AI matching — instructions for use

Version 1.5 of 25 October 2026

Status: These instructions cover the AI matching. The feature has been available since 26 July 2026 and requires an active job posting. Interview requests are either included in the job ad package or paid for individually with credits (a virtual balance).

For employers using the AI matching. Under Art. 26 of the AI Act you are a deployer of this system and carry obligations of your own. These instructions give you the information you need for that.

When this applies: The deployer obligations under Art. 26 (Chapter III, Section 3) only take effect for high-risk systems under Annex III on 2 December 2027 (Art. 113 as amended by Regulation (EU) 2026/1744). Until then these instructions are preparation: the right way to handle the system already today, legally binding from that date.

Provider: Actilize UG (haftungsbeschränkt), Am Isarkanal 24, 81379 Munich, Germany
System: AI-assisted matching of job ads and candidate profiles
Classification: High-risk AI system, Annex III No. 4(a)

1) What the system is intended for

The matching suggests candidate profiles for one of your job ads and sorts them by a score. It is a suggestion feature alongside your other channels, not a replacement for them: you can still search yourself, and candidates can apply to your ad directly and independently of it.

The same score on applications: For ads on a paid package or with the “Matching with interview requests” upgrade, the candidates page shows the same score with its breakdown on every application you receive once you have confirmed these instructions, and you can sort the cards by it. Sorting is an option, not the default; no application is hidden by it. While an evaluation is still pending, the card shows no value. A card reading “No match score” says nothing about suitability: either there was no published profile, or the person objected to the AI assessment — we do not show which.

What it is not intended for: It makes no hiring decision and replaces no interview. The order is a suggestion that supports your own review — not a verdict on anyone's suitability, and not intended to serve as the sole basis for a rejection.

2) How the score is produced

In two stages:

Pre-selection without AI. Only details from your ad are compared with details from the profile: language, weekly hours, type of employment, working arrangement, customer meetings on site and distance. Anyone who does not fit here does not appear — with no model involved.
AI assessment. For the remaining pairs, a language model reports findings: how similar the job titles are, which required skills are covered, how relevant the professional experience is. The points themselves are calculated by our own code using fixed weights, not by the model.

The strongest influence comes from the required skills, followed by the relevance of professional experience, the mandatory requirements and the similarity of the job title. Field of work, industry, distance, activity on the platform, an existing profile video and an existing membership have less influence. You can see the full breakdown for every suggestion in the score tooltip.

The model cannot remove anyone. Missing mandatory skills or missing years of experience cost points and lead to a lower rank, never to exclusion.

3) What you can influence

Distance: You choose in the ad which radius is taken into account. Candidates also state how far they are willing to commute; the stricter of the two applies.
Requirements: Which skills are mandatory and how many years you expect is set by you in the ad — this has a direct effect on the scores.
The more precise your ad, the better the result. Incomplete requirements lead to less precise suggestions.

4) Your obligations as a deployer (Art. 26)

Human oversight: Review every suggestion yourself before contacting anyone. If you switch on the auto mode for a job posting, the platform sends interview requests to profiles from a fixed minimum score without your individual review. Your oversight then consists of switching the mode on deliberately — which requires confirming these instructions once beforehand —, reading the posting's log and the weekly report, and switching it off as soon as something looks wrong. The auto mode also applies Applicant Boost suggestions to your posting's title, text and address (only those without open details); review these changes in the posting's log and adjust them if they do not suit you. The score is an indication, not a verdict. Assign this task to people who are professionally suited to it and have the necessary authority.
No decision on the basis of the score alone. Use it as one source of information among several.
Monitoring: If you notice anything unusual — systematically implausible assessments, for instance, or a pattern suggesting disadvantage — suspend use and inform us without delay (contact below). If you have reason to consider that use in accordance with these instructions presents a risk within the meaning of Art. 79(1) — that is, to health, safety or fundamental rights — Art. 26(5) additionally requires you to inform the competent market surveillance authority without delay. In Germany this is normally the Bundesnetzagentur, which under the German AI implementation act also serves as the central contact and complaints point; for particular sectors — supervised financial institutions, for instance — a sector authority may take its place.
Informing the people concerned: Where you use the system to make, or to help make, decisions about natural persons, you must inform those persons that a high-risk AI system is being used (Art. 26(11)).
Keeping logs: Automatically generated logs under your control must be kept for at least six months, unless other rules provide otherwise (Art. 26(6)). We keep the matching's own logs; logs arise on your side only insofar as you carry results over into your own systems.
Informing workers: If you use the system in an employment context, inform the affected workers' representatives and workers beforehand (Art. 26(7)).
Data protection: You are responsible for your own processing of the profile data shown. A data protection impact assessment may be required for your use; you may use the information in these instructions for it (Art. 26(9)).

5) Limits and known weaknesses

Not every profile holds experience details. Many candidates have not entered any description of their work under “My story” — the model then lacks that basis, and the assessment rests on job title, activities and skills. A low score can be due to this rather than to the person's suitability.
Language models err. A required skill may go unrecognised if it is named unusually in the profile. This is why a low rank excludes no one.
The order is not a ranking of qualification. It is a fit for this ad. The same person may rank at the very top elsewhere.
No statement about personality. The system assesses no character traits and makes no prediction about behaviour at work.

6) Which data is processed

Only the following is sent to the AI provider (OpenAI Ireland Ltd.): internal identifiers, profile type, job title, fields of work, desired job titles, activities, experience entries, education and training, skills with years, preferred industries. Not transmitted are name, address, photo, video, date of birth, contact details, salary expectations and hobbies. Before transmission, free-text fields are additionally checked for details that do not belong there — phone numbers or third-party names, for instance; hits are removed. Use for model training is contractually excluded.

Candidates can object to the matching at any time, through a switch in their account settings under privacy. The assessments and AI summaries we store are then deleted immediately, and the person no longer appears in any list of suggestions. What was transmitted to the AI provider beforehand is governed by the processing agreement with them and is not used for training; backups age out on their own cycle.

7) Oversight, shutdown, contact

We can switch the matching off platform-wide. If that happens, assessment, delivery and contacting stop at the same time — interview requests already sent are unaffected.

Anything unusual, questions and complaints: [email protected]. A report under Art. 26(5) additionally goes to the competent market surveillance authority, not to us instead (see section 4).