From submitted to fully scored
Median across 63 applications. The promise had been 30 seconds.
Applications to interview invitations
Every one scored against the same ten weighted criteria.
To decide 52 candidates
Each with a personalised email previewed and sent.
The volume
More applicants than anyone expected to read
The role went up on LinkedIn and Indeed and drew more than 144 inquiries in three weeks. 63 of those became complete, structured applications through the form, with a parsed resume and written answers to eight questions. On the busiest single day, 26 arrived.
For a team that had never run a formal round, that volume is the problem rather than the win. Sixty-three applications, each with a resume and roughly 1,700 characters of written answers, is a full working day of reading before anyone can be ranked, and it arrives while the rest of the business still needs running.
Eleven candidates were invited to interview. Every one of the other fifty-one got a written answer rather than silence.
The evaluation
Seventeen seconds, ten criteria, every applicant
A submitted application has its resume parsed and is scored against ten weighted criteria, each returning a band and the evidence it was drawn from. Median time from submit to fully scored was 17.2 seconds; nine in ten finished inside 26. That produced 616 individual criterion scores across the round, every one traceable to a line in the application.
A second pass checks whether the written answers look machine-generated. It flagged 12 of the 63. Those are not auto-rejected, they are surfaced, which leaves the judgment where it belongs.
The scores were advice, not a verdict. The highest-scoring rejected candidate outranked several who were invited, because the operator overrode the ranking on judgment, which is exactly what the tool is supposed to allow.
How it got better
The operator reshaped it, and we shipped in days
Eight changes came out of real use, each live within three days. A single ranked list in place of a strong and weak split, because the operator kept approving candidates from the weak side. Bulk decisions. A focus mode that hides anyone already decided. Scheduling links in the invitation email. Rubric editing on a role that is already published.
Two things went wrong in front of real people, and both are worth naming. A candidate could not submit because an optional field was validating too strictly: reported at 18:28, acknowledged at 18:32, a fix live at 18:35, root cause found and corrected the next day. Separately, a candidate’s reply bounced because the sending subdomain had no mailbox; that was reported, fixed, and apology notices sent to all eleven invited candidates within twelve minutes.
Seventy-three emails went out across the round with no delivery failures, and 59 of the 63 applicants opted to stay in the pool for future roles.
Kynox delivered an AI solution that elevated our recruiting process. Their intelligent review system cut our time to hire dramatically.
How these numbers were counted
Application counts, scoring times, criterion scores and email outcomes are measured from the production database and the system’s own logs, read on 21 Oct 2026, for one role run end to end. No candidate is named. The reduction in time to hire is the client’s own description: the previous baseline was never recorded, so there is no percentage to quote and none is given here. The API key for the evaluation model belongs to the client, as does the database.