What to check before choosing an agentic AI HR tool

AI agents are arriving in recruitment teams: automatic job ad distribution, application qualification, follow-ups, reporting. But not all tools that call themselves "agentic" are equal, and some choices commit the company on sensitive issues: candidates' personal data, automated decisions, regulatory compliance. Here is the complete checklist to run before signing.

What is an agentic AI tool in HR?

An agentic AI tool carries out, end to end, a task assigned by the recruiter, instead of limiting itself to suggestions. It chains the steps itself: setting up a distribution campaign, enriching an application, scoring it, triggering the next step of the process. The recruiter supervises, the agent operates.

The difference with "assistant" AI is fundamental for the evaluation:

• Assistant AI: generates a text, suggests a score, proposes an action. The human executes.
• Agentic AI: receives an objective ("distribute this job offer to the right profiles", "qualify these applications") and carries out the chain of actions. The human sets the framework and checks the results.

Concrete examples: an attraction agent that turns a job offer into a multichannel advertising campaign and runs it, or a qualification agent that enriches each application and produces a reasoned score. This is the model of tools like Seeqle, which combines a distribution agent and a qualification agent. The more autonomous the tool, the more critical the checkpoints below become.

Which questions should you ask about data?

Ask which data feeds the agent, where it comes from, where it is hosted, how long it is kept and who can access it. A serious vendor answers these five questions precisely; vague answers are a warning sign.

The data checklist:

• Origin: is the targeting or enrichment data collected lawfully? Can the vendor document it?
• Hosting: in which country is candidate data stored? Hosting in the European Union simplifies compliance.
• Retention: how long are applications and profiles kept? Can it be configured according to your internal policy?
• Separation: is your candidate data used to train models used by other clients?
• Sub-processors: which sub-processors (cloud, language models, enrichment) are involved, and are they listed in the contract?

Require the vendor's record of processing activities and a data processing agreement compliant with Article 28 of the GDPR before any proof of concept involving real applications.

Why is explainability non-negotiable?

Because a score without a justification is unusable and indefensible. The recruiter must be able to tell a candidate, a manager or an auditor why an application was prioritised. Explainable scoring details the criteria, their weighting and the elements of the profile that carried weight.

What to check during the demo:

1. Does each score come with a readable justification (criteria met, criteria missing) rather than a raw number?
2. Can your teams configure the criteria, or are they fixed in an opaque model?
3. Can you find out afterwards why a candidate was ranked at a given level three months ago?
4. Does the human stay in control: does the agent recommend and keep a trace, or does it reject applications on its own?

This last point is also a legal requirement: Article 22 of the GDPR strictly governs fully automated decisions producing significant effects on individuals, which includes rejecting an application. A good qualification agent ranks and explains; the decision remains human.

Sources: Regulation (EU) 2016/679 (GDPR), Article 22.

What should you check on the GDPR and the AI Act?

Check that the tool lets you inform candidates and answer their rights requests, and that the vendor takes a clear position on the AI Act: AI systems used for recruitment are classified as high risk by the European regulation, with stronger obligations to come.

The timetable to know, as amended in 2026:

• Since February 2025: ban on prohibited AI practices and AI literacy obligation (training the teams that use these systems).
• 2 August 2026: transparency obligations apply, including informing people who interact with an AI system.
• 2 December 2027: obligations for standalone high-risk systems apply (including recruitment and application screening tools), a date postponed by the AI Omnibus (Regulation (EU) 2026/1744), in force since 27 July 2026, instead of 2 August 2026 as originally planned.

Questions to ask the vendor:

• Is your qualification system classified as high risk within the meaning of Annex III of the AI Act, and what is your compliance plan to December 2027?
• Do you provide the technical documentation and the information notices to include in our candidate journeys?
• How do you handle candidates' requests for access, rectification and erasure?
• Is a template data protection impact assessment (DPIA) available to speed up validation by our DPO?

Sources: Regulation (EU) 2024/1689 (AI Act); Regulation (EU) 2026/1744 (AI Omnibus), in force since 27 July 2026; Quantic Avocats, 2026.

How do you assess the integration with your ATS?

Test the real flow in both directions: job offers must go from the ATS to the agent without re-keying, and enriched and scored applications must come back into the ATS with their justification. An "integration" limited to a manual CSV export is not one.

Checkpoints:

Ask for a demo on your own ATS instance, not on the vendor's demo environment.

What is reversibility and why should you require it?

Reversibility is the ability to leave the tool without losing your data or your history: full export of applications, scores and campaign statistics, in a usable format. Without a reversibility clause, you build your talent pool at a third party that can hold on to it.

To write into the contract:

• Full export on request: candidates, scores, campaign histories, in a standard format (CSV, JSON).
• Export lead time and cost: free or flat-rate, within 30 days at most.
• Certified deletion: written commitment to destroy the data after return.
• Service continuity: sufficient notice if a feature or the service is discontinued.
• No hidden dependency: audiences, targeting models or settings built with your data must be documented, so they can be rebuilt elsewhere.

Which hidden costs should you watch?

Beyond the subscription, watch setup fees, costs per additional user, how the media budget is billed, paid ATS connectors and volume overruns. Require an all-inclusive price simulated on your real annual volumes.

The cost items to have quoted in writing:

1. Setup and onboarding: initial configuration, team training.
2. Licences: per user, per site, per legal entity?
3. Media budget: in programmatic, is the media budget (distribution credit) billed separately, with or without a commission? What percentage?
4. Connectors: is the ATS integration included or billed in development days?
5. Volumes: what happens beyond the included volume of campaigns or applications?
6. Options that have become essential: advanced reporting, multi-brand, SSO, often reserved for higher tiers.

A useful benchmark: transparent platforms show a public entry price (at Seeqle, for example, a base plan from €500 per month, detailed on the pricing page) and a trial period. The complete absence of a public price list forces you to negotiate blind.

What level of support and guidance should you require?

An agentic tool automates execution, not strategy: require human support at launch (setting up the first campaigns, defining qualification criteria), responsive support in your teams' language, and regular performance reviews.

The support checklist:

• Structured onboarding: who sets up the first campaigns, how quickly, with which deliverables?
• Dedicated contact: an identified customer success manager, or a generic queue?
• Committed response times: contractual time to first response for blocking incidents.
• Business guidance: help defining audiences and criteria, sector-specific feedback.
• Progressive autonomy: documentation, help centre, training for new users included.

A last piece of advice: validate everything in real conditions. A two-week trial on a real open role, with your data and your ATS, tells you more than any demo. Run this checklist during the trial, not after signing: that is when you have negotiating power.