There’s a useful distinction at the heart of every AI feature in a recruiting tool, and almost no one is making it cleanly: is the AI assisting a human decision or making a decision and presenting it as assistance?
The difference matters. One is a productivity multiplier. The other is a category of risk you don’t want sitting between you and your next hire.
Here’s the framework we use internally at Join when evaluating any AI feature, ours or anyone else’s.
Three questions
- Could a human do this in 30 seconds with the same information? If yes, the AI is a productivity assist. Drafting the first version of a job ad from a few bullet points. Reformatting a candidate’s experience into a structured timeline. Summarizing a five-page CV into the parts a hiring manager actually needs. These are 30-second tasks for a focused human. The AI just removes the focus tax.
- Is the AI making a judgment that a human will later need to defend? If yes, you’ve moved into decision territory, and the AI shouldn’t be silently in charge. Ranking candidates by “fit”. Auto-rejecting based on a screening score. Predicting performance from a resume. These look like assistance, but if you ever have to explain to a candidate (or a regulator) why they were rejected, “the AI said so” is not an answer that holds.
- Is the AI legible? Could a candidate ask exactly what role it played? This is the EU AI Act question in plain language. Under Annex III, recruitment AI is classified high-risk, and a candidate has the right to know whether an automated system played a meaningful part in their rejection. If you can’t answer cleanly (not in your privacy policy, but in a sentence to the person), the system is too autonomous. Gartner’s 2025 survey found 87% of applicants want employers to be transparent about how AI is used in hiring; the Act now makes that a legal requirement, not just a candidate preference.
Where this lands in practice
Applying the three questions to the common AI features in a recruiting tool sorts them into four verdicts: drafting goes full speed, sorting and scoring are conditional, and deciding is off the table.
| Use case | Verdict | Condition |
|---|---|---|
| Drafting (job ads, screening questions, interview scorecards, rejection emails) | Full speed ahead | Human reviews before send. AI is a faster blank-page solver; the human still owns the final text. |
| Sorting | Yes, but transparent | Sort by recency, completeness, or a clearly-labeled “match score” the user can inspect and override. Never silently auto-rank, never auto-reject. |
| Scoring | Only when the scorer is explainable | Assessment-based scoring (the candidate took a test, here’s their result) is fine. The test is the substance. “Personality fit score” from a resume is not. |
| Deciding | Never | ”You should hire/reject this person” is not the AI’s call. That’s a category of liability we won’t absorb on your behalf. |
The actual product implication
This sounds idealistic. It’s not. It’s a product roadmap with consequences.
It means we’ve shipped fewer “wow” AI features than competitors who are willing to put a confidence score next to a candidate and call it done. It also means we sleep at night when regulators audit. The penalties under Article 99 of the Act reach up to €15 million or 3% of annual turnover for high-risk system failures; for an SMB hiring, that is not a fine, that is a company-ending event. And our customers can describe their hiring process in a sentence that doesn’t end with “…and then the AI does some stuff”.
For an SMB hiring its next account manager or junior engineer, that tradeoff is the right one. You’re not running a 10,000-applicant funnel where automated rejection is the only viable path. You’re running a 40-applicant funnel where the difference between hiring well and hiring badly is a human paying attention.
We’d rather build tools that help that human pay better attention, faster, than tools that try to replace the attention entirely.
That’s what “AI as assistant, not as decision” means in practice. It’s not a marketing slogan. It’s a feature filter. And it’s why our roadmap looks different from the next vendor’s.
