Your interview completion rate only means something when you measure it one channel at a time and split it into two steps: how many invited candidates start, and how many starters finish. A blended number across phone screens, async video, AI voice, chat, and assessments hides where people leave, because each format asks for a different commitment. This article covers the formulas, directional benchmarks from independent research, and a rule for when a drop is worth investigating.
What the interview completion rate measures, and what belongs in the denominator
Interview completion rate = completed interviews / valid delivered invitations × 100
The word doing the work is valid. Validity gets decided once, at the moment the invitation is sent and delivered, and it never gets revisited. That rule is what keeps the metric honest, because the tempting move is always to go back and remove the people who disappointed you.
Exclude a record only if it was already invalid at or before send:
- Test records and internal QA runs.
- True duplicate records for the same person and the same requisition.
- Invitations that bounced or were never delivered.
- Candidates who had already withdrawn or been rejected before the invitation went out.
Everything that happens after a valid invitation lands stays in the denominator. A candidate who withdraws, accepts another offer, ignores you, or lets a working link expire received a fair opportunity, and a metric that quietly deletes them will improve every time your pipeline leaks. Classify those records instead of removing them:
| Outcome after a valid invite | In denominator | What it suggests |
|---|---|---|
| Completed | Yes | The screen worked |
| Started, abandoned | Yes | Something inside the interview lost them |
| Never started, no response | Yes | Invitation, timing, or interest problem |
| Withdrew or took another offer | Yes | Speed and competitiveness problem |
| Link expired unused | Yes | Window, urgency, or reminder problem |
Broken links and links that expired before the candidate had a fair window are process defects. Count them in a process-health view you actually read each week: name the cause, fix the cause. They stay in the completion denominator too, because a candidate you locked out is a candidate you lost.
You can report an adjusted diagnostic cut, for example the raw denominator minus post-invite withdrawals and confirmed technical failures, to isolate what the interview itself is doing. Two conditions apply. The raw operating rate stays the headline, and both numbers must be auditable, meaning anyone can pull the candidate list behind them and reconstruct the math. If the adjusted cut becomes the only figure anyone quotes, you have built a metric that goes up as candidate loss goes up.
Two mechanics matter. Count unique candidates, not invitation events, because reminders, resends, and reschedules generate several sends per person. Deduplicate by candidate ID rather than session, so someone who opens the link on a phone and finishes on a laptop that evening counts as one candidate, not one abandonment plus one completion.
Then pick a completion window and write it down. Count a completion if it lands before the stated deadline or before the requisition closes, whichever comes first. Review late completions, but keep them outside the metric so the number stays comparable week to week.
Split it into start rate and finish rate
Start rate = unique starts / valid delivered invitations × 100 Finish rate = unique completions / unique starts × 100
From those, two loss rates that are not interchangeable:
Within-interview abandonment rate = 100 - finish rate Overall non-completion rate = 100 - interview completion rate
Within-interview abandonment counts only people who got inside and quit, so it reads on the interview: length, question quality, device path. Overall non-completion counts everyone who received a valid invitation and never finished, including people who never opened the door. Overall non-completion is equal to or larger than within-interview abandonment. It is larger whenever any valid invitee never starts, and equal only if every invitee starts. Reporting it as drop-off makes a working interview look broken when the failure sits upstream.
Live phone and video appointments do not fit this shape. There is no partial attendance, so use a show rate:
Show rate = attended appointments / scheduled appointments × 100
Track reschedules as their own line. A candidate who moves a slot once and then shows up is not a no-show, but three reschedules across a pipeline is a scheduling problem worth naming. A no-show often points to a calendar or commitment failure and an abandonment often points to the interview itself, though neither one identifies the cause on its own. If you are comparing formats, the tradeoffs in phone screen versus video interview affect which number to watch.
A worked example
One requisition, one channel:
- 200 invitations sent
- 180 valid and delivered after the send-time check
- 126 unique starts
- 108 unique completions
Completion rate: 108 / 180 = 60%. Start rate: 126 / 180 = 70%. Finish rate: 108 / 126 = 85.7%. Within-interview abandonment: 14.3%. Overall non-completion: 40%.
The 60% headline looks mediocre. The split says something more specific: nearly everyone who starts gets to the end, so the interview is probably not the bottleneck. Thirty percent never opened the door, which points at the invitation, the sender, the stated time cost, or the delay between application and invite. Rewriting interview questions here would burn a sprint on the part that already works.
Benchmarks by screening channel
The independent Candidate Voice Report surveyed 2,587 recent U.S. and U.K. applicants who met an AI touchpoint during hiring, plus a 614-person comparison group with no AI contact. Its modality breakdown is the closest thing to a channel-by-channel baseline that is not vendor marketing.
| Channel | Did not start | Started, not completed | Completed |
|---|---|---|---|
| AI voice screen | 8% | 9% | 83% |
| Async video interview | 16% | 16% | 68% |
| Skills assessment | 14% | 22% | 64% |
| Chat AI screen | 12% | 28% | 60% |
Read this as direction, not as a target you owe your VP. Three limits. The data is self-reported by candidates rather than pulled from employer systems. The report found differences between vendors and between deployments inside a single modality were larger than the differences between modality averages, so a badly built voice screen will lose to a well built assessment. And the sample spans roles, employers, and seniority levels that probably do not match your mix.
The report also found that disclosing AI use and offering a human alternative improved candidate experience ratings. That is a completion lever and a fairness lever at once.
Live phone screens need a show rate, not a completion benchmark
That channel research does not provide a comparable live phone-screen benchmark, and the research behind this article did not turn up a sound apples-to-apples figure either. Pasting an async number onto a live screen would not work anyway. An async interview stays open for days and abandonment is silent. A live phone screen is a fixed appointment where the alternative is a missed call. The commitment points differ, the failure modes differ, and the denominators are not comparable.
Use your own rolling baseline. Four weeks of scheduled-to-attended show rate, sliced by role family and source, is a more honest comparison than any external figure. Keep that baseline alongside your other recruitment funnel metrics so a movement in show rate is visible next to what is happening upstream.
Diagnose the drop before you change anything
Match the symptom to a likely cause, then run one test.
| Funnel symptom | Likely cause | First test |
|---|---|---|
| Low start rate, healthy finish rate | Invitation is unclear, unbranded, or arrives late after application | Rewrite the invite with format, duration, and deadline in the first two lines |
| Healthy start rate, poor finish rate | Interview is too long, repetitive, or questions get harder without warning | Cut question count by a third; time the actual median session |
| Both weak | Role or offer mismatch, or a broken link and device path | Complete the interview yourself on an unlisted browser and phone before touching copy |
| Mobile completions far below desktop | Browser, microphone permission, or upload failure on mobile | Test on iOS Safari and Android Chrome, check where the session dies |
| One source underperforms | Different candidate intent from agency, job board, or referral traffic | Segment by source before drawing conclusions about the format |
| Completions spike in the final hours | Deadline is doing the persuading, not the invitation | Extend the window on one requisition and compare total completions, not timing |
Two segments deserve a standing look regardless of symptom: device and demographic subgroup. A gap that shows up for one group should not wait for the next quarterly review. Check whether the format creates an accessibility barrier, whether the accommodation path and human alternative are reachable from the invitation itself rather than buried in a careers page footer, and whether the gap holds once you control for role and source. Bring in legal or compliance review where the jurisdiction, role, or selection stakes call for it. Sometimes the cause is a broken device path and sometimes it is a real fairness problem, and the percentage alone will not tell you which.
A pass, clarify, stop rule
This is an operator's triage rule, not statistical doctrine. Tune the thresholds to your volume.
- Pass: within 5 percentage points of the four-week baseline for that channel and role family. Log it and move on.
- Clarify: moved 5 to 10 points, or fewer than 50 valid invitations sit behind it. Small samples swing wildly. Check the source mix and role mix first.
- Stop and inspect: fell more than 10 points on at least 50 valid invitations, or one device, source, or subgroup shows a material gap. Pull ten candidate records and look at where the sessions ended.
The sample-size guard is the part teams skip. A channel with 18 invitations produces a dramatic percentage every week, and reacting to it wastes more time than ignoring the metric.
Tests worth running
Change one variable at a time and give it a full requisition cycle. Stacked changes produce a number you cannot explain.
- State the format, honest duration, deadline, and whether AI is involved in the first two lines of the invitation. Disclosure costs nothing and correlates with better experience ratings in the survey above.
- Test a reminder schedule rather than assuming one. Run a single reminder partway through the window against a no-reminder control on comparable requisitions, and add a second only if the first moved start rate. Watch opt-outs and complaints alongside the rate.
- Match the promise to reality. If you promise 10 minutes and the median session runs 19, the drop-off is a broken promise, not attention span.
- Cut any question a resume already answers. SHRM reports that in the Monster Work Watch survey, 36% of candidates left a hiring process because they were asked to jump through hoops, and 47% cited poor communication when they withdrew.
- Run the whole thing on a phone you did not configure, on a network you do not control.
- Where your data shows scheduling is the friction, invitations opened but slots never booked, test a format that removes the calendar step on one requisition against a live-scheduling control. Kira-AI runs browser-based, audio-only AI voice interviews that candidates finish on their own time and turns them into structured summaries, with recruiters still deciding who advances.
- Shorten outcome communication on the previous requisition. Candidates who never heard back may be less willing to engage with your next invitation.
When a high completion rate is a bad sign
A rising rate can mean the screen got better. It can also mean you loosened the invitation criteria and now pay recruiters to review unqualified completions, or that an aggressive deadline pressured people into finishing something they resented, or that the interview got so shallow it collects no evidence worth reading. It can also mean someone started cleaning the denominator.
Pair completion with pass-through quality to the next stage, time-to-screen, candidate feedback scores, and slices by device, source, and subgroup. Judged alone it is a vanity number, which is why it belongs in a small set of hiring KPIs rather than on its own dashboard tile. The point is a better process for the people on the other end of the link, measured through candidate experience as much as through conversion.
Key Takeaways
- Measure interview completion rate per channel, not blended, because live screens, async video, AI voice, chat, and assessments ask for different levels of commitment.
- Freeze eligibility at send. Exclude only records already invalid then: test records, true duplicates, bounces, and candidates who had withdrawn or been rejected before the invite. Everything after a valid invite stays in the denominator and gets classified by outcome, including post-invite withdrawals, expired working links, and technical failures.
- Report the raw rate as the headline and keep both raw and adjusted numbers auditable. An adjusted diagnostic cut never replaces the raw figure, or the metric improves whenever the process loses candidates.
- Split the funnel into start rate and finish rate. Within-interview abandonment is 100 minus finish rate; overall non-completion is 100 minus completion rate, and it is equal to or larger than abandonment depending on whether any valid invitees never start.
- Treat the Candidate Voice Report modality figures as direction only, since vendor and deployment differences inside a modality were larger than the gaps between modalities. Live phone screens get a show rate against your own four-week baseline instead.
- Triage with a pass, clarify, stop rule, guard against samples under 50 valid invitations, and investigate device or subgroup gaps promptly, including accessibility and, where applicable, legal review.