Three moments from a screening call. Auto-extracted, recruiter-approved.

A screening call is forty-five minutes. The hiring manager has two.

The recruiter's job, in that fifteen-second gap between sending a submission and the hiring manager opening it, is to find the three moments from the call that would make the HM say "yes, let's talk to them" — and put them at the top.

Most recruiters do this with a paragraph of written context. Some upload a clip or two if they remembered to record. The truly diligent ones edit a highlight reel by hand. That takes forty-five more minutes and rarely happens twice.

What shipped today

Every Submio submission now has a "Send recording bot" affordance. Paste the meeting URL from Zoom, Google Meet, Microsoft Teams, or Webex. The Submio bot joins on schedule, announces itself, captures the audio and video, and disappears when the call ends.

When the transcript is ready, Claude reads it against the role's scorecard — the same scorecard the recruiter sees in the submission editor — and proposes the three most relevant candidate moments. Each clip comes with a one-line rationale and a mapping back to the specific requirement it evidences.

The recruiter sees the clips on a dedicated review screen. Each one is playable inline. The recruiter can edit the label, approve, or reject — with an optional rejection reason that helps us improve extraction over time. Approved clips render on the public HM review page alongside the existing manual clip strip. Nothing auto-publishes.

The judgment stays with the recruiter. AI does the curation work; the recruiter does the editing work.

The design call that made it work

Most AI-recruiting features shipping in 2026 are fully automated. They source candidates, write outreach, score interviews, and route reactions — all without the recruiter in the loop. The pitch is "save the recruiter time." The implicit promise is "the recruiter is the bottleneck."

We do not believe that. The recruiter is the judgment. They know which candidate is right for which client because they have spent years calibrating that intuition through real placements. A model trained on aggregated transcripts does not have that signal. It has aggregate competence and zero context for any one deal.

So we drew a line. Claude does the work that scales — read the transcript, find the candidate moments, label them, map them to the rubric. The recruiter does the work that does not — decide whether each clip belongs in front of this specific hiring manager for this specific candidate.

Specific things we decided not to do

We did not build automatic clip publishing. Every clip needs an explicit recruiter approval before it attaches to the submission.

We did not build per-clip video extraction. The full call recording is stored once; each clip references a start and end timestamp in the same source file. The browser handles the windowing. This means clips are technically scrubbable — a determined hiring manager could play past the end of an approved clip — and we accept that for V1. The review page is per-token shareable, not a public broadcast. If we see real misuse, we will swap to per-clip mp4 extraction.

We did not gate AI clip extraction behind a paid tier. Every Submio plan — including Free — ships every feature. That is the locked product principle and we are not going to break it because a feature is expensive to run.

The privacy details

Recall.ai is our notetaker bot vendor. They join the call on the recruiter's behalf and produce the recording + transcript. The bot announces itself when it joins; meeting participants know they are being recorded. The candidate can refuse or ask the recruiter to disconnect the bot at any time.

Anthropic's Claude is the model that reads the transcript and proposes clips. Anthropic does not train on Submio data (this is contractual, not a courtesy).

Both vendors are now listed in our privacy policy at /privacy-policy under section 5 (Who we share data with). Worth a read.

What is next

In the weeks after launch we are watching three things: (1) what percent of extracted clips recruiters approve unchanged, (2) what percent get label-edited, (3) what percent get rejected with reason. Those three numbers tell us whether the extraction prompt is calibrated for the actual work or whether it is generating clips that look plausible but miss.

If the approval rate is low on a specific scorecard dimension, we tune the prompt for that pattern. If it is high across the board, we extend to four or five clips per call.

If you ship submissions and want to try the feature on your next screening call, sign up at gosubmio.com — Free tier supports one active role and includes every feature including this one.