Use cases
First-pass CV screening that stays honest
Tobiloba Odejinmi · 21 May 2026 · 6 min · 1,317 words
Direct answer
An honest first-pass screen reads the CV against a role you wrote down and returns facts, gaps, and a packet for a person. It does not rank people by vibe and it does not send the reject. Hiring tools are high-risk under NYC LL144 and the EU AI Act. Escalate every yes and every close no. Do not score culture or scrape social first. Measure agreement with a human sample and time-to-shortlist.
- Cite the line or say “not stated”. Never fill a gap with a guess.
- A person owns the yes, the close no, and the rejection.
- NYC LL144 and the EU AI Act are reasons to keep records and go slower.
- Measure packet quality and misses, not “AI-screened hires”.
What is first-pass screening, and what is it not?
It is the Monday pile. Fifty CVs. A role that is supposed to be clear. A recruiter who will otherwise skim and hope. The employee reads each one against the must-haves and writes a packet. That is the job. The job is not to find “the best candidate in the market”. That sentence is how people justify a ranker they cannot explain.
I will say this as plainly as I can. If the system decides who gets seen, you are in high-risk territory. If it prepares a packet and a person decides who gets seen, you are doing review work. Those are different products. Do not market the second and ship the first.
The packet should be short. Who they are. Which must-haves are on the page. Which are not stated. Two questions for the screen call. If a recruiter still opens the PDF for every field, the packet failed.
How do you keep a screen honest?
Write the criteria first. Date them. Put them where the hiring manager can argue with them. Then extract. “Six years of Postgres, listed under Experience, 2018–2024.” Or “Postgres not stated.” That second phrase is the product. It stops the model from filling a hole with a vibe.
Do not score culture, energy, or “polish”. Those words are where bias hides. Do not read photos. Do not scrape social unless counsel said yes and a person still reviews every hit. Paper and the form are enough for a first pass.
I use the same loop I used at SmartComply. Structured output. A person on the odd case. A miss log you can read. Review work there fell by about half because the loop was tight, not because we trusted the model. Hiring needs the same loop and a slower gate.
- Criteria versioned before any CV is read
- Every must-have is met, missing, or quoted
- No culture, accent, photo, or social score
- A person sees every yes and every close no
What laws already treat this as high-risk?
NYC LL144 treats automated employment decision tools as something you have to audit and disclose if you use them in New York City. The EU AI Act puts many hiring systems in a high-risk class. You do not get to call a ranker “just a helper” if it decides who is seen.
I am not your lawyer. I am the person who has to build a file someone can open later. Keep the role profile, the version, the fields the system may read, the fields it must ignore, and who clicked yes. If that is in a Slack thread, you are hoping.
This is why I go slower here than on a support queue. A wrong support answer is a ticket. A wrong no is a person. At 10mg a wrong row can stop a clinic. Both need a name on the gate. Hiring already has statutes that say so.
What must a person still decide?
The yes. The close no. The rejection that goes out. Any internal or referred candidate. Anyone already in process. Any file that mentions a complaint, a disability, or a legal matter. The employee prepares. It does not close the door.
Escalation is a state handoff. The recruiter gets the packet, the criteria version, and why it is in their queue. A zip of PDFs is not a handoff. If they rebuild the screen, you did not save time.
Do not let “the AI said no” become a sentence in the building. If a hiring manager uses it, the design failed. The decision is theirs. The tool is a first pass.
What should you not score first?
Do not start with interviews, voice, or video. Do not start with a model that ranks the whole market. Do not start with “fit” against a successful employee’s writing style. That is how you copy last year’s bias and give it a dashboard.
Start with one role that has volume and a profile the team already agrees on. Licensed roles with a hard must-have are cleaner than “generalist operator”. If you cannot name the role, you are not ready.
Skip auto-reject on soft skills. Skip auto-generated rejection reasons that invent a gap. “We went another way” is honest. A fabricated skill miss is not.
How do you measure a shortlist that is not a trap?
Time from application close to a shortlist a manager will open. Share of packets a recruiter did not rewrite. Agreement with a holdout a senior recruiter screens blind. Those numbers tell you if the pile moved and if the packet is true.
Do not measure “hires made by AI”. A person made the hire. Watch for the quiet failure: the shortlist gets narrower in a way nobody can explain. That is a reason to stop, not to tune in the dark.
If you cannot explain a miss, turn that rule off. Same as document review. Harder here, because the miss has a name.
What do you write down before you go live?
The role profile and its date. Allowed fields. Forbidden fields. Who reviews yeses and close noes. How a candidate can ask what happened. Where the logs live. Counsel should see this before volume.
If setup feels like a project, good. You can still ship a first-pass screen in a week if the role is clear and the human gate is real. You cannot ship a silent ranker and call it a week of work.
When we sold Insurpass, buyers asked who could explain the system. A hiring screen gets the same question from a regulator or a candidate. Leave the file in a shape a stranger can read.
Questions people ask
What is first-pass screening, and what is it not?
It is a read of the application against a written role, with quotes. It is not a hire, not a culture score, and not a secret ranking of the market. If a manager cannot see the criteria, it is not a first pass. It is a black box.
How do you keep the model from inventing experience?
Structured fields and “not stated”. Every must-have is met, missing, or quoted. If the model cannot point to a line, the field is missing. That rule does more than a better prompt.
What do the laws change in practice?
They change what you owe people and what you must be able to show. NYC LL144 expects audits and notice for automated employment decision tools in New York City. The EU AI Act treats many hiring systems as high-risk. Build as if someone will ask for the file.
Can we auto-reject the obvious noes?
Only for must-haves you would already reject on paper, and only with a person sampling the bucket. A missing licence for a licensed role is a must-have. “Did not sound energetic” is not. When in doubt, it is a close no and a person looks.
What should we measure?
Time to a shortlist a hiring manager will open. Share of packets a recruiter did not rebuild. Agreement with a blind human sample. Demographic surprise is a reason to stop and look, not a KPI to optimise in the dark.
Written by
Tobiloba Odejinmi
Head of Engineering at 10mg Health. I have run engineering at Zeeh Africa and sold Insurpass and Shopl. I still write the code. If you have one process that still runs on people copying things, we can look at it in thirty minutes.
