AI resume screening that explains itself
Most AI screening tools return a number. WadeCV returns a number, the reasons behind it, the values it extracted to get there, and a second score you can recompute yourself. This page explains exactly what happens to a CV between upload and ranking, and what the AI is not allowed to do.
- Screening starts from a rubric, not a keyword list: the criteria you care about for this role, each with a measurement (linear over a range, or closest to a target) and a weight. WadeCV drafts it from your job description and you edit it.
- For each CV the AI extracts the value of every criterion, summarises education, experience and relevance in about a hundred words each, and gives a fit score out of 100 with named strengths and gaps. You can read its reasoning on every candidate.
- A second score, the criteria score, is computed arithmetically from the extracted values and your weights. Change a value or a weight and it recomputes without another model call, so recruiters can correct the AI and see the effect.
- The model is instructed not to infer age, gender, ethnicity, religion, disability, nationality or family status, and that instruction is covered by automated tests against the live model. Names and photos do not affect the score.
- Every screening is a decision aid. Scores can be overridden, every stage move is made by a person and recorded, and candidates are told on the apply form that AI reads their CV.
What happens to a CV
- 1
The rubric comes first
Paste a job description. WadeCV proposes the criteria that matter (always including job description match, average tenure and years of relevant experience, plus up to five more), how each is measured and how much it weighs. You approve or change it.
- 2
The CV is read, not matched
The text is extracted (page images are used only for scanned files) and sent to the model together with the rubric. It returns structured data: contact details, summaries, years of experience, per-criterion values, free-form attributes and a fit assessment.
- 3
Two scores, one explained
The fit score is the model's own judgement, written up as strengths and gaps. The criteria score is deterministic: each criterion's value is mapped through the measurement you chose, weighted, and summed. Both are shown side by side.
- 4
Then people take over
Recruiters override values, add remarks, move candidates between stages, and decide. Every action is recorded against a named member of the organisation.
The two scores
Fit score (0 to 100). The model's judgement of the whole CV against the job description, returned with a list of strengths and a list of gaps. It is holistic and it is opinionated; that is its value, and it is why the reasons are always shown next to it.
Criteria score (0 to 100). Arithmetic. For each criterion in the rubric the model extracts a value (years of relevant experience: 6; job description match: 0.8). A linear criterion maps the value across its range; a target criterion scores by closeness to the target you set. Values are weighted by the weights you chose and summed. Change a weight, or correct a value the model got wrong, and the score recomputes instantly with no model call.
When the two disagree, that is information: the model saw something in the prose that the rubric did not ask about. The candidate sheet shows both.
What the AI is not allowed to do
- No prediction of tenure or performance. Models that predict how long someone will stay are trained on past decisions and inherit past bias; WadeCV removed this feature deliberately.
- No inference of protected characteristics. The model is told not to estimate age, gender, ethnicity, religion, disability, nationality or family status, and the test suite checks a CV that invites those inferences.
- No automatic rejection. A low score changes nothing about a candidate's status; only a person moves them to Rejected.
- No learning from your decisions. Scores come from the rubric and the CV, so the hundredth candidate is screened the same way as the first, and a rubric change re-screens on the same footing.
Where the candidates come from
Uploads by your team, and applications to a free job posting. A CV uploaded without a role is routed to the best open role by the same model, and that routing counts as part of the one credit. Applications are stored the moment they arrive and screened a few minutes later, so an applicant never waits on the model.
After screening, every candidate is searchable by meaning ("payments engineers who have led a team") and by distance from a place, and the same screened summaries feed video introductions and the pipeline analytics.
Frequently asked questions
Which AI model reads the CVs?
Google's Gemini, through Vertex AI, with a fixed prompt and a strict output schema. The same model powers WadeCV's job-seeker tools. Prompts are versioned, and every call is traced so a surprising score can be investigated.
How accurate is the fit score?
It is as good as the rubric. A vague job description produces vague criteria and a score that tells you little; a specific rubric produces a ranking recruiters usually agree with. That is why the rubric is shown and editable before the first CV is screened, and why the criteria score exists as an arithmetic check on the model's judgement.
What does a screening cost?
One credit per CV, whether it arrives by upload or through your job posting. Re-screening after you change the rubric or move a candidate to another role costs another credit, and the preview tells you how many before you confirm.
Is automated CV screening legal?
Rules differ by place. New York City's Local Law 144 requires bias audits and notice for automated employment decision tools; the EU AI Act classifies recruitment AI as high risk with transparency and human-oversight duties. WadeCV is built as a decision aid with a human in the loop and an explanation on every score, and the apply form tells candidates that AI reads their CV. Whether a given use needs a notice, an audit or an impact assessment is a question for your own counsel.
Can candidates see their score?
No. Scores, summaries and remarks are visible only to members of your organisation. What a candidate sees is the job page, the consent notice on the apply form, and whatever you tell them.
What if the AI gets a CV wrong?
Open the candidate, correct the value it extracted, and the criteria score recomputes. If the CV was a scan the text came from page images, which is where most misreads come from; the original file is always one click away.