A model can tell you that a CV does not mention SQL. It cannot tell you that the person learned SQL last year in a job the CV describes badly, or that the requirement was written by someone who has since left. Let it rank and explain. Let a person decide.

We build CV screening and AI interviews, so this is a strange thing for us to argue. It is also the rule the whole product is shaped around, and it costs us something: our software is slower than one that would simply reject the bottom 60% overnight. Here is why we accept that.

What the score actually is

Each CV is scored against one job description on six weighted dimensions: skills, experience relevance, education, achievement, role alignment and stability. Every dimension quotes the line from the CV that earned it. The weights are yours to set per role, because "analyst" at one company is not "analyst" at another.

That produces a number between 0 and 100 that means one specific thing: how much of what this job description asked for is visible in this CV. It is a measure of the match between two documents. It is not a measure of the person.

Three things it cannot see

1. A gap with a reason

An eighteen-month gap reads the same whether someone was ill, caring for a parent, running a business that failed, or studying. The CV rarely says. A recruiter asks in the first call and gets an answer in a sentence. A scoring model has nothing to ask.

2. A good worker who writes a bad CV

In our own screenings the median CV matched 38% of what its job description asked for. Some of that is genuine mismatch. Some of it is people who do the work daily and never thought to write it down, because nobody taught them that "handled the monthly reporting" should say which tool, which volume, and what changed because of them. The score sees the writing, not the work.

3. Requirements nobody stated

Job descriptions leave out the things everyone in the team already knows: that this role sits between two arguing departments, that the previous person burned out, that the tool listed is being replaced next quarter. The model scores against what was written down, which is always less than what is true.

What a machine is genuinely good at

None of this means the scoring is useless. It is very good at the parts people do badly:

• Consistency. CV number 400 is read with the same attention as CV number 4. A person cannot do this, and honest recruiters know it.
• Order independence. In a manual pile, arriving early is an advantage. In a scored list, it is not.
• Showing its working. Every number points at the sentence that produced it, so your team can disagree with it specifically rather than vaguely.

That is a ranking tool and an argument starter. It is not a judge.

What we do instead

The rules below are mechanisms in the product, not intentions:

• Held is never rejected. A candidate below your cut-off sits in a held state, with their place kept, until a person decides. Nothing expires them.
• Every bulk action names the people. Before you reject or shortlist a group, the confirm screen lists every person affected and who is being skipped. Above fifty, you type the number to confirm.
• Undo. Reject, hire and shortlist can be undone as a batch, and everyone goes back to the exact stage they were in.
• Candidates never receive a score. The emails we send carry no numbers and no reasons. A score is an internal working note, not a verdict to hand someone.

The one thing the system does on its own is invite people to an interview when they clear a threshold you set. An invitation is not a decision — nobody is worse off for being asked — and you can switch it off per job.

The honest cost of this rule

A person still has to look. On a 500-applicant posting that is perhaps twenty minutes with a ranked list, instead of the eight hours it would take unsorted — but it is not zero, and any vendor who tells you it is zero is selling you a liability. When a rejection turns out to be wrong, "the system did it" is not a defence anyone wants to give a candidate, a client, or a court.

What to ask any vendor

1. Can your software move a candidate to rejected without a human click? If yes, walk away.
2. Show me the screen where the decision happens. Does it name the people affected?
3. Can that action be undone, and does the person go back to where they were?
4. What does the candidate receive, and does it contain a score?

We answer those four the same way every time, and you can check each one inside the product rather than taking our word for it.

Those four are about the decision. There are six more worth asking before you sign anything — about where candidate data lives, what you can take with you, and what happens on the day you exceed your plan.