What are recruitment metrics?

The change above the card compares the selected period with the one before it: pick 30 days and it is the 30 days immediately preceding them. Not the same period a year earlier. Seasonality in hiring is real, so the distinction matters; a number that compares summer with autumn carries the season inside it.

In two cases the direction is not read. If there were no records at all in the previous period, no percentage change is produced, because a ratio off zero is undefined. If the change is under 0.5 per cent in absolute terms, the direction counts as no change and the card stays neutral.

The number the card leads with is the absolute difference rather than the percentage, and the comparison baseline sits next to it: an up arrow, plus 48, previous period 4. The percentage appears only in the tooltip, when you hover the card.

The reasoning is simple. Going from 4 records in the previous period to 52 in this one is an increase of 1200 per cent, and that number says nothing; the sentence it went from 4 to 52 says something. Percentages off a small base are where recruitment reports go wrong most often.

The error card adds three separate events together: a failed CV profile, a failed AI analysis run and a failed email. The three merge into one number, because the question asked in the morning is not which layer broke, it is how many jobs were left half done today.

The error text you see in the panel is not the raw provider error, it is the matching entry picked from the catalogue. The raw text never leaves the server; that is a readability decision as much as a security one.

While no candidate is selected, the right pane of the application list opens a pool summary: Total Applications, AI Analysis (analysed over total), Average Score and Waiting. These four numbers belong to that pool, not to the dashboard, and the trend window is fixed at 30 days.

The waiting threshold is seven days and it is baked into the code: applications in Pending or Reviewing that are older than seven days are counted. It cannot be adjusted from the panel, and at most 12 people appear in the list. A pool scan reads at most 5000 records, and the breakdown by posting shows at most 8 postings.

Average Score can measure two different things, and which one it is stands on the label under the value. If the scanned pool holds cached position fit scores, you see Position fit avg.; if not, CV score avg. A single list always means one thing, the two score types never mix inside one average; but the averages of two screens cannot be compared without reading that label.

• The status filter in the pool does not touch the funnel. The funnel is deliberately computed without filters, so context is not lost while you look at a single status.
• A fit record with no score does not count as zero, it drops out of the average entirely. A missing score is never read anywhere as the worst candidate.

The score distribution report on the AI advisor screen gives four sets of numbers per posting: the band distribution (strong, review, weak, missing), the result of the required criterion gate (passed, closed, could not decide, not computed), the number of records flagged for human review and the number of overrides.

The default band thresholds are strong 55, review 35, weak 20 and they can be changed per posting. The job of the report is to show whether those thresholds are right for that posting: if everyone on one posting lands in the weak band, the problem is usually not the candidates, it is the threshold or the rubric.

The report returns numbers only. It carries no candidate name, no email address and no individual score, which is why it needs no separate masking step. A posting that has never been analysed is left out of the report on purpose; it says nothing about calibration and only pads the list.

• An override is two separate numbers: how many records are overridden right now and how many override events happened in total. Overriding the same record a second time raises only the event count.
• Rescoring after a criterion change is free. A candidate and position pair is charged once, so calibrating a threshold by trying it spends no extra credit.
• A recalculated score does not erase the old one. The previous score, the band, the gate result and the reason for the recalculation are written to a separate history row.

Your dashboard shows AI spend in credits: spent, remaining (the plan allowance plus add-on packs) and a breakdown by operation. The unit fits in one sentence: one AI evaluation is one credit. The same candidate and posting pair is charged once.

When the remaining allowance falls under 10 per cent of the monthly allowance, the panel warns you that your allowance is running low. Cost analysis in dollars and tokens is not on the company dashboard; that calculation lives on the platform administrator screen and is not a surface sold to customers.

The accuracy of the metrics is a property of the process, not of the dashboard. When the status of an application changes, the record update and the audit entry enter a single database transaction: the changed field, the old value, the new value and the user who made the change are kept together. The activity feed keeps who, what, which module and when; the dashboard shows the eight most recent of those entries.

One condition is left and it sits on the human side: when the status really changes, let it change in the panel too. If a candidate has been invited to an interview while the status stays Pending, no number is correct; the funnel, the average score and the waiting count all carry the same delay. AI never changes the status of an application on its own, it produces suggestions and the team makes the call.