Most bad AI screening is not the software's fault. Someone turned a wish list into a set of filters, cut 500 people down to nine, and four of those nine were wrong.
Two kinds of rule
This is the most important idea in the module, and it is simple.
- A filter removes people. They are gone.
- A score moves people up or down the list. They are still there.
Filters should only be used for things that are genuinely non-negotiable and easy to check: the legal right to work there, a licence the law requires, a language the job is done in.
Everything else should be a score. Skills, tools, industry background, seniority, education — all of it.
Most tools make everything a filter by default, because filters are easy to build and produce satisfyingly short lists. Resist that. A list of 25 you can skim beats a list of 6 you cannot check, and the skim takes ten minutes.
Six things never to filter on
| Do not filter on | Why not | Do this instead |
|---|---|---|
| Years of experience | Built on dates the tool may have read wrongly. Punishes career changers | Score it, and stop rewarding extra years past what you need |
| A degree | Weak predictor for most jobs, and closely tied to background | Filter on the skill the degree stands for |
| Current job title | Every company names jobs differently | Look for what they actually did |
| Gaps in employment | Usually caring, illness, moving country, or redundancy | Ignore at this stage. Ask at interview if it matters |
| Location, for remote jobs | You said it was remote | Only filter on time zone, and only if the job needs it |
| Staying a long time at one employer | Punishes contractors and agency people | Score the relevant work, not the shape of the career |
Some rules look fair and are not
"Years of experience" feels like a fact. It is not. It is a sum calculated from dates the tool guessed at, using an assumption about what counts as relevant.
"Worked at a similar sized company" sounds neutral. It mostly reflects where someone lives and who they know.
"Steady career progression" rewards people whose lives allowed a straight line.
The test: can you point at the sentence in the CV that proves it, and would two people agree that it does? If not, the rule is doing something other than what its name suggests.
Turning your scorecard into rules
Take the scorecard you built in Module 2. Sort every line into one of three places:
- A filter. Three at most. Each has to survive this question: would I really reject an outstanding candidate for missing this?
- A score. With a weight, and a note saying what counts as proof.
- An interview question. Anything you cannot check from a document. Most personal qualities belong here. Put them in the screening and you are really scoring how well someone writes about themselves.
Writing a "what counts as proof" note
One line per score. Compare these:
| Rule | Weak note | Usable note |
|---|---|---|
| Pipeline experience | Has worked with data pipelines | Names a specific tool and describes what they built with it |
| Leading a team | Leadership experience | Gives a team size, or names people they managed |
| Client-facing | Good with clients | Describes working with customers outside the company, not colleagues |
The usable ones are longer and more specific. That is the point — you can check them against a CV in ten seconds.
Test on 50 before you run on 500
Write your rules. Run them on 50 CVs. Read the top five and the bottom five. Would you have made those calls? Adjust once. Then run the full set.
Two rounds is usually enough. Testing on 50 keeps it quick enough that you will actually do it. The temptation is to go straight to all 500 because the software makes it cheap — but then you are judging 500 results you have already committed to in your head.
Download the screening rules sheet · all templates
You get three filters
If your list of must-haves has more than three items, they are not requirements. They describe a person who does not exist, and you are about to spend six weeks finding that out.