Instructions for Use — HRBlade

Provider: AMISCON GLOBAL S.L., NIF B70862099, Valencia, Spain Last updated: 19 August 2026 Contact: info@hrblade.com

Prepared under Article 13 of Regulation (EU) 2024/1689. Article 13 obligations apply from 2 December 2027 for Annex III systems; this document is issued early so that deployers can prepare. Send to any customer on request, no NDA needed.

1. Identity and intended purpose

HRBLADE is a hiring platform whose AI functions assist a human hiring team in screening and evaluating candidates. It is intended to be used by an employer or a recruitment agency, by trained members of a hiring team, to produce scores, summaries and rankings that a person then acts on.

We classify the screening, ranking, matching and evaluation functions as high-risk under Annex III point 4 (employment, recruitment and selection). We do not rely on the Article 6(3) exemption, because it is unavailable to any system performing profiling of natural persons and these functions do profile.

2. What the system is not for

Using it for any of the following is outside the intended purpose and is not supported:

as the sole basis for a hiring decision, with no human assessment;
to infer emotions, personality or health from a candidate's appearance, voice or behaviour. The platform has no such capability and must not be represented as having one;
to screen on any protected characteristic, or on a proxy for one;
in a jurisdiction where the deployer has not satisfied its own obligations on informing candidates and worker representatives;
to assess people for a purpose unrelated to a specific role, such as general ranking of a workforce.

3. Functions, inputs and outputs

Function	Input	Output	Role in the decision
CV screening	CV text, structured profile fields, vacancy text	Score 0-100, sub-scores, matched and missing skills, summary	Advisory. Ranks and highlights.
Interview answer analysis	Answer transcript, the question asked	Score 0-100, sub-scores for relevance, depth, clarity, strengths, weaknesses	Advisory.
Application summary and competencies	All answers and their analyses	Written summary, score per competency	Advisory.
Speech to text	Recorded audio	Text transcript	Preparatory. No voice properties assessed.
Semantic search	Candidate profile text	Vector, used for retrieval	Surfaces candidates for a human.
AI voice interview	Live call audio	Transcript, structured assessment	Advisory. Agent discloses it is an AI.
AI-authorship detection	Written answer	Likelihood score, indicators	Signal only. Never triggers rejection.
Pipeline automation	Scores above	Stage change, or a rejection proposal	Acts, within the limits in section 5.

Cognitive assessments are scored by a deterministic psychometric engine and are not a language-model output.

4. Human oversight the deployer must provide

Article 14 requires the system to be designed for effective oversight, and Article 26 requires the deployer to actually exercise it. Concretely:

Assign named reviewers who have the authority to overturn any AI output. A reviewer who cannot change the outcome is not oversight.
Give them time. The AEPD's test for meaningful human intervention includes having sufficient time per decision; a reviewer processing a hundred long reports a day does not meet it. Instrument this if you can.
Train them on automation bias, the tendency to defer to a score because it is a number. The score is one input among several.
Review the evidence, not just the score. Every score links to the answers that produced it.
Keep automated rejection off unless you have assessed your legal basis under Article 22 GDPR and can offer candidates human review. It is off by default and the platform proposes rather than rejects.

5. Limitations and known risks

The models are probabilistic. The same answer can score slightly differently on repeat runs. Do not treat a two-point difference between candidates as meaningful.
They assess text, not people. A candidate who writes fluently may score above an equally capable candidate who does not. This matters most for non-native speakers, and is one of the dimensions in our bias testing.
Free-text input can carry information we do not ask for. If a CV states age, gender or marital status, that text reaches the model. Prompts instruct the model to disregard it, and testing checks that instruction, but the cleanest mitigation is on the deployer's side: do not require such fields.
AI-authorship detection is unreliable at the individual level. It is known to misfire on non-native speakers. Treat it as a prompt to look closer, never as evidence.
The system does not verify facts. It does not confirm that a claimed qualification is real.
Out-of-distribution roles. Accuracy is lowest for roles very unlike those the underlying models saw in training, for example highly specialised technical or regulated professions.

6. Accuracy and metrics

Article 15(3) requires the declared accuracy and its metrics to appear here. We report on two axes, both from the controlled-comparison audit described in the AI Transparency Statement:

Impact ratio by dimension, against the four-fifths rule from 29 CFR 1607.4(D). Our internal threshold for action is 0.8.
Consistency, the share of matched profile pairs receiving the same outcome.

Current figures are published in the audit summary, available on request. They are refreshed whenever the underlying model, a scoring prompt or a threshold changes.

We deliberately do not publish a single "accuracy" percentage against hiring outcomes. Nobody has reliable ground truth for who would have succeeded in a role, and a number presented without that caveat would be misleading.

7. Input data the deployer controls

Article 26(4) makes the deployer responsible for input data being relevant and sufficiently representative where they control it. In practice:

write vacancy requirements that describe the role, not a person;
do not paste protected characteristics into vacancy text or interview questions;
keep company values text focused on ways of working, not on background;
use the same question set for every candidate for a given role.

8. Logging

The system logs every AI-influenced decision with the feature, model, score, threshold, rule fired, outcome, and reviewer where a person acted. Logs are retained for 12 months, above the six-month minimum in Article 19, and are exportable per candidate and per vacancy. Deployers have their own six-month retention duty under Article 26(6); the export satisfies it.

9. Changes

Changes to the model backing a function and to scoring logic are recorded and dated. Material changes are announced to affected customers at least 30 days before they take effect, so a deployer can re-run its own validation. Every such change triggers the bias audit before release.

10. Human review and candidate rights

Candidates are informed that AI is used before an interview and when they apply. They can request human review and an explanation through a token-based page that does not require an account. Requests reach the deployer, who is the controller; we supply the underlying decision record so the deployer can answer under Article 86 of the AI Act and Article 15(1)(h) GDPR.

11. Incidents

Report any suspected serious incident, including any outcome that appears discriminatory, to info@hrblade.com without delay. Article 73 deadlines run from the moment either party becomes aware, so prompt notice from the deployer is what makes the provider's clock workable: 15 days generally, 10 for a death, 2 for a widespread infringement.