Use cases
AI employee for recruiting
Tobiloba Odejinmi · 12 Dec 2025 · 6 min · 1,397 words
Direct answer
A recruiting AI employee reads the pile and returns a shortlist against criteria you wrote down. It does not hire. Good looks like a packet a recruiter can defend. Escalate anything that is a yes, a close no, or a legal risk. Do not automate interviews or rejection reasons first. Hiring AI is high-risk work under NYC LL144 and the EU AI Act. Measure time-to-shortlist and whether a person still owns the decision.
- The job is a first-pass shortlist against written criteria, not a hire.
- A person still owns every yes and every close no.
- NYC LL144 and the EU AI Act treat this as high-risk. Act like it.
- Measure time-to-shortlist and audit misses. Do not measure “AI hires”.
What part of recruiting is safe to hand over?
The pile. CVs, forms, the same missing-cover-letter problem every Monday. A person should not spend a career copying years of experience into a spreadsheet. The employee can pull the facts that are on the page and sort them against a role you defined. That is the safe job.
Safe does not mean unsupervised. I will not ship a system that auto-rejects at volume without a person who can explain why. The first pass can be automatic. The miss still needs a name next to it. If that sounds slow, try explaining a silent reject to counsel.
Write the role profile before you touch a model. Must-haves that are actually must-haves. Nice-to-haves that are not secret must-haves. If the team cannot agree on that page, the employee will encode the argument, and you will not see it until someone complains.
Why is hiring AI treated as high-risk?
Because the harm is a person, not a ticket. NYC LL144 already treats automated employment decision tools as something you have to audit and disclose in New York City. The EU AI Act puts many hiring systems in a high-risk class. You do not get to call this “just a helper” if it ranks or filters people.
That is not a reason to do nothing. It is a reason to keep a person on the decision, keep the criteria written, and keep a log of what the system saw. If you cannot replay a screen, you are not ready to go live.
I treat this with the same seriousness I treat money movement at 10mg. We have more than 6,000 providers and $3.4 million in loans. A wrong row there is a clinic and a patient. A wrong row in hiring is a career. Both need a name on the gate.
What does an honest first pass look like?
An honest first pass cites the line on the CV. “Four years in Nest and Postgres, listed under Experience.” Not “seems senior”. Not a culture fit score. If the fact is not on the page, the packet says “not stated”. That phrase is a feature. It stops the model from filling gaps with a vibe.
The output is a shortlist packet: who, the criteria they met, the criteria they missed, and the open questions. A recruiter should be able to disagree in a minute. If they have to re-read the whole CV to trust the packet, you did not save time. You added a layer.
I wrote a longer piece on keeping CV screens honest. The short version is the same as SmartComply: structured output, a person on the weird cases, and a miss log you can read. The model is not the hero. The loop is.
- Criteria written before any CV is scored
- Every flag cites a line or “not stated”
- No culture, accent, or photo signals
- A person reviews every yes and every close no
What must still go to a person?
Every yes. Every close no. Any candidate who is internal, referred, or already in process. Any case that smells like a legal risk: disability, leave, a complaint in the file. The employee can prepare the packet. It does not send the offer and it does not close the door on a near miss.
Escalation here is a state handoff too. The recruiter gets the packet, the criteria version, and why it is in their queue. Not a zip of CVs and a smile. If they have to rebuild the work, the employee failed.
Do not let “the AI said no” become a phrase in the building. If a hiring manager uses that sentence, you have already lost the plot. The decision is theirs. The tool is a first pass.
What should you not automate first?
Do not start with interviews. Do not start with offer letters. Do not start with a model that scrapes social profiles for “fit”. That is how you bake in junk and then spend a year unbaking it.
Do not start by ranking the whole market. Start with one role that has volume and a profile the team already agrees on. If you cannot name that role, you are not looking for an employee. You are looking for a story.
Skip video analysis and voice scoring. You do not need them for a first pass, and they are a magnet for bias claims. Paper, form, written criteria. That is enough work for a first month.
How do you measure without cooking the shortlist?
Time from application close to a shortlist a manager will open. Share of packets a recruiter did not rewrite. Agreement rate on a sample the team already labelled. Those are process numbers. They tell you if the pile moved.
Do not measure “hires made by AI”. A person made the hire. If you need a vanity number, you are in the wrong meeting. Watch for the failure mode where the shortlist gets quieter and quieter because the model learned your last bias.
Keep a holdout. Ten applications a week a senior recruiter screens blind. Compare. Write down the misses. If you cannot explain a miss, turn the thing down until you can. That is the same rule I use for document review. It applies harder here.
What records do you need before you go live?
The role profile and its version date. The fields the employee may read. The fields it must ignore. Who reviews yeses and close noes. How a candidate can ask what happened. Where the logs live. If that list is in someone’s head, you are not live. You are hoping.
Counsel should see this before volume. I am not a lawyer. I am the person who has to build a system that still makes sense when someone asks for the file. Buyers asked me that when we sold Insurpass. Regulators will ask it here.
If setup feels like a project, good. Hiring is not a weekend chatbot. 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 done.
Questions people ask
What can a recruiting AI employee safely do?
Read applications against a role profile you wrote, pull the facts that are on the page, and group them for a person. It can flag missing documents. It cannot decide who gets the job. If you cannot show the criteria, do not run the screen.
Why is this treated as high-risk?
Because a wrong no is a person who never gets a chance, and the law already noticed. NYC LL144 and the EU AI Act put automated hiring tools in a different bucket from a support bot. You need records, a review, and a human who can explain the outcome.
Should it send rejection emails?
Not at the start. A rejection is a decision. If you later send a template, a person should have marked the stage. Do not let the model invent a reason. “We went another way” is honest. A fabricated skill gap is not.
What about interviews?
Leave live interviews to people. An employee can schedule, send the brief, and collect a score sheet. It should not run a scored interview that becomes the hire decision unless counsel has signed off and a person still reviews every outcome.
How do we measure it?
Hours from close of applications to a shortlist a hiring manager will look at. Share of packets a recruiter did not have to rebuild. Miss rate on a held-out set you already labelled. Time-to-hire is downstream. Do not credit the model for a hire a person made.
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.
