📘

Who should read this

This document is for every organisation that uses the Jaicob platform to recruit ("you", the deployer in AI Act terms). It is the "instructions for use" that Jaicob, as provider of a high-risk AI system, must supply under Article 13 of the EU AI Act. Share it with the people in your organisation who own recruitment, compliance, and data protection.

This document is informational and does not constitute legal advice. Consult your own counsel for your specific situation.

Document version: 3.0 (October 2026). Version 3.0 replaces version 2.0 in full. It adds the AI agents, which act on their own within the limits set out in Section 3.1, and corrects version 2.0 where it said that a person reviews all AI output before it affects a candidate. The version number identifies the revision; the platform changes continuously and this document is updated when a change matters for how you use it.

1. Provider identification

Jaicob B.V. is the provider, within the meaning of Article 3(3) of the EU AI Act, of the AI systems embedded in the Jaicob platform. Registration and conformity obligations that rest on the provider of a high-risk AI system are handled by Jaicob on the timeline the Regulation sets; obligations that rest on the deployer are yours and are described in Section 7 and in the companion document EU AI Act: Deployer Checklist.

2. Classification and intended purpose

2.1 Classification

The AI features of the Jaicob platform are classified as a high-risk AI system under Article 6(2) of the AI Act in conjunction with Annex III, point 4(a): AI systems intended to be used for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates.

Dates. Under Regulation (EU) 2026/1744, the Digital Omnibus on AI (https://eur-lex.europa.eu/eli/reg/2026/1744/oj), the obligations for the high-risk systems listed in Annex III, this one included, apply from 2 December 2027. The transparency duties of Article 50, among them telling people that they are dealing with an AI system, have applied since 2 August 2026.

2.2 Intended purpose

The platform's AI features are intended to be used by professional recruiters and hiring teams in the recruitment and selection of candidates:

• identifying potential candidates for a specific vacancy from public and licensed professional sources (AI Sourcing), including an AI pre-selection of which sourced profiles are worth a recruiter's attention (Section 3);
• comparing candidate profiles and applications against the requirements of a specific vacancy and producing advisory match scores with written reasons (AI Matching);
• conducting structured screening conversations with candidates by SMS, WhatsApp or phone (AI interviews);
• assisting recruiters with questions, drafts and tasks on their own data (Jaicob, the recruiter assistant), including scheduled tasks that run without a recruiter present;
• carrying out recruitment steps for your team through AI agents (under "Studio" in the product), if your company uses them: contacting candidates and replying to them, screening and qualifying them, moving them through your stages, booking intakes, answering client contacts and finding prospects, within the limits set out in Section 3.1; and
• extracting structured data from documents such as CVs (document understanding).

Workflow automations ("Automations" in the product) are also part of the platform. They send messages you have written as templates and move applications between the stages you configure. They contain no AI-generated text and make no AI decision of their own; they are described here because the AI pre-selection in Section 3 decides which sourced people they reach. AI agents are a different feature: they write their own messages and act on their own (Section 3.1).

Two kinds of AI use. Match scores, the recruiter assistant, interview summaries and document understanding are decision support: a recruiter decides what to do with them. Three features act without a person reviewing each step: the AI pre-selection and the score threshold decide which sourced people you see (Section 6.3); AI interviews hold the screening conversation with the candidate; and AI agents act on their own within the limits in Section 3.1. None of them rejects anyone: every rejection is made or approved by a person (Section 6.1). The intended operator is a trained recruiter who sets these features up, monitors them, and approves what the platform asks a person to approve.

2.3 Reasonably foreseeable misuse (do not do this)

• Treating approvals as a formality: approving an agent's rejection, disqualification or offer, or an assistant action, without reading it, for example by approving a whole list at once (Section 6). The platform keeps every rejection behind a person; approving without reading defeats that.
• Emotion inference: using the platform, or attempting to configure it, to infer candidates' emotions. Emotion inference in the workplace context is prohibited by Article 5(1)(f) of the AI Act. The platform contains no such capability and none may be recreated via custom instructions.
• Filtering on protected characteristics: instructing the assistant, the search features or an agent's must-haves and screening questions to select or exclude candidates by gender, nationality, ethnicity, age, religion, disability, or other protected characteristics. The platform withholds gender, nationality and date of birth from its AI features by design (Section 8); do not attempt to reintroduce them.
• Use for other HR decisions: the system is intended for recruitment and selection. Using its outputs for decisions about existing employees (promotion, termination, performance evaluation) is a different high-risk category (Annex III, point 4(b)) and is not the intended purpose.
• Non-recruitment evaluation of people: creditworthiness, insurance, tenant screening, or any other evaluation of natural persons.
• Undisclosed AI conversations: disabling or circumventing the AI disclosure the platform adds to screening conversations and to messages an agent writes on its own. It is added by the platform, not by prompt; you choose how an agent's line reads, not whether it appears.
• Use on minors: the platform is intended for the recruitment of adults.

3. Capabilities and features

3.1 AI agents: what happens on its own and what waits for you

This section applies if your company uses AI agents. You set up an agent on a canvas of cards in Studio; each card is a job, such as finding candidates, a LinkedIn conversation or booking intakes. Once you activate the agent, it decides when to act and which step to take, within what its cards allow, its working hours and its daily limit on first messages (25 by default for a new agent). No person starts or reads each step. Replies to client contacts, booking links and client-interview invitations go out at once, also outside working hours.

The second row holds in every region and whatever an agent's settings say; no setting moves a step out of it.

Making an agent ask you first. On the agent's canvas:

• Cards whose work reaches someone outside your agency have a switch with two options, Asks you and On its own: Find candidates, WhatsApp interview, Text interview, Book intake, Book client interview, Answer clients and Find prospects. On Asks you, the agent pauses and waits for your approval before it acts.
• Message cards (LinkedIn invite, LinkedIn conversation, LinkedIn InMail, Email conversation, WhatsApp conversation, Re-engage candidates) ask who writes the messages: I write everything (the agent sends your own first message as written and passes every reply to you), I start, the agent replies, or The agent writes everything. None of the three makes the agent ask you before each message it writes; choose I write everything to write every message yourself.
• Steps that only change records in your own workspace (recording screening answers, marking someone qualified, moving a candidate or an application to a later stage short of an offer, closing a quiet process) have no switch. They happen on their own, are recorded with the agent as the one who acted, and can be undone.

The region setting (Section 10) changes none of this.

4. Limitations of performance and known risks

• Probabilistic output. Outputs of the language models are probabilistic. Scores are estimates, not measurements. Never treat a match score as an objective fact about a person.
• Errors and fabrication. Language models can produce plausible but incorrect statements, including in scoring reasons, summaries and the messages an agent writes. Verify substantive claims against the source (the CV, the transcript, the application) before relying on them.
• Agents act before anyone reads. An agent's message, reply, screening record or stage move reaches the candidate, or changes the record, before a person sees it (Section 3.1). A sent message cannot be recalled; a stage move or a qualification can be undone.
• Document extraction errors. CV parsing can misread layout-heavy, scanned or unusual documents.
• Language coverage. Best in English, Dutch, German and French; quality can degrade elsewhere.
• Voice limits. Speech recognition varies with connection quality, noise and accents. Voice processing is transcription and conversation only; the platform creates no voice prints and analyses no biometric characteristics.
• Bias risk. Language models can reflect biases in their training data. The platform withholds protected attributes and the candidate's name from the evaluations described in Section 8, but no technical measure eliminates this risk. Testing found small score shifts on details that can stand in for a protected characteristic; they were corrected by no longer giving the AI those details, and a small difference remains for some school names. The platform's own measurement of differences between groups is off until you switch it on and covers few candidates (both in Section 8). Your own monitoring and fair-hiring practices are necessary, not optional.
• Data recency. Sourced profile data reflects the source at query time.
• Input quality dependence. Vague vacancy requirements produce vague evaluations.

Accuracy metrics. Every individual scoring decision is reconstructable: each score is stored with its six dimension scores, the written reason for each, the prompt version and the model that produced it, and every re-score is kept as a new version. For companies with AI agents, the platform has also computed aggregate measures every day since 12 September 2026, over the candidates an agent scored: how scores are spread, how well they line up with your recruiters' decisions, how they drift over time, and, if you switch it on, how outcomes differ between groups by sex and age band (Section 8). You find them under Settings > Security > Scoring audit. No target values are declared yet; Jaicob reviews the first baseline in February 2027 and updates this section then.

5. Transparency toward candidates

• AI screening conversations by SMS and WhatsApp open with a disclosure, inserted by the platform on the model's first message in the company language, that the candidate is talking to an AI assistant and can ask for a person at any time. This cannot be disabled.
• AI phone interviews open with a fixed spoken message, in the company language and naming your company, that the candidate is talking to an AI assistant. It plays before the AI joins the call. The AI is also instructed to introduce itself as an AI assistant.
• Outreach messages sent by workflow automations contain no AI-generated text, so no AI system is speaking to the candidate there. Messages an AI agent writes on its own carry a line, added by the platform, saying that AI helped write them: on the first message to a person on each channel (LinkedIn, email, WhatsApp), not repeated after that. Replies an agent writes to your client contacts carry it on the first AI reply in each conversation, and again in a new conversation. You can choose how the line reads, not switch it off. It is left out of a template you wrote and of a message a person approved, one at a time or with Approve all: the person who approves a message sends it as their own. Outreach to people you sourced carries the GDPR Article 14 notice described in Section 10 when you turn it on.
• Consent. No AI screening conversation starts, on any channel, unless the candidate has opted in through a platform flow that names the AI conversation and its channels; an agent cannot start one without that consent either. STOP ends an SMS or WhatsApp conversation.

What remains yours: informing candidates that AI is used in your recruitment process before it is used on them, in your job postings, application flows and privacy notices, including, if you use AI agents, that an AI assistant may contact them and move their application forward. Section 7 contains a sample notice text.

6. Human oversight measures (Article 14 AI Act)

6.1 Built into the platform

• No rejection without a person. Every rejection is made or approved by a person, in every region and whatever an agent's settings say. A recruiter can reject an application themselves. An agent can only propose a rejection or a disqualification; the proposal waits until a person approves it, and the platform records who did. When an agent proposes a rejection email, its text is prepared before the request reaches you, so you see, and can change, the exact email the candidate will receive; after approval the agent sends that text and nothing else. If no email could be prepared, the request carries none and no email is sent. Approve all never includes a rejection or a disqualification: each stays on the list until you open it and decide it on its own. No score, filter or assistant action rejects anyone.
• Approval on assistant actions. Assistant actions that message a candidate, send email, delete a record or commit to a multi-step plan are presented to you for approval, one at a time. Approval requests are cryptographically signed by the platform, so an approval cannot be fabricated by a modified client. A refusal carries a written reason that is returned to the model. Scheduled assistant tasks that run without you present park these actions for your approval, with one narrow exception: a single message to a single recipient that the task itself named.
• Approval on agent steps. What an agent may not do on its own (Section 3.1), and every step of a card you set to Asks you, waits under Waiting on you on the agent's page. You see the proposed step with its details, can change it, and approve or decline it. The platform records the decision, who made it and whether several requests were decided at once. Approve all approves every request on the list in one step, except rejections and disqualifications, which you open and decide one by one. The list opens the first request and shows the others as one line until you open them, so open and read each one first.
• Scores you can read. The match detail shows all six dimension scores with the model's written reason for each, and the history of re-scores. For a candidate an agent scored, the match explanation also shows the weights, the must-have checks and the decision trail.
• Interruptibility. Workflow automations can be paused or stopped. An agent can be paused at any time; a paused agent starts no new work. Candidates can end AI conversations at any time. A STOP reply or the removal link takes a person out of all outreach at once; when a person asks an agent in their own words not to be contacted, the agent asks you to confirm before adding them to your do-not-contact list.

6.2 What is not an approval step

Two parts of the platform act without a per-action approval.

Workflow automations act when they run. Once you configure and start one, its steps send your templated messages and change application stages without a per-action approval. Oversight here is at configuration time: review what an automation does before you start it. A stage change made by an automation also appears in your connected ATS.

AI agents take most steps on their own (Section 3.1). Oversight here is in how you set them up (which cards, which of them ask you, who writes the messages, working hours and limits) and in the safeguards that hold without a person: a person on every rejection, the AI disclosure, the check on every message an agent sends on its own (Section 11), your do-not-contact list and the daily limit on first messages. When an agent moves an application to a stage that needs action, a task is created for your team. A stage change an agent makes also appears in your connected ATS.

6.3 The visibility gate on sourced people

For people found by a workflow's AI Sourcing step, the platform applies two AI decisions before you see them: an AI pre-selection (accept or reject) and, for accepted profiles, a match score. A workflow-sourced person whose score is below 55 is not shown to you, and your automations do not send them invites or messages. Manually saved, imported and assistant-saved people are never gated. People an agent finds pass the same AI pre-selection, but the score threshold does not apply to them: for them the score only ranks. No human sees a gated person before the gate applies. Jaicob records the pre-selection's decisions with their reason for six months; to limit volume, at most 30 rejections are stored per search and filter pass, so not every rejection has a stored reason. Gate states are repaired automatically every hour so nobody stays hidden for a technical reason. Jaicob has registered the gate as a risk and has an open decision on making hidden people visible behind an explicit control; this document is updated when that changes.

6.4 What you must do

• Assign oversight to named people with the competence, training and authority to exercise it, including the authority to overrule or discard AI output and to pause an agent.
• Ensure reviewers understand Sections 3.1 and 4.
• Decide, per agent, which steps happen on its own and which ask you first (Section 3.1), and check this again whenever you change the agent.
• Guard against automation bias (Article 14(4)(b)): an approval is a real decision. Reviewers must be able to explain why they approved an action, not merely that they did. Read each rejection request, including its email, before approving it.
• Keep a human decision-maker responsible for every hire and reject decision, and never present those to candidates as AI decisions.

7. Your obligations as deployer

Sample candidate notice (adapt with your counsel; leave out the second paragraph if you do not use AI agents):

We use the Jaicob platform in our recruitment process. It uses artificial intelligence (AI) to compare profiles and applications with the requirements of a vacancy, to select which profiles we approach, and to hold screening conversations.

An AI assistant may contact you, answer your questions, ask you screening questions and move your application to the next step. It may also close your application if you stop responding.

Only a person can decide to reject your application or offer you a job. We tell you the first time AI writes to you or calls you. You can ask to deal with a person instead, and you can ask us how AI was used in your application. See our privacy notice for your data protection rights, and https://jaicob.ai/legal/ai for how Jaicob's AI works.

8. Data governance, fairness, and bias monitoring

• Protected attributes are withheld from AI features. Gender, nationality and date of birth are not passed to the match-scoring prompts, are not part of what the recruiter assistant sees or can filter on, and are not extracted from CVs. Demographic answers received from job boards and connected ATS systems are discarded on receipt. Only if you switch on the comparison between groups (below), and only to measure it, the platform works out sex and age band from the gender and date of birth already on a person's record; no AI model receives them.
• The AI pre-selection of sourced people does not see the candidate's name. It judges the role, skills, location and industry signals, keyed on an opaque identifier.
• No emotion inference exists anywhere in the platform (Article 5(1)(f)).
• Vacancy-specific evaluation. Candidates are assessed against the requirements of the specific vacancy, with a written reason per dimension.
• Measuring differences between groups, if you switch it on. For companies with AI agents, Settings > Security > Scoring audit has a switch to compare outcomes by sex and by age band (under 40, 40 and over). It is off until you switch it on. Switching it on is your instruction to Jaicob, as your processor, to work out each candidate's sex and age band from the gender and date of birth you already hold, only to measure bias: decide it, record it, and make sure your DPA or written instructions cover it (deployer checklist). While it is on, a daily job compares outcomes over the candidates an agent scored, flags a group whose rate is below four fifths of the highest group's, and leaves out groups of fewer than five; while it is off, the page says so and shows no comparison. The bias report, and an export of the scoring records without names (CSV or JSON), are on the same page. Coverage is low even when it is on: most sourced people have no gender or date of birth on record, so the figures rest on a small share of candidates. Candidates scored outside an agent are not included. Records made before the switch existed, when the comparison ran for every company with AI agents, keep the sex and age band they carry, but count only while the comparison is on.
• Testing on details that stand in for a protected characteristic. Before a change to the scoring instructions or model reaches you, and every night, an automated test scores sample CVs in pairs that differ in one detail: a graduation year, a school, an association, a native language the vacancy does not require, a parental leave, or a district. Taken together over ten repeats of the test on 1 October 2026, some of these details lowered scores slightly. The largest shift in the overall score was about 4 points out of 100, for a student association linked to a country of origin, on a CV scoring just under the threshold of 55 (Section 6.3). Since 2 October 2026 the scoring no longer receives these details: it sees the level and field of an education but not the school, no years, the city but not the district, only the languages the vacancy asks for, and no career breaks. Repeated tests show no measurable shift from the details tested, except about 1 point for some school names that cannot be recognised as a school. A test every night fails if that difference grows beyond 2 points; it catches a shift of 8 points almost every night and smaller shifts less often. Details in free text, such as a summary, are removed by recognising words, which misses some phrasings. Older scoring versions, which only Jaicob can set for a company, still receive these details.
• Your own review remains necessary. Neither measure replaces your own review of outcomes across candidate groups (Section 7, Art. 26(5)).

9. Logging and record keeping (Articles 12, 19, 26(6) AI Act)

• The platform records: every match score with its dimension scores, reasons, prompt version and model id, per version; AI pre-selection decisions, accepted and rejected, with their reason, prompt version and model (rejections up to the limit in Section 6.3); workflow execution and step events; outbound message delivery; for AI agents, the agent's work history (each time it worked, with what it read, decided and did), every step with what it sent or changed, every approval request with the decision and who made it, and per candidate and vacancy the score, the qualification and the final decision with who made each; and an append-only audit log of authentication, permission changes, consent, erasure and field-level changes to candidate data.
• Logs of high-risk functionality are retained for at least six months (183 days, enforced in code as a floor no configuration can lower); the audit log is kept for one year. An agent's work history, with its approvals and the steps recorded in it, is kept for one year and then deleted; for a company on US rules (Section 10) it is not deleted automatically. Message content in delivery logs is redacted at six months, keeping the delivery record itself, where that redaction is switched on; in production it is not switched on yet and runs daily as a count only until it is.
• The audit log is available to you in-product with export; match versions are visible on each candidate's match detail; an agent's activity and approvals are visible on its page; and the scoring records can be exported from the scoring audit page.
• Provide log extracts to competent authorities when lawfully requested; we will assist.

10. Data protection and data residency

• Roles. You are the data controller for candidate data, including the people you source and contact through the platform: you decide whether and how they are informed (GDPR Article 14). As the sender of messages to business prospects, the opt-out duty of Telecommunicatiewet article 11.7 is also yours. Jaicob is your processor under the DPA: it supplies the candidate and prospect notices as settings and, whatever those settings, adds the AI disclosure to messages an agent writes and acts on STOP replies to agents and on removal links. For client account data, Jaicob is the controller.
• Hosting. Platform data is hosted on AWS in the European Union (Frankfurt).
• AI model providers. The platform uses general-purpose models from third-party providers as sub-processors under data processing agreements with no-training commitments. OpenAI traffic is pinned to OpenAI's EU endpoint. Traffic to other model providers is not pinned to an EU region; international transfers are safeguarded as described in the Privacy Policy. The LLM telemetry provider is a sub-processor and is listed in the DPA's sub-processor register.
• Sourced candidates and GDPR Article 14. The candidate notice is a company setting you turn on and word yourself (Settings > Company > Privacy); it is off until you turn it on. The default wording, in your company language, names your company, says the profile was found through an external source and why you are making contact, and links your privacy information and a one-click removal page; your own text can use the same three placeholders. When the notice is on, the first message to a person you sourced carries it, whether a workflow, an agent or the recruiter assistant sends it; a LinkedIn connection invitation does not. A first message that cannot carry it is not sent: a text that uses the privacy link needs a privacy URL, and the settings will not let you turn the notice on, or remove the URL, without one. Whatever the setting, a removal link already sent keeps working, and a STOP reply to an agent puts that person on your do-not-contact list.
• Erasure. Deleting an applicant or lead removes the record, its match scores, interview transcripts, and stored CV and attachments, and writes an erasure entry to the audit log. Delivery logs and the AI pre-selection's records survive for the six-month floor: the profiles it rejected, and the profiles it accepted, each with the person's name, headline and profile link. Content in delivery logs is redacted where that redaction is switched on. A person who objects through the notice link is removed from your outreach immediately; a lead removed this way is deleted after thirty days where that deletion is switched on, and a person an agent contacted is kept on your do-not-contact list instead. In production neither the redaction nor the thirty-day deletion is switched on yet; each runs daily as a count only until it is.
• Region. A company with AI agents chooses EU or US rules under Settings > Company > Data retention; EU is the default. The choice changes how long data is kept and nothing else. Under EU rules, contacts expire after the periods you set (the settings page says whether expired contacts are then erased automatically or stay until you remove them). Under US rules nothing expires automatically, and an agent's work history is not deleted after a year. The region does not change what agents do on their own, what waits for a person, or the AI disclosure. Choosing US rules for people in the EU leaves you without automatic expiry; as controller, that choice is yours.

11. Accuracy, robustness, and cybersecurity (Article 15 AI Act)

• Outputs are constrained by structured prompting, curated data projections and output validation; candidate-facing messages are sanitised before sending.
• Text written outside your company (CVs, profiles, and messages from candidates and clients) is passed to the AI marked as untrusted data, with the instruction not to follow anything written in it, wherever the AI scores, filters or writes to a person. Four steps still read it unmarked: the check on automated LinkedIn messages, choosing the language of an AI text interview, reading a CV, and the interview summary. Jaicob plans to mark these too. Marking lowers the risk of manipulation; it does not remove it.
• Messages an agent sends on its own are checked before they go out: a message with a link, email address or phone number that is not yours, the recipient's or in the agent's settings, or that repeats the agent's instructions, is not sent. Replies to client contacts get the same check. So do messages you approve with Approve all, because approving a list is not reading each message. A message a person approved on its own is theirs and goes out as approved. A rejection email is not checked this way, and Approve all never sends one: you open each rejection and read its email before you approve it. The check looks at links, addresses and numbers, not at tone or facts.
• Replies in AI text interviews (SMS and WhatsApp) treat a message that claims to come from the system, the recruiter or the hiring team as the candidate's own words, and are checked before sending: a reply with a link, address or number that is neither yours nor the candidate's, or a reference code you never wrote, is replaced by a standard message.
• The platform runs on infrastructure certified to ISO 27001:2022, with encryption at rest and in transit, role-based access control and tenant isolation by company. See the Privacy Policy for the security description and breach procedure (notification within 72 hours).
• Assistant tool actions are permission-gated and, for outward actions, approval-gated with signed approvals. Model-supplied record identifiers are verified against your tenant before any write.

12. Maintenance, updates, and substantial modifications

• The platform is continuously maintained; changes are recorded in the product changelog.
• We may update or replace underlying AI models. When Jaicob changes a model or its instructions on a path that affects candidates, the change must pass the automated quality and safety tests before it is released; the tests are repeated every night, and for companies with AI agents the daily drift measure shows a change in scores (Section 4). The safeguards in this document (disclosure, a person on every rejection, approval signing, withheld attributes, log retention) are independent of the model in use.
• Substantial modifications are announced to clients in advance and this document is updated.
• There is no fixed end-of-life date for the system; if a feature is retired, clients are notified with a migration and data-export path.

13. Incidents

• Your duty: if you identify a serious incident, or suspect that use of the platform presents a risk to health, safety or fundamental rights (a suspected discriminatory pattern, an undisclosed AI conversation, an action taken without required approval, an agent acting outside its settings), suspend the affected use and inform us without undue delay at [email protected] with "AI incident" in the subject line (Article 26(5)).
• Our duty: Jaicob operates post-market monitoring and reports serious incidents to the competent market surveillance authority within the deadlines of Article 73. We inform affected clients and cooperate on corrective actions.

14. Other jurisdictions

The safeguards above apply in every region; the region setting changes only how long data is kept (Section 10). Depending on where you hire, local rules may add obligations for you as employer, such as New York City Local Law 144, the Colorado AI Act, Illinois rules on AI in hiring, and California bot-disclosure rules. The platform's AI disclosure is consistent with the California rule. For companies with AI agents, the scoring audit page exports the scoring records with sex and age band where known and where you switched the comparison on (Section 8); whether that is enough for an independent bias audit under Local Law 144 or a similar duty depends on your auditor and on coverage, which is low. Verify current obligations for your hiring locations with your counsel.

15. Support and contact

• Compliance and documentation questions: [email protected]
• Incident reports: [email protected] (subject "AI incident"), see Section 13
• Data protection (DPA, sub-processors, data subject requests): [email protected]
• Public transparency statement for candidates: https://jaicob.ai/legal/ai

Read next: EU AI Act: Deployer Checklist.