How to build a candidate scoring model you can defend
Of the 33,753 résumés that went through relevance scoring in our published hiring data, 49.5 percent matched less than 30 percent of what the role asked for. At the top end, 1.7 percent cleared 70 percent. Read those two numbers together and you have the whole case for a candidate scoring model: half the pile is clearly off target, a sliver is clearly on it, and everything in between is where recruiters lose their week.
The instinct is to handle that middle by reading harder. At volume it doesn’t work. What works is deciding, before the first résumé is opened, exactly what earns a point, how much each thing is worth, and where the line sits. This post is how to build that model, how to weight it, where to set the cut-offs and how to keep every score explainable, so the shortlist can go to interview in one step instead of a week of second-guessing. It’s the hub for a five-part series on getting from résumé to interview with job-matched scoring.
What is a candidate scoring model?
A candidate scoring model is a fixed set of criteria drawn from the job description, each with a weight and a clear rule for what earns points, applied identically to every applicant. It turns a pile of résumés into one score per person, a ranked list, and a cut-off that decides who is invited to interview and who is reviewed by a person.
The important word is fixed. A model that changes depending on who is reading, or on how tired they are at résumé 400, is not a model. It is a mood. The whole value comes from scoring the first and the last applicant against the same yardstick.
A scoring model is also not the same thing as a filter. A filter answers yes or no: does the candidate have the licence, can they work from the office, will they relocate. A score answers how much: how closely does this person’s track record match what the role needs. Mixing the two is the most common reason a model ranks badly, and we’ll come back to it.
Why gut-feel ranking breaks at volume
Picture a support role with 1,200 applicants and two recruiters. That is a hypothetical, but the arithmetic is familiar. At even two minutes a résumé, that’s 40 hours of reading before anyone is called. Nobody spends two minutes on résumé 900. By then the reading has collapsed into pattern matching: a known college, a known employer, a title that looks right.
Those proxies are fast because they’re visible, not because they predict anything. A 2019 meta-analysis in Personnel Psychology pooled 81 samples and found that pre-hire work experience correlated just .06 with later job performance, and .00 with turnover. The authors cautioned against selecting on such measures “unless more positive evidence emerges.” Total years is the single most common thing a tired reader counts, and it is close to noise.
The second failure is drift. Recruiter A thinks a B2B background matters. Recruiter B doesn’t. Both are reasonable, and the candidates in A’s half of the pile now face a different bar from the candidates in B’s half. Nobody notices, because nobody wrote the bar down.
A model fixes both problems at once. It makes the criteria explicit, which forces the argument about B2B experience to happen once, before screening, between the recruiter and the hiring manager, instead of 600 times inside two heads.
Pick criteria from the job, not from the résumé
The criteria come from the job description, and only from it. The temptation is to start from what résumés make easy to read: degree, college, years, company names. Resist it. Start from the question “what will this person spend their first six months doing?” and work backwards.
For a customer support role, that might give you:
- Handling customer conversations as the main part of a past role, by phone, chat or email.
- Experience with a ticketing or CRM tool, any brand.
- Written English shown in the résumé itself or in a described task.
- Working in a shift or target-driven environment.
- Domain familiarity with the product category, as a plus.
Notice what’s missing. No college tier. No “minimum three years”. No named employer. Each criterion describes work, and each can be checked against something the candidate wrote.
Two tests for every criterion before it goes in. First, can you point to the line in a résumé that would earn it? If not, the criterion is a wish, not a criterion. Second, would the hiring manager reject a strong candidate who lacked it? If yes, it’s not a scored criterion at all. It’s a hard requirement, and it belongs in a separate yes-or-no check. How to write those requirements so they can be checked at all is the subject of the job specification post in this series.
How to weight the criteria
Weights are where the hiring manager’s real priorities come out. Ask them to split 100 points across the criteria. Most will try to give everything 20. Push back until one or two criteria clearly lead.
A worked example for the support role:
| Criterion | Weight | Full points for | Partial points for |
|---|---|---|---|
| Customer conversations as main work | 35 | 2+ years where this was the job | Some customer contact inside another role |
| Ticketing or CRM tool | 15 | Named tool used daily | Mentions a tool without detail |
| Written English | 20 | Clear, specific résumé with described outcomes | Readable but generic |
| Shift or target environment | 15 | Stated targets, shifts or SLAs | Implied by the type of employer |
| Product-category familiarity | 15 | Same category | Adjacent category |
A few rules keep weights honest:
- No criterion below 10. If it’s worth less than 10 out of 100, it won’t change anyone’s rank. Drop it.
- No criterion above 40. Above that, the model is one question with decoration.
- Write the partial-credit rule. “Some customer contact inside another role” is the line that decides most middle candidates. Leaving it to judgment brings the drift straight back.
- Score relevance, not quantity. Ten years in an unrelated function should earn little against a role that needs the specific work. That is exactly what the .06 finding warns about.
Keep the weights in the same document as the job description. When the role reopens in six months, you want the model to come back with it, not be rebuilt from memory.
Where to set the cut-offs
A single cut-off is a trap. Set two, and you get three bands:
- Above the upper line: invite to interview without further reading.
- Between the lines: a person reads these, with the score and its evidence in front of them.
- Below the lower line: decline, or hold for another role.
Where the lines go depends on how many interviews you can run, not on a magic number. If you can interview 150 people this week and 1,200 applied, the upper line sits wherever the 150th-ranked score falls, rounded to something the team can remember. Then check it: open five résumés just above the line and five just below. If the five below look as good as the five above, the model is not separating them, and the weights need another look before the line does.
Don’t set the lower line by the same method. Set it where the evidence runs out. In our data, half the pile sat below 30 percent relevance. Those are the facility managers applying to a software role, and the bottom line can sit near there without much risk. The risk is all in the middle band, and that’s where your people’s reading time should go. We wrote separately about why a score filter needs a review queue for exactly that band.
Make every score show its evidence
A number with no reason behind it is unusable. When a hiring manager asks why candidate 47 scored 62 and candidate 48 scored 58, “the model said so” ends the conversation badly. The answer should be a line from each résumé, quoted, next to the criterion it earned.
Evidence does three jobs. It lets a reviewer in the middle band decide in 30 seconds instead of three minutes, because the reason is already highlighted. It lets the team catch a model that is scoring the wrong thing: if “customer conversations” keeps being awarded for a line that says “managed client relationships” at a bank, you’ve found a criterion that needs a sharper rule. And it gives you something to say to a candidate who asks why they weren’t shortlisted.
This is the part HireQwik was built around. Every résumé is scored against the job’s own requirements, with the line from the résumé behind each score quoted. Deal-breakers can be written as plain sentences, such as “must be a graduate”, and sit apart from the score as the yes-or-no checks described above. More on how that looks in practice in explainable scoring and the evidence behind a score.
Scoring freshers, career-changers and returners
Most scoring models are built with an experienced applicant in mind, then quietly fail the three groups that make up a large share of volume hiring.
Freshers have no work history to score, so a model weighted on past roles ranks every one of them near zero, and the ranking among them is random. Build a separate, lighter model for campus and entry-level roles: internships, projects, coursework that matches the work, and written clarity in the résumé itself. Then compare freshers with each other, not with someone who has three years in the role.
Career-changers often describe the right work under the wrong titles. A school administrator who ran timetables, parent communication and fee follow-ups has done a lot of what an operations coordinator does. A model that scores titles misses them; a model that scores described work finds them. This is the clearest case for writing criteria as work, not as job names.
Returners after a break carry their experience with a date gap in front of it. If the model has a recency weight (“in the last two years”), it will push strong returners down for the gap alone. Decide deliberately whether recency matters for the role. For most skills it matters far less than the gap suggests, and the résumé red flags post in this series covers why a gap belongs in an interview question, not in the score.
A quick check after each drive: pull the top 20 from each of these groups and read them. If the strongest people in a group are all sitting in the bottom band, the model is scoring the group, not the person.
Common mistakes that wreck a scoring model
Putting deal-breakers inside the score. If “must relocate to Pune” is worth 20 points, a candidate who won’t relocate but is strong everywhere else still scores 80 and lands on your shortlist. Hard requirements are filters, not weights.
Scoring the résumé’s polish. A beautifully formatted résumé with no specifics should not outscore a plain one that says “handled 60 tickets a day on Freshdesk, 92 percent CSAT”. Score content, not layout.
Copying the model across roles. A sales model reused for an operations role will rank people on the wrong things with total confidence. Reuse is good when the role really is the same. A role’s screening setup can be saved and reused for a new job in HireQwik for that reason, but check the criteria against the new description before you do.
Never checking the model against outcomes. Six weeks after the drive, look at who was hired and how they scored. If hires came from the middle band more than the top, the weights are off. In HireQwik, HR’s decisions sit next to the AI’s recommendations for each job, which is the quickest way to see where the two part ways.
Treating the score as the decision. The score ranks. People decide. The point of a good model is that the deciding gets faster and fairer, not that it disappears.
From ranked list to booked interview
Once the model is built and the cut-offs are set, the last step should be the shortest. The candidates above the upper line don’t need another meeting about them. Approve them as a group, and they should each get an invitation with a link to book their own interview slot.
In HireQwik that is the recruiter inbox: filter to the band above the line and approve in bulk. Each approved candidate receives an email link to choose an interview time, with reminders until the booking window closes, and then speaks with an AI interviewer for 10-20 minutes in an ordinary browser tab. Nothing needs installing, and the questions are drawn from the job and from what that person’s résumé says. The résumé score and the interview score are then combined into one recommendation, with a flag when the two disagree. HR still makes the final call.
The rest of this series takes each part of the model in turn: writing a job specification a screen can score, screening applicants in before screening them out, re-scoring your talent pool for a new role, and which résumé red flags a score should weigh.
Build it once, before the next drive
If you build one thing this week, make it the weights table for your highest-volume role. Five criteria, 100 points, a written rule for partial credit, and two cut-offs. Score twenty résumés from your last drive by hand against it and see whether the ranking matches who you actually hired.
If you’d like to see a model like this run on one of your own job descriptions, book a demo and we’ll score a sample set with you. Or take the live AI interview yourself to see what the shortlisted candidates experience next.
Frequently asked questions
What is a candidate scoring model?
A candidate scoring model is a fixed set of criteria taken from the job, each with a weight and a rule for what earns points, applied the same way to every applicant. It produces one score per candidate, a ranked list, and a cut-off that separates who is invited to interview from who is reviewed or declined.
How many criteria should a candidate scoring model have?
Four to six is a workable range for a first screen. Fewer than four and the score is decided by one or two things, usually years and title. More than eight and each criterion carries so little weight that the total stops telling candidates apart. Hard requirements belong outside the score as yes-or-no checks.
Should years of experience carry a big weight in candidate scoring?
Usually not on their own. A 2019 meta-analysis of 81 samples found pre-hire experience correlated only .06 with later job performance. Score relevant experience instead: the years spent doing the work this job needs, shown by what the candidate did, not total years in any role.
See your own candidates screened
Book a 30-minute demo. Bring a live JD and we'll screen against it, then start with a pilot on your own candidates before committing to anything.
Existing customer? Sign in