FAQ
Straight answers, first paragraph
No preamble before the point. If a question has an uncomfortable answer, the uncomfortable answer is the one here.
Ten questions
The ones that decide the purchase
Grouped roughly in the order they get asked.
No, and it is deliberately built to be visibly unlike it. There is no webcam, no room scan, no biometric identity check and no full-desktop recording. Capture is scoped to the assessment workspace tab and nothing else.
This is a design decision with evidence behind it: remote proctoring has drawn sustained litigation and campus-level bans, and the backlash attaches to the surveillance apparatus rather than to assessment itself. An instrument that looks like proctoring inherits its reputation whether or not it deserves it.
Honestly: less than we intend to. What is running today is the reviewer. A person reads the workspace timeline, the AI transcript and the deliverable together against the bank's answer key, and inconsistency between those streams is visible to them. An invite link is also bound to one candidate — the first session it opens is the one it keeps, so a forwarded link is refused rather than given an attempt of its own.
What is not running: rotation within a role, and a debrief bound to the candidate's own recorded moments. Both are the next integrity work. A bank now ships twelve authored cases and a role opens on one of them, so a leak costs that case rather than the bank — but everyone invited to that role meets it, and the debrief questions are the same for everyone who takes it. We would rather you knew that than discovered it. No biometrics, ever — that part is not a roadmap item, it is a refusal.
Because that is the thing being measured. Sandboxing our own model would make the assessment partly a test of how quickly a candidate adapts to an unfamiliar tool — a confound, not a signal.
Being precise about what exists today: the workspace has an AI panel with two modes. One is a provided assistant, which is built but not yet switched on in production. The other is bring-your-own, and it works by the candidate pasting each exchange into the transcript as it happens. So today the transcript is candidate-supplied text, not a proxied capture, and it should be read that way.
Proxying the candidate's own assistant so the transcript is first-party evidence is the version we are building toward. It is not what ships this week.
About 50 to 70 minutes all in: a 40 to 60 minute assignment hard-capped 15 to 20 minutes past whatever the case states and never above 75, plus a short written debrief. Async, on the candidate's own clock, pausable and resumable within a 48-hour window — and paused time does not count against either the timebox or the cap.
The length is evidence-bound rather than chosen. Take-home completion holds at roughly 55 to 75% for one-to-two-hour tasks and collapses past three hours; one well-documented case saw 15-hour take-homes produce around 20% outright non-completion and force a redesign. We target 70% completion and measure it by funnel stage.
A human writes it. There is no automated scorer in the product at all — every word of every finding is typed by a reviewer reading the session against the bank's answer key, and a release is refused until all six dimensions are written and the reviewer has confirmed the report says nothing the candidate should not read about themselves.
This is structural rather than promised: the database refuses every client write to a result and every read of one that is not released, and a single staff-gated code path is the only thing that can release it. Under the automated-decision rules now in force in several US jurisdictions, the obligations key off whether a tool substantially assists or replaces the decision — and today the tool does not score at all.
Because a composite is easy to skim and impossible to defend. It cannot tell you which judgment failed. It averages away the single finding you actually needed. And it is exactly the shape that attracts regulatory weight.
Six findings with evidence excerpts take longer to read and are far more useful: you can see the moment, disagree with it, and show it to the candidate.
No, and we think measuring that is a category error. A candidate who correctly judged that the model was the wrong instrument for a step and did it themselves has demonstrated judgment, and a metric that penalizes them is measuring compliance with a tool rather than competence at a job.
The public record is instructive here: at least one large organization publicly reversed an AI-usage mandate after the backlash. "Used AI a lot" is contested as a virtue. "Demanded a source for the rate and opened it" is not.
Typed annotation and a text debrief are first-class equivalents, built in the same sprint as the voice paths rather than added later. Speech recognition error rates for deaf and disordered speech have been measured around 78% against roughly 18% for typical speech; a product where voice is the only road to a score has built a discriminatory instrument regardless of intent.
There is also scale evidence that chat-based assessment completes better than video, so the accessible path is not a degraded product.
No ATS integration and no webhooks yet — those are roadmap, and we would rather list them as roadmap than as a paywall. Export does exist: every released report downloads as a readable file or as archival JSON, and an employer can take all of them at once.
The flow today is link-based and manual: you open a role, the dashboard mints an invite link, and you send it to the candidate yourself, or compose a letter the service delivers for you. Olive never writes to a candidate on its own initiative. Results appear in the dashboard. That is enough to run a real pilot without an integration project, which is the point, but it is not an integration.
Records are kept while a session is reviewed and for as long as your account holds them. Deletion is a request rather than a switch: it opens a 30-day window, can be withdrawn at any point inside it, and then removes the session, its events, the AI exchanges, the deliverable history, the capture, the report, and the reviewer's private notes with it.
The window is deliberate. A deletion that fired instantly could not be taken back, and a record you are obliged to keep is easier to notice before it is gone than after. Everything released can also be exported first — a readable file and an archival JSON carrying rubricVersion, scorerVersion and bankVersion, because a record you cannot reproduce is not a record.
Ask directly
Still something unanswered?
There is no salesperson below the Org tier, so you will get an answer rather than a discovery call.