Resume parsing software reads resume files (PDF, Word, sometimes scans) and turns them into structured fields such as roles, dates, education and skills, so an ATS or screening tool can search and filter them. To choose one, test it on your own CVs and compare five things: which layouts and languages it reads, field accuracy, output format and ATS connection, data protection, and cost per CV including the cost of errors. Status: November 2026.
What does "parse a resume" mean?
To parse a resume is to read the file and sort its content into fields: name, contact details, each job with employer, title and date range, education, languages, skills. The result is data instead of a document. Parsing does not judge whether someone fits a job; that is a separate step. How the two differ, step by step, is covered in CV parsing: how it works and where it fails. This page is about the buying decision: which software, and how to check it.
Which kinds of resume parsing software exist?
| Kind | What it is | Fits when | Watch out for |
|---|---|---|---|
| ATS built-in parser | The parser that comes with your applicant tracking system | You want the leanest setup and your CVs are mostly clean files | You usually cannot switch it or see its accuracy; ask how it handles columns and scans |
| Standalone parser or API | A separate service you call with a file and get structured data back | You build your own workflow, a talent database or a career site | You carry the integration, the data-processing contract and the monitoring |
| AI screening tool that reads the CV itself | Reading, evaluating and a shortlist in one step, often with a language model | Your problem is shortlisting at volume, not data entry | Make sure the rating shows evidence per requirement, see CV screening software |
No source we found compares the accuracy of built-in and standalone parsers on the same files, so treat any claim that one kind is "more accurate" as unproven until you test it. Many vendors now also use language models for the reading step; that changes how they cope with unusual layouts, but not the need to measure.
What our own test showed
Every parser starts by getting text out of the file. We built four test CVs for the same fictional person (German names with ö, ß and Ä, month/year dates) and ran the text layer through three common open-source readers, pypdf 6.14.2, PyMuPDF 1.28.2 and python-docx, on 8 November 2026. Commercial parsers add layout analysis and OCR on top, so this is not a vendor ranking; it shows which file traits make the first step fail.
| Test file | What the text reader returned | What it means for a parser |
|---|---|---|
| Single-column PDF | Complete, in order, umlauts intact | The easy case; every parser should pass |
| Two-column PDF (sidebar and experience) | Rows were merged line by line: "Kontakt", then "03/2021 – heute", then an e-mail address, then "Teamleiter Produktion" | Dates and employers can end up on the wrong job unless the parser detects columns |
| Word file with a table layout | The normal paragraph reader saw only the heading; dates, titles and employers sat inside table cells | A parser that reads only paragraphs loses the whole work history |
| Scanned PDF (image only) | Empty text | Without OCR the CV counts as blank, and a blank record can look like "no experience" |
A fifth trap is invisible on screen. The same word "Müller" can be stored as one character (ü) or as "u" plus a combining mark. The two look identical but are not equal in a string comparison, as the Unicode normalization rules describe. In our check the two forms had 6 and 7 characters and did not match until normalized. If a parser or a duplicate check does not normalize, the same person or the same employer can be counted twice.
The test files are small enough to rebuild yourself; the point is the method, which you can apply to your own CVs in the next section.
How to test a parser with your own CVs
- Collect 30 real CVs from your last hiring rounds (remove personal data or get permission): clean ones, two-column designs, Word with tables, scans or photos of paper, English and German, one very long one.
- Define the fields that matter for your jobs: employer, title, start and end date, degree, languages, key skills.
- Run every file through each tool you are testing and write down, per field, whether it is right, wrong or empty. Wrong is worse than empty: it looks like data.
- Count errors per field, not per CV. A parser that gets names perfect and dates wrong is dangerous for experience filters.
- Check what happens on failure: does the system flag a file it could not read, or silently store a blank?
What should you compare?
| Criterion | Question to ask | Why it matters |
|---|---|---|
| Input coverage | Which formats, layouts and scans, which languages? | Your applicants do not send template files |
| Field accuracy | Measured on my 30 CVs, per field | Published percentages come from the vendor's own test set |
| Failure handling | Is an unreadable file flagged? | Silent blanks cause wrong rejections |
| Output | Fixed schema, API or ATS connector, can you correct a field? | A fix in the UI should not be overwritten on re-import |
| Data protection | Hosting location, retention, deletion, sub-processors, training on your data? | Needed for the contract, see below |
| Price model | Per CV, per user, per volume tier, extras for OCR or API? | Compare on your annual volume |
| Updates | Are model updates announced, can accuracy drift? | Accuracy can change without notice after a vendor update |
What does it cost: price per CV and price per error
Per-CV prices vary and are often quoted on request, so we do not print vendor prices here. What you can calculate yourself is the price of an error. An illustrative example, with assumptions you should replace: a person needs 2 minutes to fix a wrongly parsed CV, and time costs 40 € per hour (0.67 € per minute).
| Share of CVs needing a fix | Cost of fixing per CV | Per year at 3,000 CVs |
|---|---|---|
| 5 % | 0.07 € | about 200 € |
| 10 % | 0.13 € | about 400 € |
| 25 % | 0.33 € | about 1,000 € |
Two points follow. First, at small volume the accuracy difference can matter more than a few cents of licence price. Second, the table only counts errors someone notices. The costly ones are silent: a date on the wrong job, a scan stored as blank. They never get fixed; they get a good candidate filtered out, which is why a field test beats a price comparison.
What do you need to check in Germany and the EU?
- Data processing contract. A parser run by an external provider processes applicant data on your behalf; GDPR Art. 28 requires a contract with the provider. Ask for hosting location, sub-processors and deletion.
- Data minimisation. German CVs often carry a photo, date of birth and marital status. Under GDPR Art. 5(1)(c) you should only keep what you need. Decide which fields your parser may store: § 7 AGG bans disadvantage on protected grounds, and stored fields like age or photo make that harder to defend (burden of proof, § 22 AGG).
- Works council. In Germany, technical systems that can monitor behaviour or performance trigger co-determination, § 87 (1) no. 6 BetrVG. Check early whether a parser or its scoring counts.
- AI Act. Systems that filter or evaluate applications fall under Annex III of Regulation (EU) 2024/1689 (high risk); the obligations apply from 2 December 2027 (Regulation (EU) 2026/1744). Pure field extraction is not a rating, but as soon as a score is added it can be, see EU AI Act in recruiting.
This is orientation, not legal advice.
When parsing is not enough
A parser gives you clean data and no shortlist. If your real bottleneck is too many applications, you need a step that compares each CV with the requirements of the job. In Sprad's screening, reading the CV is part of the evaluation rather than a separate product: the CV check costs 2 credits per application, every requirement is marked must, nice-to-have or knockout, and each rating is backed by the exact sentence from the CV. A person confirms every rejection, and results go back into your ATS (Personio, Ashby and d.vinci are documented connections). Details on the screening product page and the pricing.
Notice: Sprad is the vendor of this site and of the product mentioned. We did not rank other tools here; the test above describes file traits, not products.
FAQ
What is the best resume parsing software?
There is no single best one. The best parser is the one with the lowest error rate on your own CVs, in your languages and layouts, at a price you can afford at your volume. Run the 30-CV test with two or three tools.
Is an ATS parser good enough?
For clean, text-based files often yes. Check how it deals with columns, tables and scans on your sample, and whether you can see which files it failed to read.
How accurate is resume parsing?
Vendors publish figures between 90 and 99 percent, but on their own test sets. Treat them as a starting point and measure per field on your own files.
Can a parser read scanned resumes?
Only with OCR (text recognition on images). Without it, as in our test, a scan returns empty text. Ask whether OCR is included or an extra.
Does a parser reject candidates?
No, parsing only extracts data. Rejections come from rules or ratings applied afterwards. GDPR Art. 22 limits decisions based solely on automated processing; see Can AI reject candidates automatically?.
Do I need parsing and screening?
For a shortlist, yes, but they can be one product. Ask the vendor which step produces the score and whether you can see the evidence.







