AI for HR and recruitment
Four hundred applications for one role. The first sixty are read carefully, the next hundred quickly, and the rest are skimmed by someone who has already decided. Everyone in HR knows this. Very few processes are designed around admitting it.
Consistency is the real product
The argument for automated screening is not speed. It is that the four hundredth application is assessed against the same criteria as the first, by the same reasoning, with the reasoning written down. That is both a better candidate experience and a far stronger position if a decision is ever challenged.
Which means the criteria have to be explicit before anything is built. Turning a vague sense of “a good fit” into a written, defensible rubric is half the engagement, and it improves the process whether or not you automate afterwards.
Where we deliberately hold back
No automated rejections, and no scoring people on protected characteristics. The agent ranks and explains; a person decides. In several jurisdictions this is the legal line, and where it is not yet, it is still the right one. Any vendor happy to fully automate rejection is a liability you are being asked to carry.
We also keep the model away from signals that proxy for protected characteristics — name, photograph, age, address, university where it functions as a class marker — unless the role genuinely requires it, and we show you what was excluded and why.
What the agent does
- Structured extraction. Every CV, in any format or language, reduced to the same set of fields so that like is compared with like.
- Rubric scoring with reasons. A score per criterion and a sentence for each, so a recruiter can disagree with something specific.
- Candidate communication. Acknowledgements, status updates and scheduling in the candidate’s own language — the part everyone intends to do and nobody has time for.
- Screening questions. The three role-specific questions asked consistently up front, with answers attached to the file.
- An audit trail. Who saw what, when, what the system suggested and what the human decided.
Multilingual hiring
If you recruit across borders, a substantial share of good candidates write imperfect English and are filtered out for reasons that have nothing to do with the job. Reading and replying in the candidate’s language widens the pool measurably, and it costs nothing extra once the pipeline exists.
A reasonable first project
One high-volume role. Run the agent alongside your existing process for a full cycle without letting it touch a decision, then compare: where did it agree with your recruiters, where did it disagree, and who was right. That comparison is worth having even if you stop there.
Start with one high-volume role.
Run it in parallel with your current process for a cycle. You keep the rubric and the comparison regardless of what you decide afterwards.