Resume parsing: what it extracts, what breaks it, and what the regulators say about matching

Most large employers read a CV twice: once by machine, to turn it into database fields, and once by a person, if it survives the first reading. Knowing what the first reading does explains a great deal about why good candidates disappear.

This article covers resume parsing: what it extracts, what breaks it, how parsing differs from the matching and scoring built on top of it, what the UK guidance says about the risks, and what employers and candidates should each do. New Dynamics does not include an applicant tracking system, so this is general background and names no products.

Parsing ends at the offer; what follows decides whether the hire works. See how New Dynamics onboarding gives every new starter the same first weeks.

What resume parsing is

The UK government's guide to responsible AI in recruitment defines the category in one sentence, under the heading CV matching: "also known as resume parsing or CV scraping, these systems are a type of screening tool that extract key data from CVs to find semantic similarities between applicants and pre-determined ideal CVs."

It is worth separating two things that sentence puts together:

• Parsing is extraction. Reading a document and filling in fields.
• Matching or scoring is judgement. Comparing those fields against a job, a set of keywords or a "pre-determined ideal".

Parsing is a data quality question. Matching is a fairness question. Most articles about "beating the ATS" conflate them, and most procurement decisions do too.

What a parser extracts

Typically:

1. Identity and contact details. Name, email, phone, location.
2. Employment history. Employer, job title, dates, and sometimes the description under each role.
3. Education and qualifications. Institution, course, level, dates, and professional registrations.
4. Skills and keywords. Either a list the candidate wrote or terms recognised from the body text.
5. Dates and derived values. Length in each role, gaps between roles, total experience.

That last group matters more than candidates expect, because it is computed rather than stated, and the government's guide notes that a sifting tool "may also consider gaps in employment in a way that disproportionately affects parents, people with care-giving responsibilities, people with disabilities or long-term health conditions and neurodivergent candidates".

What breaks it

Parsers read structure. Anything that carries meaning visually rather than structurally is at risk:

• Tables and columns. A two-column CV can be read in the wrong order, interleaving a skills sidebar with employment history.
• Text inside images. A name, a header or an entire CV saved as an image extracts as nothing.
• Headers and footers. Contact details placed in a document header are sometimes skipped entirely.
• Unusual headings. "Where I've worked" instead of "Experience" may not be recognised as the employment section.
• Inconsistent dates. Mixing "Mar 2019", "03/19" and "Spring 2019" produces gaps that were never there.
• Decorative fonts, icons and text boxes. Graphical elements that a person reads easily and a parser does not.

None of this is the candidate's fault, and a good employer treats a mangled record as a system problem rather than a signal about the applicant.

Parsing, matching and the regulator's concern

Extraction rarely causes unfairness. What sits on top of it can.

The responsible AI guide notes that organisations use AI-enabled technologies "across every stage of the recruitment process, including: Sourcing, Screening, Interview, Selection", and that "at all stages there is a risk of unfair bias or discrimination against applicants". It also identifies "a risk of digital exclusion for applicants who may not be proficient in, or have access to, technology due to age, disability, socio-economic status or religion".

On CV matching specifically it warns that these tools "risk perpetuating existing biases – for example reinforcing patterns of employment where certain groups are underrepresented (ethnic minorities, women, disabled people) – particularly in sectors that are less diverse, such as engineering, policing, and construction". The mechanism is easy to miss: if the "pre-determined ideal CV" is built from the people you already hired, the tool reproduces your existing workforce with more efficiency than before.

The guide's own worked example shows what a careful buyer does. An organisation procuring a screening system that scores applicants on CVs and cover letters asks the supplier for evidence of a bias audit. The audit evaluates performance by sex, ethnicity, age and disability, and finds "applicants with disabilities or who are neurodivergent may be disadvantaged by the system". The supplier states this as a limitation, and the buyer proceeds but routes applicants who declare a disability through manual review instead of scoring. It also adds that "bias audits should be repeated at regular intervals after the system is in operation".

Our ATS post sets out the ICO's six questions to ask a provider before procuring an AI recruitment tool, which apply here directly.

Before you switch parsing and matching on

1. Separate the two features. Parsing to populate a record is a different decision from scoring to rank people. Buy and approve them separately.
2. Ask for the bias audit, by protected characteristic, and ask when it was last repeated.
3. Decide the human step. Who looks at what, at which stage, and what happens to candidates the system ranks low. A route to manual review is the guide's own suggestion.
4. Tell candidates. Transparency is a data protection requirement, not a courtesy, and it is also the only way someone can flag that their CV parsed badly.

The data protection layer

Recruitment generates more personal information than most employers realise.

The ICO's draft guidance on recruitment and selection is aimed at employers and at organisations recruiting on their behalf, and covers "all potential employment relationships, including employees, contractors, volunteers or gig and platform workers". It notes that recruiters process information about "candidates, prospective candidates, temporary workers, contractors, referees, emergency contacts, and dependants", some of which "may be sensitive and include, for example, details about health, diversity, or criminal convictions". And it makes the point that matters here: "the use of novel technologies in recruitment processes means that organisations are processing increasingly large amounts of information about people."

Practical consequences: know how long parsed records are kept, know who can see them, and remember that a parsed CV creates a searchable database of people who never became employees. This is general information rather than advice on your circumstances.

What candidates should actually do

The advice is dull and it works.

• Use one column, with standard headings: Experience, Education, Skills.
• Real text, not images. Check by selecting the text in your CV; if you cannot select your name, neither can the parser.
• Consistent dates in one format throughout.
• Put contact details in the body, not in a document header.
• Use the words in the advert, where they are honestly true of you. This is not keyword stuffing; it is answering the question as asked.
• Keep a plain version alongside a designed one, and send the plain version to portals.

GOV.UK's guidance on publishing accessible documents is the same advice from a different direction: publish in HTML "wherever possible", because PDFs "can make your content harder to find, use and maintain" and "do not work well with assistive technologies like screen readers a lot of the time", keep the language and structure simple, and explain abbreviations the first time you use them. A CV that parses well is usually a CV that a screen reader handles well, which is a better reason to write it that way.

What parsing cannot tell you

• Whether someone can do the job. It extracts claims, it does not verify them.
• Why a gap exists. Only a conversation does that.
• Whether unusual experience is relevant. Semantic similarity to an ideal CV punishes the career that took an interesting route.
• Whether the person is any good. Which is why the shortlist is a hypothesis, not a verdict.

Frequently asked questions

What is resume parsing?

Software reading a CV and turning it into structured data (name, contact details, employers, job titles, dates, education, skills) so that a recruitment system can store, search and compare it. The UK government's responsible AI in recruitment guide describes such systems as screening tools that "extract key data from CVs to find semantic similarities between applicants and pre-determined ideal CVs", noting that resume parsing and CV scraping are other names for the same thing.

What is the difference between parsing and matching?

Parsing extracts information; matching or scoring makes a judgement with it. Parsing problems show up as wrong data in a record. Matching problems show up as the wrong people being shortlisted, which is where the UK guidance identifies a risk of perpetuating existing bias, especially in sectors that are already less diverse.

What CV formatting breaks a parser?

Tables and columns, text inside images, contact details in a document header, unusual section headings, inconsistent date formats, and decorative elements such as icons and text boxes. A single-column layout with standard headings, selectable text and one date format avoids nearly all of it.

Does resume parsing reject candidates?

Parsing itself does not; it populates fields. Rejection comes from the screening, scoring or filtering applied afterwards. That distinction matters when choosing software, because the two functions carry very different risks and should be approved separately.

What should employers check before using CV matching software?

Ask for evidence of a bias audit covering sex, ethnicity, age and disability, and when it was last repeated; decide where a human reviews candidates and what happens to those the system ranks low; and tell candidates that a tool is being used and how to challenge an automated decision. The government's own example routes applicants who declare a disability through manual review rather than scoring.

Is CV parsing a data protection issue?

Yes. The ICO notes that recruitment involves information about candidates, prospective candidates, referees and others, some of it sensitive, including details about health, diversity or criminal convictions, and that new technologies mean organisations are processing increasingly large amounts of information about people. Retention, access and transparency all need deciding before the feature is switched on, and this is general information rather than legal advice.

Your next step: separate the two decisions

1. Decide whether you are buying extraction, judgement, or both, and approve them separately.
2. Ask any supplier for a bias audit by protected characteristic, and its date.
3. Write down where a person reviews candidates and what happens to low-ranked applicants.
4. Tell candidates what is used and how to raise a problem with a parsed record.

For the wider question of using AI in hiring without losing fairness or defensibility, read and download our AI recruitment guide. The guide is free to read, and the PDF uses our short download form.

Want the first weeks after the offer to be as well designed as the screening before it? Book a New Dynamics demo and bring your current starter checklist. You can also email contact@new-dynamics.com.