Overview

The AI Companion Dashboard is for TA teams and TA operations. It tracks how far the AI Companion has been taken up, what it does to interview feedback, and how the interviews it sat in read on quality and integrity. The Overview tab answers "is it being used, and is it saving the interviewer work": adoption of the Companion toggle, and what happens to the feedback afterwards. The Quality and Integrity tab answers "what did the interviews themselves look like": the readings the Companion generated after each interview, the sensitive topics it heard, and the same measures broken down by business unit and by interviewer.

How interviews are counted

An interview here is one interview slot inside a calendar event, not a calendar event and not a candidate. A slot is counted once however many interviewers are on it and however many feedback forms it produces.

Cancelled, deleted and declined slots, and slots with no video conference attached, are left out of every interview count on this tab: the Companion cannot join a slot that has no conference, so counting those would understate adoption.

Interviews run by an AI Interviewer are not part of this dashboard at all. They are dropped from every metric on both tabs, so nothing here mixes the Companion with the AI Interviewer.

Metrics

Total Interviews Scheduled

Definition: Total Interviews Scheduled measures all interviews scheduled, with and without the Companion, filtered by the interview's meeting date.

How it is calculated: Counts distinct interview slots.

Note:

• This is the denominator of Companion Adoption Rate, so the two always move together.

Companion Interviews Scheduled

Definition: Companion Interviews Scheduled measures the interviews scheduled with the Companion toggle turned on, filtered by the interview's meeting date.

How it is calculated: Counts distinct interview slots whose Companion toggle was on.

Note:

• The toggle is what is counted, not what the Companion managed to do. An interview counts here even if the bot was never admitted; Companion Bot Admissions is the metric for that.

Companion Adoption Rate %

Definition: Companion Adoption Rate measures the share of scheduled interviews that had the Companion enabled, filtered by the interview's meeting date.

How it is calculated: Divides Companion Interviews Scheduled by Total Interviews Scheduled.

Note:

• Both halves use the same exclusions, so the rate is unaffected by cancelled slots or by AI Interviewer interviews.

Weekly Trend in Companion Interviews Scheduled

Definition: Weekly Trend in Companion Interviews Scheduled measures Companion interviews per week, filtered by the interview's meeting date.

How it is calculated: Counts distinct interview slots with the Companion toggle on, grouped by week.

Note:

• This is a count, not the adoption rate: a flat line with falling hiring volume still means adoption is rising.

Feedbacks Generated by Companion

Definition: Feedbacks Generated by Companion measures the interview feedbacks the Companion drafted, filtered by the date the feedback was created.

How it is calculated: Counts distinct interview feedbacks that are linked back to a Companion session.

Note:

• Drafted, not submitted. A draft the interviewer never sent still counts here.
• Only interview feedback is counted; other feedback types are not in the dataset.

Feedback Generation Rate

Definition: Feedback Generation Rate measures, of the interviews the Companion recorded, the share it went on to draft feedback for, filtered by the date the Companion was due to join.

How it is calculated: Divides the sessions that reached feedback generated by the sessions the Companion recorded.

Note:

• The denominator is interviews the bot was admitted to and recorded, not interviews it was booked for. Interviews where the bot was never let in do not drag this number down; they show up in Companion Bot Admissions instead.

Companion Bot Admissions

Definition: Companion Bot Admissions measures how many times the bot was admitted to an interview and recorded it, filtered by the date the Companion was due to join.

How it is calculated: Counts distinct Companion sessions that recorded the interview.

Note:

• Read this against Companion Interviews Scheduled. A wide gap is an admission problem (the host never let the bot in), not an adoption problem.

Candidate Opt Outs

Definition: Candidate Opt Outs measures the candidates who opted out of the Companion during their interview, filtered by the date the Companion was due to join.

How it is calculated: Counts distinct Companion sessions whose last bot status is candidate opted out.

Note:

• This reads the session's latest status, so an interview is counted once however many times the bot tried to join.

How feedback cohorts work

Every feedback metric on this tab is split into two cohorts: Companion, meaning the feedback carries a link back to a Companion session, and No Companion, meaning it does not. The cohorts are the comparison the tab exists to make, so a number on its own is less useful than the gap between the two.

Only interview feedback is in scope. One interview produces one feedback per interviewer, so a panel interview contributes several rows.

Feedback Outcomes by Companion Usage

Definition: Feedback Outcomes by Companion Usage measures feedback volume, submission rate and time to submit for each cohort, filtered by the date the feedback was created.

How it is calculated: Counts distinct feedbacks and submitted feedbacks per cohort, and takes the median time between a feedback being requested and submitted.

Note:

• The median ignores feedback that was never submitted, so a cohort that submits rarely but quickly can show a better time than one that always submits.
• Time to submit is measured from the request, not from the interview.

Weekly Trend in Feedback Submission %

Definition: Weekly Trend in Feedback Submission % measures the share of feedbacks submitted per week in each cohort, filtered by the date the feedback was created.

How it is calculated: Divides submitted feedbacks by all feedbacks, grouped by week and cohort.

Note:

• The most recent week is always understated: feedback raised in it has had less time to come back. Read the trend, not the last point.

Weekly Trend in Time to Submit Feedback

Definition: Weekly Trend in Time to Submit Feedback measures the median hours between a feedback being requested and submitted, per week and cohort, filtered by the date the feedback was created.

How it is calculated: Takes the median time between request and submission, grouped by week and cohort.

Note:

• Only submitted feedback is measured, so slow weeks can look fast until the stragglers land.

Quality and Integrity

How interview readings are counted

Every measure on this tab comes from a reading the Companion generated after the interview, and only interviews it recorded have one. Interviews the Companion never joined are not measured and are not in the denominator.

Each reading carries its own state: generated, skipped or failed. Only generated readings are averaged or counted. A skipped reading is never treated as a zero, which is why two cards on the same page can be measured over different numbers of interviews.

A panel interview has one set of readings shared by everyone on the panel. Every card, trend and the business unit table count that interview once; only the interviewer table repeats it per interviewer.

Interviewer Talk-to-Listen Ratio

Definition: Interviewer Talk-to-Listen Ratio measures the interviewer side's share of the speaking time, filtered by the date the Companion was due to join.

How it is calculated: Averages the interviewer share of interviewer plus candidate speech across measured interviews.

Note:

• Either extreme reads badly: an interviewer who dominates, and one who barely speaks. All interviewers on a panel count as one side, and speech that could not be attributed to either side is left out.

Probing Depth

Definition: Probing Depth measures the substantive follow-up questions asked per topic opened, filtered by the date the Companion was due to join.

How it is calculated: Averages follow-ups per topic opened across measured interviews.

Note:

• Shown as a number, not a percentage: 2.0 means two follow-up questions for every topic opened.
• A follow-up has to build on the answer just given; filler and repeats do not count. Interviews with fewer than three interviewer questions or fewer than two topics are not measured at all, so very short screens drop out.

Actual vs Scheduled Duration

Definition: Actual vs Scheduled Duration measures the minutes an interview ran over or under its booked slot, filtered by the date the Companion was due to join.

How it is calculated: Averages the signed minutes between the interview's spoken span and the slot it was booked into.

Note:

• The number is signed, so interviews that run long and interviews that finish early cancel each other out. A cohort split between +10 and -10 reads as zero, exactly like a cohort that always ran to time. Use the breakdown tables to see which it is.

Interview Start Delay

Definition: Interview Start Delay measures how long after the scheduled start the interviewer began speaking, filtered by the date the Companion was due to join.

How it is calculated: Averages the seconds from the scheduled start to the interviewer's first words, and shows them in minutes.

Note:

• Not measured when the Companion itself joined more than a minute late, so the metric never reports the bot's own lateness as the interviewer's.

Feedback Form Coverage

Definition: Feedback Form Coverage measures the share of the feedback form's fields the interview produced usable evidence for, filtered by the date the Companion was due to join.

How it is calculated: Averages the per-interview field score, where full evidence counts 1, thin evidence 0.5 and none 0.

Note:

• Measured only for interviews that have a feedback form, so interviews on the default form with no fields do not appear.
• Low coverage is a signal about the interview, not about the form: it means the conversation did not reach what the form asks for.

Weekly Trend in Feedback Form Coverage %

Definition: Weekly Trend in Feedback Form Coverage % measures average coverage per week, filtered by the date the Companion was due to join.

How it is calculated: Averages the per-interview coverage score, grouped by week.

Note:

• Weeks with very few measured interviews swing hard; read them against the interview counts in the breakdown table.

How sensitive topics work

Every sensitive topic is one the organization configured, for the interviewer side, the candidate side, or both. Nothing ships by default, so an empty chart can mean no list was configured rather than no risk.

The two cards count interviews: the share of measured interviews where at least one topic came up. The two category charts count interviews per topic, so an interview that touched three topics appears in three bars and the bars do not add to the card.

Interviewer Sensitive Topics

Definition: Interviewer Sensitive Topics measures the share of measured interviews in which the interviewer raised a listed topic, filtered by the date the Companion was due to join.

How it is calculated: Divides the interviews with at least one interviewer finding by all measured interviews.

Note:

• This is the metric with legal and fairness exposure behind it; treat a rising number as a training signal, not a reporting artifact.

Interviewer Sensitive Topics by Category

Definition: Interviewer Sensitive Topics by Category measures how many interviews the interviewer raised each listed topic in, filtered by the date the Companion was due to join.

How it is calculated: Counts distinct interviews per topic, keeping interviewer findings only.

Note:

• Bars group on the topic's stable id, not its label, because an admin can reword a label and a label can arrive in the interview's language. The bar shows the most recent label for that topic.
• Interviews that raised no topic produce no bar, so this chart cannot be used to work out how many interviews were measured.

Candidate Sensitive Topics

Definition: Candidate Sensitive Topics measures the share of measured interviews in which the candidate raised, or was drawn into answering, a listed topic, filtered by the date the Companion was due to join.

How it is calculated: Divides the interviews with at least one candidate finding by all measured interviews.

Note:

• This is a risk signal for the interviewer, not a finding about candidate conduct: what matters is how the interviewer handled it.

Candidate Sensitive Topics by Category

Definition: Candidate Sensitive Topics by Category measures how many interviews the candidate raised each listed topic in, filtered by the date the Companion was due to join.

How it is calculated: Counts distinct interviews per topic, keeping candidate findings only.

Note:

• Same grouping rule as the interviewer chart: the stable topic id groups the bars, the label is display only.

Interview Quality and Integrity by Business Unit

Definition: Interview Quality and Integrity by Business Unit measures the interview count and the average of every quality and integrity measure for each business unit, filtered by the date the Companion was due to join.

How it is calculated: Counts distinct interviews per business unit, then averages each measure over the interviews where that measure was generated. Column by column:

• Interviews — distinct interviews in that business unit.
• Coverage % — average share of the feedback form's fields the interviews produced evidence for.
• Probing Depth — average follow-up questions per topic opened.
• Talk To Listen Ratio % — average share of the speaking time held by the interviewer side.
• Duration Drift (mins) — average signed minutes over or under the booked slot.
• Start Delay (mins) — average time from the scheduled start to the interviewer's first words.
• Interviewer Topics % — share of measured interviews where the interviewer raised a listed topic.
• Candidate Topics % — share of measured interviews where the candidate raised a listed topic.

Note:

• Each column is measured over its own set of interviews, so the counts behind two columns in one row can differ; the Interviews column is the widest of them, not the denominator of every column.
• Interviews with no position appear in a "(no position)" row, so the table adds up to the cards.
• Every row is shown however few interviews it has, so small business units are noisy.

Interview Quality and Integrity by Interviewer

Definition: Interview Quality and Integrity by Interviewer measures the interview count and the average of every quality and integrity measure for each interviewer, filtered by the date the Companion was due to join.

How it is calculated: Counts distinct interviews per interviewer, then averages each measure over the interviews where that measure was generated. Column by column:

• Interviews — distinct interviews that interviewer sat in.
• Coverage % — average share of the feedback form's fields the interviews produced evidence for.
• Probing Depth — average follow-up questions per topic opened.
• Duration Drift (mins) — average signed minutes over or under the booked slot.
• Start Delay (mins) — average time from the scheduled start to the interviewer's first words.
• Interviewer Topics % — share of measured interviews where the interviewer raised a listed topic.
• Candidate Topics % — share of measured interviews where the candidate raised a listed topic.
• Interviewer Talk To Listen % — average share of the speaking time held by that interviewer alone, measured against the candidate.

Note:

• A panel interview appears once per interviewer, so this table must not be used to average across interviews or to total the dashboard: the cards above are the interview-level truth.
• Interviewer Talk To Listen % is the only column about the person: it is that interviewer's own speaking time measured against the candidate's, so on a panel it tells two interviewers apart. Every other column is the interview's own value shared by everyone on the panel, so read those as the context an interviewer worked in rather than as that person's score.
• This column does not match the Interview Talk-to-Listen Ratio card at the top of the tab on panel interviews, and is not meant to: the card measures the whole interviewer side against the candidate, this column measures one person against the candidate. In an interview with a single interviewer the two agree.
• It is averaged only over the interviews where that interviewer's own speaking time was captured, which is fewer interviews than the columns beside it cover.
• An interviewer appears only once feedback has been raised against the interview for them, which is where their identity comes from.