A screening score is just arithmetic done on the data pulled out of a CV. If that data is wrong, the score is wrong, the ranking is wrong, and a good candidate quietly disappears. Nothing tells you.
How one wrong field loses a candidate
Follow one mistake all the way through.
A candidate lists a contract job as "Jun 2021 – Present" in a sidebar. The software reads the sidebar in the wrong order and misses that job. Their total experience drops from six years to two and a half. Your rules weight experience heavily, so they land in the bottom third. Nobody opens the bottom third of 500.
Look at what happened there. No error appeared. The tool did what it was built to do. Your rules were sensible. The person simply vanished, and the only trace is a low score that looks like every other low score.
This is the difference between software being wrong and software being unreliable. Software that crashes tells you something. Software that produces a confident number from bad data does not.
Which fields matter most
| Field | How much damage | Why |
|---|---|---|
| Dates | Severe | They decide total experience, gaps, how recent the work is — several rules at once |
| Job titles | Severe | They decide how senior the tool thinks someone is |
| Employer names | Medium | Used for industry and company size |
| Skills | Medium | Usually findable by searching the full text anyway |
| Education | Low | Only matters if you filter on it, which you mostly should not |
| Email, phone | Low for scoring | But you cannot contact someone you cannot email |
The pattern: fields used by several rules do the most damage. Dates are the single most important thing a tool does, and also one of the hardest.
Check 25 CVs by hand
Do this once for any tool, and again whenever the vendor updates it. Half an hour.
- Pick 25 CVs at random. Not the tidy ones. Include the scans and the two-column ones.
- For each, put the original next to the extracted data.
- Mark every field: right, partly right, wrong, missing.
- Count by field, not by CV. You want to know which field fails.
- Note the layout of any CV that failed badly. Patterns show up fast, and they are usually template families rather than individual documents.
What your results mean
| What you found | What it means | What to do |
|---|---|---|
| Dates wrong more than 1 in 10 | Experience numbers cannot be trusted | Never filter on experience with this tool |
| Titles wrong more than 1 in 7 | Seniority guesses are unreliable | Look for responsibilities in the text instead |
| One CV template failing again and again | A pattern, not bad luck | Treat it as a fairness problem, not just a quality one |
| Scans failing completely | Expected | Handle them as a separate manual pile |
| Everything above 95% | Your sample was too tidy | Try again with harder CVs |
Why a layout problem is a fairness problem
CV templates are not chosen at random. Which one someone uses depends on their country, how long they have been working, whether their university careers service gave them one, whether they paid for one, and what language they wrote in.
So if your tool reliably fails on one kind of template, it reliably disadvantages whoever uses it. And that disadvantage shows up in your results as a low score — which looks exactly like a weak candidate.
This is the most common way a purely technical fault turns into an unfair outcome. No check that only looks at scores will ever catch it.
When the tool gets it wrong
Fix the field by hand, and find out whether your tool learns from corrections. Some do. Most do not. If it does not, correcting is just tidying up one record — so fix it for people you are moving forward, not for the whole pile.
Download the 25-CV checking sheet · all templates
The rule
Never use a field to reject people until you have checked it. Use it to sort, to score, to rank — but do not let an unchecked field remove a person from your pile.