← Blog
ai-toolsproduct

Mistral AI in recruiting: how Umamy reads candidate resumes

Mistral AI in recruiting: at Umamy, Mistral OCR reads the resumes and documents candidates upload, scans included. What it does, and what it does not do.

Jonathan BouazizCo-founder, Umamy · · 6 min read

At Umamy, Mistral AI has one specific job in recruiting: reading the documents candidates upload, a resume, a cover letter or notes, and turning them into text. The model doing the reading is Mistral OCR. It also reads scans and photos, where basic text extraction finds nothing at all.

Tuesday evening. Karim, a maintenance technician with eight years on the job, uploads his resume to Umamy. It is a PDF built from an online template: two columns, icons instead of section titles, a progress bar for every skill. He adds a phone photo of his electrical safety certification. A recruiter gets the picture in ten seconds. Software is another story.

Mistral AI in recruiting: what does Mistral do at Umamy?

A document uploaded by a candidate goes through three steps.

1. The candidate uploads a document

Resume, cover letter, notes: the file goes in as it is, PDF or image, with no reformatting or retyping.

2. Mistral OCR reads it

Mistral OCR, Mistral AI’s document-reading model, reads the file and returns its text while keeping the reading order, the columns and the tables. A scan or a photo is read just like a PDF.

3. Umamy works from the text

The extracted text is passed to Umamy’s system. Mistral plays no further part.

So Mistral OCR only shows up at step 2. It is not the core of the product: it is a reading tool, placed at the entrance of the flow.

Why a resume PDF is hard for software to read

A PDF mostly describes where each character sits on the page, not the logical order of the text. And resumes are among the hardest documents to read: they are designed to stand out visually, not to be parsed by software.

Naive text extraction often trips over:

  • columns, merged line by line, so skills end up glued to work history;
  • icons and skill bars, which contain no text at all: a phone icon or a bar filled to 80% means nothing to a machine;
  • tables, where dates, employers and job titles lose track of each other;
  • scans and photos: a diploma, a certification, a resume printed then scanned. With no text layer, there is nothing to extract.

Take Karim’s file. With naive extraction, his employment dates end up stuck to the skills column, and the electrical certification, which only exists in the photo, vanishes. Some of the text is there, but eight years of experience become hard to piece together, and the certification that set him apart is nowhere to be found.

Reading a document while keeping its structure is exactly what document recognition (OCR) models are built for. Mistral AI, a French company founded in April 2023 and headquartered in Paris, has made it one of its areas of focus: its Mistral OCR model, announced in March 2025, reads images and PDFs while keeping tables and layout intact.

A badly read resume is an invisible candidate

Whatever the tool, ATS, job board or AI hiring tool, the first step is always the same: reading the document. If it fails, everything downstream works on incomplete text. A missing certification, a misattributed date, a skill lost in a column: the candidate never knows, and neither does the recruiter.

Three questions to ask any tool that reads resumes:

  • How does it read a two-column resume with icons?
  • What does it do with a scan or a photo?
  • What happens when reading fails: is the document flagged, or silently skipped?

On the hiring side, the same standard applies to the job description: clear criteria make for a clear comparison. Related reading: why your hiring data should be reachable from any AI.

📄 A well-read resume is not enough: you also need clear criteria to compare it against. The job brief template designed to be understood by an AI: sections, must-have and nice-to-have criteria, filled-in examples.
The AI-readable job brief template

Candidate data: what reading a resume involves

A resume holds personal data: name, contact details, career history, sometimes a photo. Reading it, even just to extract the text, counts as processing under the GDPR.

Four questions to ask any tool that reads candidate documents:

  • Who reads the document, and through which provider?
  • What is the extracted text used for?
  • How long are the document and the text kept?
  • How can the candidate access their data or ask for it to be deleted?

At Umamy, reading with Mistral OCR has one purpose only: turning the document into text. Umamy is currently filing for AI Act and GDPR compliance.

What Mistral does not do at Umamy

“AI partner” can suggest much more, so it is worth stating plainly.

  • Mistral does not search for candidates. Sourcing is handled by Umamy’s agents.
  • Mistral does not score profiles. Ranking relies on Umamy’s Dynamic Fit Score, a context-enriched scoring built on accumulated context (scorecards, decisions to keep or set aside a profile).
  • Mistral does not message anyone. Messages to candidates go out in the name of the founder or hiring manager, after approval.
  • Mistral makes no hiring decision. The agents do the work; people decide.

In short: Mistral OCR reads a document, once, when it is uploaded.

A tool picked for a task, not for show

Umamy puts its agents together like a team: each step uses the component suited to its task. Reading a formatted document is a problem of its own, with its own tools. Choosing that component separately means documents are read better, without changing how Umamy prepares and runs a search.

What makes a search good in Umamy is not the model that reads a file. It is the context that builds up around each role: criteria, feedback on profiles, hiring history. We call this context enrichment.

Frequently asked questions

Does the candidate need a Mistral account?

No. Document reading is built into Umamy. The candidate uploads a file and the rest happens without any setup on their side.

Is Mistral used to choose candidates?

No. Mistral OCR reads documents and extracts their text. Search, profile ranking and messages are handled by Umamy, and every message waits for human approval.

Can Mistral OCR read a scanned or photographed resume?

Yes. Mistral OCR reads images as well as PDFs. A clean scan reads well; a blurry or skewed phone photo, much less so.

Is Mistral AI the only AI model Umamy uses?

No. Mistral OCR reads the documents candidates upload. The other steps, from preparing the search to writing messages, rely on other components run by Umamy.

What Umamy does next

On the hiring side, Umamy’s agents map the profiles that fit the role, rank them with the Dynamic Fit Score and contact them in the name of your founder or hiring manager, after your approval. First candidates arrive 10 minutes after the brief. Pricing is flat, with no success fees: from €249 per month, excluding VAT. To see how the agents work, explore AI agent candidate sourcing. Umamy is currently filing for AI Act and GDPR compliance.

📄 Before you open your next role. The brief template that gives AI clear criteria, ready to fill in.
The AI-readable job brief template

Sources: Mistral AI, Mistral OCR announcement (March 6, 2025); Mistral AI, Wikipedia

See who Umamy finds on one of your roles.First candidates in 10 minutes, scored on evidence, not on the CV.Talk to our team
Jonathan Bouaziz is co-founder of Umamy, the agentic hiring platform where AI agents do the legwork and your team makes the hires.