← Guides

The agentic problem: how AI actually reads your CV

Your CV doesn't go to a person first

When you apply today, your CV usually doesn't land on a recruiter's desk first — it lands in an Applicant Tracking System, gets parsed into structured fields, and is scored against the job description before a human opens it. If you're applying to more than a handful of roles, some of those scoring passes are AI: a recruiter (or an automated tool acting on their behalf) drops a stack of CVs and a job posting into a model and asks it to rank them. Only the top few get a human look.

This isn't a conspiracy or a shortcut a lazy recruiter is taking. At volume, it's a genuinely reasonable way to triage hundreds of applications for one role. But it changes what "a good CV" means, and most advice on writing one hasn't caught up.

How the matching actually works

Underneath, CV-to-job matching in 2026 is a layered pipeline, not one model doing one thing:

  • Lexical baselines (TF-IDF, keyword overlap) — the oldest and crudest layer, still used as a floor because it's cheap and bias-free relative to what's built on top of it.
  • Semantic matching — transformer-based embeddings (Sentence-BERT-style models, and domain-specific variants trained on career data) that compare the meaning of your experience to the meaning of the job description, not just shared words. This is why a CV that says "led sprint planning and backlog grooming" can match a posting asking for "Scrum Master experience" even without the exact phrase.
  • Skills ontologies — structured taxonomies (like ESCO or O*NET) that map thousands of skill terms to a shared vocabulary, so "Kubernetes" and "container orchestration" resolve to the same underlying concept.
  • LLM re-ranking — increasingly, a final pass where a language model reads the shortlist and produces a score, a gap analysis, or both.

None of this is exotic research anymore — it's shipped, production infrastructure at the ATS and recruiter-tooling level.

Why "they'll infer it" doesn't work anymore

Here's the part that actually changes how you should write: a human reader fills in gaps automatically. If your CV says you were a Scrum Master, a human recruiter assumes you supported a Product Owner, ran ceremonies, managed a backlog — they don't need you to spell it out, because they already know what the role implies.

A matching system doesn't do that. It scores what's literally present against what's literally asked for. If the job posting lists "stakeholder management" as a requirement and your CV never uses language that maps to it — even though you obviously did it, because you can't be a Scrum Master without doing it — you show up as a gap. Not because you lack the experience, but because the experience isn't in the data.

This is the actual mechanism behind a frustrating, increasingly common experience: getting fewer interviews despite having more experience, because a CV written for a human's ability to infer is being read by a system that can't.

The bias problem is real, and it's regulated now

AI-based screening doesn't just risk being incomplete — it has a documented history of being unevenly wrong. Academic audits of LLM-based resume screening have found real, measurable bias that has actually flipped direction across model generations (earlier models showing one demographic pattern, newer models showing a different one) — which is itself the problem: the bias moves, but it doesn't go away just because a newer model is used.

Regulators are responding. The EU AI Act classifies AI systems used to filter or rank job candidates as high-risk (Annex III), which brings real compliance obligations for any company deploying them, with a compliance deadline at the end of 2027. New York City's Local Law 144 already requires an independent annual bias audit for any automated tool used in hiring decisions within the city. This is a fast-moving legal area, not a settled one — treat anything you read here (including this page) as a starting point, not final legal guidance.

What this actually means for your CV

None of this is a reason to write your CV for an algorithm at the expense of writing it for a human — the good news is that the same fix helps with both. What it does mean:

  • Say the thing, don't imply the thing. If you did stakeholder management, write "stakeholder management." Don't rely on your job title to imply it.
  • Name the specific technology, tool, or method, not just the category. "Cloud infrastructure" matches less than "AWS, specifically Lambda and DynamoDB" — and a human reader loses nothing from the specificity either.
  • Completeness matters more than it used to. A tight 2-page CV that cuts everything but the highlights was optimized for a human's limited reading time. A system with no reading-time constraint doesn't reward that cut — it just sees fewer data points to match against.
  • You can't do this by hand for every application. Rewriting your entire history to be this explicit, for every role, isn't a realistic ask of your evenings. This is exactly the gap a tool that holds your complete structured story — and can generate a tailored, complete version on demand — is built to close.

The underlying shift is simple to state and annoying to live through: the audience for your CV quietly changed from "a person with context" to "a system with none." Writing for that audience isn't harder, exactly — it's just more literal, more complete, and not something most people have been taught to do.

Related guides