A recruiting agent can process a queue while the team is away. That does not tell you how much work it removed. Someone still prepared the job brief, checked the output, fixed missing details and handled the records the agent could not finish. A useful measurement includes that work. It also asks whether the result was good enough to use.
Start with one task, one definition of done and one comparison method. Do not begin with a companywide hours saved target. This guide sets out a measurement plan for sourcing, outreach, screening and scheduling without claiming a benchmark that belongs to someone else's team.
Define the task and its accepted output
Write down exactly what starts the task and what ends it. A sourcing task might begin with an approved job brief and end with a recruiter accepted candidate record. An outreach task might end with an approved draft, not a sent message or a reply. Those are different outputs and should not share one savings estimate.
Define the acceptance rule before you run the agent. For sourcing, require evidence for the agreed job criteria. For screening, require a clear summary with unknowns left as unknowns. For scheduling, require the right participants, timezone and agreed interview constraints. Keep the same rule for manual work and AI assisted work.
Measure active work and elapsed time separately
Active work is the time a person spends doing the task. Elapsed time runs from the start of the process to completion, including waits. Microsoft's agent metrics reference defines cycle time as elapsed process time and treats outcome, quality and productivity as separate metric groups. Use that separation in your recruiting report.
If a candidate confirms an interview sooner, that is a shorter scheduling cycle. Count recruiter hours saved only when the recruiter spends less active time arranging it. Agent runtime belongs in the technical report. Do not add unattended runtime to human savings.
| Measure | What to record | What it answers |
|---|---|---|
| Active human time | Preparation, execution, review and corrections | Did the task require less work? |
| Cycle time | Task start and accepted completion timestamps | Did the process finish sooner? |
| Accepted output | Results that passed the written acceptance rule | Was the completed work usable? |
| Exception work | Failed runs, manual completion and support effort | Where did the work move? |
Build a baseline from the team's own work
Observe the current process before changing it. Record active work by step rather than asking everyone for a single remembered estimate. Separate reading a brief, finding records, checking evidence, writing notes and updating systems. Pause the task timer when someone switches to unrelated work.
Keep enough context to compare similar tasks. Note the role family, task complexity, input completeness and recruiter experience. A repeat search for a familiar role is not a fair comparison with a new specialist brief. Keep difficult cases in the sample and label them instead of dropping them.
Use this baseline to choose the first workflow. The recruiting hours breakdown can help identify task boundaries. Measure the work your team actually does before deciding which part an agent should take over.
Include review, rework and failed runs
For each AI assisted task, record preparation, active operation, review, correction and exception handling. A draft that takes little time to generate but extensive checking has not removed all the writing effort. If the agent fails and a recruiter finishes manually, include both the recovery work and the manual completion.
Track initial setup separately from ongoing operation. Keep integration work, test preparation and training visible so a sponsor can judge the full investment. Record recurring maintenance, such as fixing a broken connection or updating a review rule, in the period when it happens. Do not silently move that work outside the report.
Compare matched work without double counting
Use comparable task groups and the same acceptance rules. Where practical, distribute similar incoming work between the manual and assisted process. Avoid comparing an unusually busy manual period with an unusually quiet pilot. Record changes to role mix, staffing or process rules that could explain a difference.
Aggregate total active minutes and accepted outputs for each group. Also inspect the typical task and the slow exceptions. An average alone can hide a workflow that works well for routine records but leaves a difficult review queue behind.
Do not give sourcing and screening full credit for removing the same review step. Assign every human activity to one task owner and one measurement bucket. Shared job brief preparation belongs in one place. If the team still performs it, it remains work even when several agents use the result.
Put output quality beside time saved
Choose task specific checks. For sourcing, review whether the record meets the written criteria and whether it duplicates an existing candidate. For outreach, check factual accuracy, role relevance and suppression rules. For screening, inspect the evidence and the recruiter correction. For scheduling, inspect conflicts and reschedules caused by errors.
Track the effort required to produce accepted work, not just generated drafts. Compare candidate complaints, recruiter overrides and corrections across the two processes. Decide in advance which failures stop the pilot. A faster workflow that creates unacceptable candidate handling should not pass because its time chart looks good.
Keep estimates separate from measured results
Microsoft's Copilot Studio savings calculator asks the agent owner to enter the estimated time another method would take. It supports savings estimates per run or per tool. That is useful for planning, but the entered assumption still needs local validation before it becomes a claim about your recruiters.
Label each figure as observed, estimated or projected. Project only over eligible work that actually occurs. Do not multiply a pilot result across every recruiter task. Report regained capacity separately from cash savings. Less active work does not by itself prove a budget reduction.
Measurement mistakes to avoid
- Counting every generated candidate as accepted output.
- Subtracting model runtime from a manual estimate and calling the difference human savings.
- Leaving manager review and failed run recovery outside the measurement.
- Counting the same shared task in sourcing, screening and outreach.
- Changing the acceptance standard after seeing the result.
- Presenting a vendor assumption as a measured recruiting benchmark.
Keep a short record of the baseline, comparison groups, acceptance rules and exclusions. When the workflow changes, note the version and measure again. That gives the team a reason to trust the result.
How Remqa helps measure recruiting AI work
Remqa builds agents for sourcing, outreach, screening and scheduling. We can help define task boundaries, add work and exception tracking, and keep human review visible in the workflow. Explore AI recruiting agents or start with a discussion of one queue, its acceptance rule and the work your team wants to remove.