AI transparency notice
Staffholm and Rolefit are brands of Sirin OÜ, a company registered in Estonia (registry code 16315993). Staffholm is the service employers use, and Rolefit is the name candidates see when they apply. Wherever this document says “Staffholm”, “Rolefit”, “we” or “us”, it means Sirin OÜ.
What the software does
- Reads the CV and extracts what it says, meaning employment, education, skills and evidence for each.
- Compares that against the requirements in the employer's open adverts and finds what the CV does not answer.
- Asks the candidate about those gaps and only those. An AI model writes the questions. When an employer forwards a CV or adds it by hand, no questions are asked.
- Scores the fit and writes a short explanation of each score.
What it does not do
It does not decide anything. It produces an ordered list and reasons. It cannot reject, shortlist or hire. Every outcome comes from a person at the employer.
This is not a policy choice we could quietly reverse. There is no automatic rejection anywhere in the product, and the employer's terms forbid using the output as an automatic filter. When an employer closes a job, the service can send the employer's own answer to the people still waiting. A person at the employer closes the job and chooses that answer, and it does not depend on anybody's score.
How a score is arrived at
The overall score starts from how well you can do the job, on a scale of 0 to 100. That is the evidence in your CV and answers against the requirements in the advert.
A covering letter counts only if the employer asks for one. Two more parts can then move the score up or down.
| Part | Default weight | What it measures |
|---|---|---|
| Motivation | 25% | What the letter says about wanting this role |
| Specificity | 15% | Whether the letter is about this employer or generic |
A part that scores 50 moves nothing. With the default weights the sum is the skill score, plus 25% of how far motivation is from 50, plus 15% of how far specificity is from 50. Without a letter, the overall score is the skill score.
An employer can change these weights. The parts of each score are stored with the application and the weights in force are stored with the employer's account, so a score can be explained later.
Requirements come from the employer's own advert text. We do not add criteria.
What it is told to ignore
We do not ask about age, date of birth, sex, gender, ethnicity, nationality beyond a right-to-work answer, religion, disability, health, sexual orientation, marital or family status, and we do not ask for a photograph. They are not scoring criteria.
A CV may contain them because the candidate included them. The model reads the whole CV, so it sees that text, and it is instructed to ignore anything that is not in the advert, including name, age, gender, nationality, photographs and employment gaps. That is an instruction to a model and not a guarantee, which is one more reason a person decides.
Where it is wrong
Honesty here is worth more than reassurance.
- It reads what is written. A real skill left off a CV cannot be scored. The questions exist to reduce this, not to eliminate it.
- It cannot read every document. A scanned CV or an unusual layout may be read poorly. When reading fails, the application is stored anyway and flagged for a human rather than scored badly and buried.
- Language. Adverts and CVs in a language the model handles less well may be scored less reliably.
- A score is not a person. Two points apart is not a meaningful difference, and an employer should read them as level.
Your rights about this specifically
You can ask why you were ranked where you were, ask for a person to look again, put your side and object.
Ask the employer, or privacy@staffholm.com and we will pass it on.
Under the EU AI Act
Software used in recruitment and candidate selection is listed as high risk (Annex III). The duties for high-risk systems apply from 2 December 2027 (Regulation (EU) 2026/1744). They fall differently on the provider (Staffholm) and the deployer (the employer), and include risk management, data governance, logging, human oversight, transparency and accuracy.
Some of the design already points that way, namely advisory output, a human decision and stored reasons. The conformity assessment the Act will require is still ahead of us. The rules that apply today are the ban on inferring emotions at work, which this product does not do, and the duty to tell you that you are dealing with an AI system, which is what this notice and the line beside the upload are for.
Model
The reading and scoring uses Anthropic's Claude models through Anthropic's API. Anthropic's commercial terms do not allow it to train models on what we send. It deletes inputs and outputs within 30 days, and keeps them for longer only if they are flagged for breaking its usage rules (up to 2 years) or the law requires it.
Last updated: 1 November 2026