How a round runs, end to end
Nine steps, from a qualified shortlist to a recorded decision. Two of them are gates a human controls, and the last one is the only step that decides anything.
Step 01
HR uploads candidates
A spreadsheet of people already screened as qualified, or one candidate at a time.
The people you upload have already been screened as qualified — RoundWise does not source applicants or decide who is worth interviewing. A sheet is validated row by row before anything is created, and a problem is reported against its row number with the specific issue: this email is malformed, this role does not exist, this resume link does not resolve. You fix the sheet, rather than guess at what a validation error meant.
Step 02
Invitations go out
Only after a person confirms the send. Every invitation says the first round is run by AI.
Bulk sending is irreversible, so nothing leaves the system until a person explicitly confirms it — and re-uploading the same sheet sends exactly one invitation per candidate, not two. The first message a candidate receives states that the first round is conducted by AI and offers the human-alternative path. Nobody should discover that after clicking.
Step 03
Secure access
Each candidate gets a private link, good for one interview and a set window.
The link is scoped to one candidate and one interview, and it cannot be reached from the public site. If the window closes before they start, a recruiter can reissue it. An invitation that expired and one that never arrived are recorded as different things, because they say different things about whose fault it was — a hard bounce is a delivery failure, not a candidate who ignored you.
Step 04
Consent gate
The candidate reads what the AI does and agrees before anything starts, or asks for a human instead.
No session starts until consent is recorded, and that is enforced on the server rather than only in the screen. Being invited as part of a batch is not consent — the gate is identical for every candidate. The disclosure names the exact modality: voice and text, no camera, no facial or emotion analysis. A candidate who declines is routed to a human, and nothing about them is assessed.
Step 05
The interview
The AI asks the planned questions by text or by voice. There is no camera at any point.
Chat is the default mode; voice runs over the phone. There is no camera in either. Every candidate for a role is asked the same core questions — the only questions that differ are probes grounded in that candidate's own resume, the job description, or something they said earlier in the interview. A probe may never be drawn from a protected characteristic, even one the candidate wrote on their own CV. When an answer is thin the AI follows up once, then moves on. The round is time-boxed.
Step 06
Analysis
Answers are matched to the role's competencies, each tied to the transcript quote behind it.
Answers are mapped to the competencies defined for the role, not to a general impression. Every score has to carry the transcript line it rests on: a score submitted without evidence fails validation rather than being published. The model's raw output is never trusted as-is — it is parsed and checked before it becomes a report.
Step 07
Summary and scores
A structured report. Where a competency was never covered, it says so instead of guessing.
The report keeps the core questions and the candidate-specific probes separate, each probe tagged with what grounded it, so any difference between two candidates' interviews can be explained rather than assumed. A competency the conversation never reached is marked not assessed and flagged. An absent score is more useful than an invented one.
Step 08
A recruiter reviews
The summary, the evidence for every score, and the full transcript — plus the audio, if the interview was by voice.
You see the summary, the quoted evidence under each score, the full transcript, and the audio recording when the interview was conducted by voice. The AI's recommendation is advisory, is labelled as advisory, and is deliberately presented as less prominent than the evidence beneath it. It is an input to your judgement, not a verdict you are ratifying.
Step 09
A person decides
Advance, hold or do not advance — recorded against the recruiter who chose it.
The three options carry equal weight on screen, and none of them is tinted to agree with what the AI suggested. The choice is recorded against the identity of the recruiter who made it, with a timestamp. There is no code path that sends a hiring outcome to a candidate without a human. If any step of the pipeline fails, it pauses and notifies a person instead of guessing past the failure.
AI handles the interview. A person makes the hiring decision.
Every step above exists to make that sentence true in the code, not only on this page.