Work With Us
Video 380

Never Screen A CV Again With This AI Agent

Length: 26:13 Chapters: 23 Published: 22 Oct 2026

Using a Claude agent to score each applicant's CV from 1 to 10 against the job description and write the reason in monday.com: the boards, the legal question, instructions, setup, a first-run snag and the results, $0.46 for two applicants. The full written breakdown is below the video.

How Do You Screen CVs With AI In monday.com?

Keep vacancies and applicants on two boards, then have an AI agent score each applicant's CV against the job description of the vacancy they applied for. Ours are monday.com boards, but the agent would work with any system.

Whoever is hiring adds the vacancy with its name, employment type, closing date and salary range, and uploads the job description. That's the most important part, because every CV is measured against it.

Our demo job description, for a monday.com implementation consultant, is basic: a section about the business, the role and what it involves, the requirements, and some nice-to-haves.

Why Should Applicants Apply On Your Own Website?

Because there's no direct integration to get applicant data out of LinkedIn or Indeed and into monday.com or any other system. If the apply button on the job ad takes people to a form on a page you host, each submission adds the applicant straight to your applicants board.

Applications are much like a lead funnel, where you want to filter out poor fits, and a job posted online can bring in an overwhelming number of them. Your own form puts the data for every applicant in one place.

What Should A Job Application Form Capture?

The applicant's name, current role, email, phone number and location, the vacancy they're applying for, and above all their CV, uploaded on the form. Each submission adds all of it to the applicants board.

The vacancy matters because you may have several open at once, and the agent needs to know which job description to score each CV against.

How Does An AI Agent Score Each Applicant?

It compares the CV with the vacancy's job description and picks a match score from a 1 to 10 dropdown, where 1 means they won't be hired and 10 is a perfect match. It writes a match reason explaining the score, then moves the applicant's status from New to Screened.

Those two data points, a score and a reason, are what you need whatever system you use. You could add a rule that moves anyone under 5 to rejected and sends a rejection email, but that's not part of this build, and anyone who passes your threshold still goes to a manual review.

We find checking CVs the most painful part of hiring, and the time this saves grows when you get hundreds of applicants per job description and have several vacancies open.

What Should The Instructions For A CV Screening Agent Say?

Where the applicants are, what to compare and what to write back. Ours tell the agent to review each applicant's CV on the applicants board against the vacancy's job description, give a match score from 1 to 10 and a match reason, then change the status from New to Screened.

We asked for reasons short enough to scan quickly, with 10 for the best applicant and 1 for the worst, judged entirely on the job description. We also pointed out that middle scores of 4 to 6 are the hardest to justify.

Don't write the system prompt yourself, as it probably won't give the best result. In the Claude Console, press Create agent under Agents and let the summary function write it from your instructions. There's no dictate option there yet, so we talked the job through in a separate Claude chat and had Claude write the instructions.

Which Model & Tools Does A CV Screening Agent Need?

Sonnet and one connection, to monday.com. Here's how we set it up:

  1. Model: we chose Sonnet, on low or medium effort. It only compares a CV with a job description, and a person reviews the strong matches anyway, so it doesn't need much compute, and Opus or a higher effort level would cost more per run.
  2. Generate: with the instructions pasted into the summary function, press Generate. It wrote the agent's name and description as well as the system prompt, which came out shorter than the original instructions but not as condensed as we'd hoped.
  3. Tools: press Add MCP server and choose monday.com, where the applicants and vacancies both live. Press Connect, create a vault for the credentials or reuse one you have, then authorise the connection in monday.com until it shows as connected.
  4. Permissions: the monday.com MCP lists every action the agent could take, and Always deny blocks one outright, such as creating items and tasks. Always ask makes it wait for your permission every time, which takes away its autonomy, and we're wary of Auto, which decides for itself, so on our demo account everything is on Always allow.
  5. Advisor: no skills, workflows or sub-agents are needed. For applications that are a mess, far too long or empty, you could add an advisor, a bigger model than the agent's own, such as Opus, that a confused agent can escalate to and be told what to do.
  6. Check the model again before pressing Create agent, as ours wasn't on Sonnet when we got there.

How Do You Run The Agent For Every New Applicant?

Trigger it when a new applicant is created on the board. That needs a webhook through the API, which we didn't set up in the video, so we ran it manually as part of testing. The alternative is a schedule, every minute, hourly or daily, that screens whoever is in the New group.

What else it needs, in the order we set it up:

  1. Environment: let Anthropic host it in the cloud, which we'd strongly advise over hosting it yourself. Screening CVs isn't high risk, so you could limit networking to MCP servers, since that's all it should need, or leave it unrestricted, as we do on our demo account.
  2. Deployment: ours is called CV review deployment, set up under Deployments on the agent's page, with that environment, the vault holding the monday.com connection and "Please go to work" as the initial message.
  3. Budget: each session can have one, $2 for example, and a session that reaches it before finishing just stops. While you're still testing and refining the instructions we'd leave it off, and set it once the agent is live and you know what a run should cost.

What Went Wrong On The First Test Run?

The agent said neither applicant had a CV, when both had a PDF uploaded. We'd put two test applicants in the New group with no score or reason, pressed Run now and were watching the session live when it told us.

We replied in the session that both had a PDF, and it carried on. We think it searched for the PDF by its file name instead of looking in the CV column, so the fix is to say in the instructions exactly where each applicant's CV is.

This is why we test agents: the first run shows what the instructions missed.

How Much Does AI CV Screening Cost?

In our test it cost $0.46 to screen two applicants, in a run of about a minute that the snag made longer. Each applicant ended up with a match score, a match reason and the status Screened.

One test applicant scored 9 out of 10 for meeting every must-have and all four nice-to-haves. The other scored 2: their strongest match was using monday.com with basic automations and one dashboard for their own freelance business, against gaps in client-facing consulting, discovery workshops and data migration, and only single-user monday.com experience.

We'd written that second CV to suggest monday.com experience and throw the agent off, and it wasn't fooled. Both scores and reasons appeared on the applicants board in monday.com, and in the test nobody had to open a single CV.