Build an AI Agent That Screens Every Job Applicant For You

Your job post did not get 214 replies. It got 214 unread emails wearing a disguise.

It is Tuesday morning in Manchester. Priya runs a six-person physiotherapy clinic and posted for a front-desk coordinator three days ago. She now has 214 applications sitting in one folder, and if she is honest with herself, she has properly read four of them.

She is not lazy. She is not disorganised. She is one person with a clinic to run, a waitlist to manage, and a stack of CVs that all start to look the same by application forty.

This is not a hiring problem. It is an unbuilt inbox.

Chatbot or Agent? The Difference That Actually Matters

Here is the distinction the YouTuber and "buy back your time" author Dan Martell uses in a guide that has quietly passed 199,000 views: a chat is a meeting. An agent is an employee.

With a chat, you ask a question and you get an answer. You still have to do something with it. With an agent, you hand over an outcome, and it runs the whole loop on its own: it diagnoses the problem, assembles a plan, takes the action, and assesses whether the result actually landed. Martell calls it the DATA loop, and the part that matters for a small business is the last letter. Without that self-check, you just built an automation that quietly breaks the first time something unusual comes in. With it, you built something closer to a junior teammate.

An applicant inbox is a textbook case. It is repetitive. Every resume takes roughly the same shape. It is rules based: the same job description, the same must-haves, the same deal-breakers, every single time. And it fails Martell's third test, return on time, in exactly the way that makes it worth automating: reading and replying to 214 applications by hand costs hours you do not have, and building the agent once costs an afternoon.

Why This Is Suddenly Something You Can Actually Do

Two years ago, this meant hiring a developer or learning to write code. That is no longer true, and it is not marketing spin, it is a function of what the tools you already have can now do on their own.

General-purpose assistants like Claude and ChatGPT can now hold a long-running job, not just answer one question. New open-weight models are pushing the cost of running an agent down hard enough that Moonshot AI's Kimi K3, released as a direct coding-agent challenger this month, is already being tested by builders as a cheaper engine for exactly this kind of always-on task. You do not need to touch Kimi K3 to build a hiring agent. You need to know that the underlying cost of "have software watch my inbox and act on it" has fallen off a cliff, and that is why guides on building your first agent are suddenly everywhere on X and YouTube at once.

Google's own AI Overview, when you ask it directly how to screen applicants with AI, now sketches out the same architecture independently: a trigger that fires when a new application lands, a defined scoring rubric, and a routed output back to a dashboard you actually check. That is not a coincidence. It is the same pattern converging from three directions at once, because it is simply the correct shape for the job.

Isn't This Just an Autoreply?

No, and the difference is the part most guides skip. An autoreply sends the same message to everyone. An agent reads the resume, checks it against your actual rubric, decides which of three or four responses fits, drafts that response in your voice, and flags the ones that do not fit any bucket for a human to look at. One is a form letter. The other is a judgment call, made consistently, at 2 a.m., without you in the room.

The risk is real if you skip the guardrails, which is exactly what the next section is for.

How to Build Your First Hiring Screening Agent, Step by Step

You can build this in an afternoon inside Claude, ChatGPT, or any assistant that can hold a long conversation and connect to a document or inbox. Follow the framework in order.

1. Aim it at one outcome. Not "handle my hiring." Something you can picture finished: "every applicant gets a same-day response, the top five candidates are flagged with a one-line reason why, and I never open a resume I was always going to reject." If you cannot say it in one sentence, you are not ready to build.
2. Give it an identity. Out of the box, a model knows a little about everything and nothing about your clinic, your shop, or your agency specifically. Write three short plain-English files and let the AI draft them after you answer a few questions: a soul file (how it talks: warm but brief, never robotic, never "I hope this finds you well"), an identity file (its job title, for example Senior Talent Screener, and its one lane: resumes and first replies, nothing about payroll or contracts), and a user file (who it works for: your business, the specific role, your three must-haves and two instant no's). Martell cites a case where an airline support agent's success rate dropped from 33% to 11% the moment its identity file was stripped out, same model, same task, just no job description. An agent without an identity is a genius sitting at an empty desk.
3. Equip it with a real rubric. This is where most people stop too early. Give it the actual job description, not a summary. Then define a numeric scoring rubric the way Google's own AI Overview now recommends for this exact task: a 0 to 10 score for relevant experience, a 0 to 10 score for the specific skill the role needs most, and a short written reason for every score, so you can audit any decision it makes. Feed it a few resumes you would have hired and a few you would have rejected, and ask it to score those first as a calibration check before it touches anything live.
4. Narrow its scope. Do not build one mega-agent that screens, schedules, and negotiates salary. Build a screening specialist that only reads and scores, and a scheduling specialist that only books interviews for candidates the screener has already flagged. Put a manager agent above both, whose only job is to route the work and report back to you. One agent, one lane, same as Martell's own setup for his content team.
5. Trust it in stages. Week one, it only drafts replies and you approve every send. Week two, it auto-sends the polite "thank you, not a fit this time" replies and still queues anything borderline for you. Only in week three or four, once you have checked its calls against your own judgement, does it start booking first interviews on your calendar without a human touching the invite first. You never hand over the keys on day one. You loosen the leash once it has earned it.

Notice what did not change from Priya's inbox problem to your inbox, your invoice queue, or your review requests: aim, identity, equip, narrow, trust. That is the shape of every agent worth building, whichever part of the business it is watching.

What You Have After This

In a week, you have an agent drafting a same-day reply to every applicant, something almost no small business does today, and it is already the difference candidates notice and mention to other candidates.

In a month, you have stopped opening the folder at all. You get a Friday note with a ranked shortlist of three names and a one-line reason for each. That is the only hiring email you read.

In six months, the same pattern is running your invoicing follow-ups, your review requests, or your appointment reminders, because you already know how to build one. Same framework. Same afternoon of setup. A completely different business underneath it.