What was due on 2 August, and what happened instead
The AI Act places software used for recruitment, selection, promotion, termination, task allocation and performance monitoring in Annex III: the high-risk tier. High-risk duties are the substantive ones, covering risk management, data governance, logging, human oversight, and documentation a deployer can be shown.
Those duties were due on 2 August 2026. Following an agreement reached in the spring and confirmed by the Council, standalone Annex III systems now have until 2 December 2027, with systems embedded in products given until 2 August 2028. Ogletree Deakins puts it plainly: the delay is not a repeal, and employment-related AI "remains squarely in the AI Act’s high-risk category".
One thing did not move. Gibson Dunn notes that 2 August 2026 remains an active compliance date for the Article 50 transparency duties, which are the ones about telling a person they are dealing with a machine. So Europe now has a rule that you must be told a machine is involved, and no enforceable rule, until December 2027, about whether anyone can explain what the machine did.
| Obligation | Was due | Now due |
|---|---|---|
| Annex III high-risk duties, standalone systems (recruitment, selection, promotion, termination, task allocation, monitoring) | 2 August 2026 | 2 December 2027 |
| High-risk duties for systems embedded in products | 2 August 2027 | 2 August 2028 |
| Article 50 transparency duties (telling people they are dealing with an AI) | 2 August 2026 | Unchanged |
| Classification of recruitment AI as high-risk | Annex III, high-risk | Still high-risk |
The pressure that did not get an extension
Robert Half surveyed more than 2,000 US hiring managers and published the result in March. Sixty-seven per cent of HR leaders said reviewing AI-generated applications had slowed their hiring, one in five by more than two weeks. Eighty-four per cent of HR teams reported feeling overworked as a result, and sixty-five per cent of hiring managers said the surge had made it harder to verify that a candidate’s skills were real.
Greenhouse, whose software sits between applicants and employers at scale, reports an average of 254 applicants per posting across 175,000 live jobs, and a 412% rise in applications per recruiter. Its chief executive Daniel Chait has a name for the result. "Everyone’s using their own AI to solve their own problem, but it’s making the whole system worse," he told Benzinga in July. "We’ve called that the AI doom loop."
That is the part worth holding onto. Candidates did not adopt AI because they wanted to; they adopted it because the volume on the other side made a handwritten application a rounding error. Employers did not adopt screening AI out of enthusiasm either. Each side is responding rationally, and the sum is a market that Chait describes as working for neither.
Sixteen months in which explaining yourself is a choice
Put the two together and the shape is clear. Volume makes automated screening non-optional. The rule that would have obliged those screens to be documented and overseen has moved sixteen months down the road. In between sits a period in which transparency in hiring software is voluntary.
Voluntary is not nothing. It is the opposite of nothing: it is testable. A duty that arrives in December 2027 tells you what a vendor will eventually be forced to do. A choice made in August 2026 tells you what they are willing to do when no one is making them, which is considerably more informative.
And the disclosure that matters is cheap to give. Naming the fields a scoring model receives costs an afternoon. Publishing the weights costs a paragraph. Neither requires a compliance department, a certification, or a deadline. What they require is being willing to have the answer read back to you.
Four questions worth asking before December 2027
If you are a candidate, you can ask these of any employer whose process involves automated screening. If you are hiring, you can ask them of your vendor and get an answer in writing, which is worth having on file well before the obligation lands.
The fourth is the one that separates the serious from the polished. Every hiring product claims fairness. Very few will tell you which word they have decided never to use about themselves, and why.
| Ask | A real answer looks like |
|---|---|
| What does your scorer actually receive about me? | A list of fields, and a list of fields it is never passed. Not a policy statement. |
| What are the weights, and do they move? | Published numbers, plus the conditions under which they change. |
| Has anyone tested it for bias, and what failed? | A dated audit that names a result it did not like. |
| What will you not claim? | A refusal in writing. Everyone claims fairness; few will name the word they avoid. |
Where we stand, since we are asking
This desk belongs to a company that ships hiring software, so the questions above are ones we should answer rather than only pose. Our scoring model is never passed a candidate’s name, photo, age, gender, nationality, or the prestige of their school or past employers. It receives six dimensions and the weights are published: 30, 20, 15, 13, 12 and 10, for skills, experience, location, salary, job type and work style. They shift with intent, and we print that too.
The audit is public, dated, and names what it fails: on-site and hybrid roles still vary by around twenty points across cities, down from thirty-four before a fix, and the residual is on the page. The word we have decided never to use about ourselves is "unbiased", because nobody can check it.
We publish all of that because December 2027 will eventually make some of it compulsory for everyone, and the interesting question is what a company does in the sixteen months when it is not.
The audit this piece refers to, including the part it fails, is at babzituna.com/bias. It is dated, and it names the result rather than the intention.