How Evidence review works

What it is for

Evidence review supports a recruiter or hiring manager who reviews one candidate’s CV, and portfolio if there is one, against needs that the recruiter has confirmed for a role. For each need it shows quotes from the material as evidence, says whether evidence was found, partly found or not found, and suggests interview questions. A person reads the material, checks each result and makes every decision.

What it must not be used for

• Deciding on its own who goes on, who is rejected or who is hired.
• Ranking, scoring or comparing candidates, or sorting many CVs. The tool takes one CV at a time and gives no score.
• Needs that are not about the job, or that relate to age, gender, origin, religion, disability, health, family, sexual orientation, political views or union membership.
• Assessing people who have not been told that an AI-supported review is used.

How it works

1. You define the needs. The tool takes them from the job description you upload, and you check, edit and confirm them before any review.
2. Before the CV and portfolio are sent to the AI, the candidate’s name, contact details, ID numbers, links and personal-detail fields (such as date of birth, family, nationality and health) are removed. Portfolio links are read only if the pages are public.
3. The AI model (Claude, by Anthropic) looks for evidence for each need. Every quote is checked against the CV text, and a quote that is not in the CV is removed. If evidence was claimed but no quote could be verified, the result says so, and you check the CV yourself.
4. Texts that contain an overall verdict, a ranking or a reference to protected characteristics are removed.
5. You check each result against the CV, add your own notes, and can download a review record.

Limits and known risks

• The AI can miss evidence, especially when it is described in other words, in another language, or in a CV with an unusual layout. “No evidence found in the CV” only means that the CV text does not show it, not that the person lacks it.
• The removal of personal details is automatic and can miss something, or remove a line that mattered. Read the CV yourself.
• CVs differ in how much people write about themselves, which can favour some groups. Use the interview questions to give everyone the same chance to show what they can do.
• Do not let the result anchor your judgement. Read the CV first, then the result.

Human oversight

The person who uses the tool must have the competence to assess candidates for the role and the authority to disregard the tool’s result. Each result has a checkbox for your own check, and the review record shows what you confirmed. If a result looks wrong, report it from the result page: that is how the tool improves.

Your data

• The needs and the reduced CV text are sent to Anthropic’s API to produce the result. Nothing from the CV or the result is saved on mikaelafrisk.com.
• For traceability, each use is logged without personal data: time, tool and model version, the number of needs and results per status, and a one-way code of the input. The log is kept for 12 months.
• Pilot users get a personal key, which can be revoked at any time.
• You, as the employer or recruiter, are responsible for the processing of the candidate’s personal data. Before pilot access is given, we agree on terms, including how personal data is handled.

What you need to do as a user

• Tell candidates that an AI-supported review is used, and that a person decides. You can use the text below.
• Have a lawful basis under GDPR, and consider a data protection impact assessment for your recruitment process.
• Inform employee representatives where that applies.
• Keep your review records for at least six months, and be ready to explain to a candidate what the review showed and how you decided.
• Use the tool only as described here.

A text you can use to inform candidates

Edit it so that it matches your own process.

In our recruitment we may use an AI-supported tool (Evidence review by Mikaela Frisk) as decision support. It shows which parts of your CV relate to the needs of the role, and suggests interview questions. The tool gives no score and makes no decisions: a person reads your CV and decides. Contact details and personal details such as date of birth and family are removed before the review, and the tool does not save your CV. You can ask us to explain how your application was assessed, and you can ask for your application to be reviewed without the tool.

Copy the text

Status and the rules it is built for

Evidence review is a pilot. It is designed for the requirements on high-risk AI systems in the EU AI Act (Annex III, employment), which apply from 2 December 2027: a defined purpose, risk management, data minimisation, logging, instructions for use, human oversight and accuracy checks. Before the tool is offered outside the pilot, its technical documentation, conformity assessment and registration will be completed, and the terms reviewed by a lawyer. Questions: hello@mikaelafrisk.com.