A Dropped Call Is Not a Weak Candidate

For part of 2026, HireQwik could not separate a candidate who gave zero-value answers to two questions and a candidate who was never asked them. Both showed as Incomplete. We found it during a July code review: the scoring code only counted a question as assessed if its score was above zero. So a screen the rubric had clearly failed looked exactly like a call that dropped in minute three.

Fixing that made me think harder about the opposite mistake, which is the one candidates actually feel. An incomplete AI interview must never be scored as a weak one. A dead phone battery is not a communication problem. This post explains how HireQwik decides a call is Incomplete, what HR sees, what to do with each kind of partial call, and the patterns that suggest the problem sits in your setup rather than with your candidates.

What an incomplete AI interview means in HireQwik

An incomplete AI interview is a screening call that stopped before enough of the screen was covered to judge the candidate. HireQwik labels it Incomplete instead of giving it one of the four verdicts, flags it for follow-up, and does not turn the missing answers into a low score.

It sits beside the four verdict tiers, not inside them. The tiers, and the order in which the engine picks one, are covered in how every verdict tier is decided. Incomplete is the engine saying: I do not have enough to place this person in any tier.

Why a partial call must never be scored as a low one

Picture a final-year student taking the screen from a hostel room at 9 PM. Four minutes in, the Wi-Fi drops. If the system scored what it had, two answered questions and three blanks, the average would land somewhere near No Go. She would sit in the reject pile next to people who answered all five questions badly, and no recruiter scrolling a 3,000-row drive would ever know the difference.

Calls break more often than vendors like to admit. A 2026 field experiment with AI voice interviews reported technical difficulties in 7% of AI-led interviews, and 5% of applicants ending the interview because they did not want to speak to an AI. Those are two very different reasons for a short call, and neither tells you anything about ability.

There is a measurement principle underneath this too. NIST’s AI Risk Management Framework lists “valid and reliable” first among the qualities of trustworthy AI. An answer that was never given is not a low reading. It is no reading at all, and recording it as a score makes the whole result invalid.

The rule HireQwik uses to mark a call Incomplete

The exact rule depends on how the job is scored:

On the points rubric, an Incomplete result also carries an abandonment flag. Everything else about the call, the transcript, the recording and any partial notes, is kept.

Two nearby labels are easy to confuse with it. Pending (the inbox says “Awaiting eval”) is a finished call whose scores have not arrived yet; check back shortly. A knockout reject is not incomplete either. A failed must-have question ends the call deliberately, often very early, and the row becomes a No Go that quotes the failed question and what was said in reply. Candidates often describe both as “the call ended early”, which is why the ended-early explainer exists for them.

The bug that hid real rejects as Incomplete

Back to that July finding. Every rubric in HireQwik defines what a zero means for each question, something like “cannot answer” or a knockout-level response. So zero is a real result. But the old code treated any score of zero as “not asked”. A candidate who scored zero on the first two core questions was counted as having answered only one, fell under the two-question rule, and showed as Incomplete with an abandonment flag.

That sounds kind to the candidate, and in one narrow sense it was. It was also wrong in a way that could cost everyone. A recruiter seeing Incomplete would reasonably assume a dropped call and might reset the booking for someone who had in fact finished the screen and failed it. That candidate would get a second attempt nobody else got, and the recruiter would spend time on a retry the evidence did not support.

The fix was one line of logic: any numeric score, zero included, counts as assessed. A clear fail is now a No Go, and Incomplete means what it says. I mention it because “flag it rather than score it” is only fair when the flag is accurate in both directions.

The call that drops after question two

Here is a limit of the rule that I want you to know about, because it is easy to miss. The points rubric needs two of its three core questions scored to produce a verdict. So a call that drops after the second question is not Incomplete. It gets a normal verdict, and the third question, which carries 2 of the 10 points, simply adds nothing.

Take an invented candidate on a points job. She earns 2.5 of 3 on the first question, 2.5 of 3 on the second and 1.5 of 2 on communication, and then her connection dies before the third question is asked. Her total is 6.5. That is On Hold, above the 6.0 line but under the 7.5 needed for Go. Had she finished and scored even 1 out of 2 on the last question, she would have reached 7.5 and landed in Go.

Nothing about this is hidden. Open the interview and scroll to the recruiter scoring panel: the third question shows a dash where the AI’s number would be, and the transcript shows where the call stopped. But a recruiter skimming verdict pills would not see it. So when you work through On Hold and Go rows on a points job, glance at that panel for a missing third score. If one is missing and the transcript ends abruptly, treat the verdict as a floor rather than a finding, and consider resetting the booking so the candidate can finish.

Why not require all three? The trade-off runs both ways. Demanding three would push every call with one unscored question into Incomplete, including ones where the candidate finished but the evaluator could not score the final answer. Two of three keeps genuine attempts scored. The price is this edge case, and naming it seems better than pretending it is not there.

What HR sees for a partial call

In the inbox Screened tab, an Incomplete candidate carries a purple Incomplete pill in the AI screen column in place of a verdict. On points-rubric jobs, a second chip appears, marked “abandoned”, whose tooltip says the screen was flagged, not scored low. The campaign dashboard counts Incomplete calls as their own figure beside the verdict counts, so a spike on one drive is visible at a glance.

Open the interview and the cause is usually clear within a minute of reading: a transcript that stops mid-answer, a recording that ends abruptly, or a candidate saying they need to leave. If the candidate spoke for under 30 seconds in total, speech analysis is not run and any scoring leans on the transcript alone, one more reason not to read much into a very short call.

What to do with each kind of Incomplete row

Different causes need different responses:

Resetting the booking is the standard retry. It works from a screened or broken call, keeps the old interview on record marked as superseded, reopens booking for the candidate, and fully resets the reschedule limit, so someone with a genuine connection problem is not stuck against the usual cap of three reschedules in 72 hours. It is refused while the candidate is inside a live call, so nobody gets cut off by accident.

One caution about the last row. Deciding by hand on a single answer is sometimes the right call on a closing drive, but write down why in your decision notes. Your call sits on record next to the machine’s, as set out in keeping both decisions on file. And a partial call that also carries a knockout quote missing from the transcript will already show as On Hold, thanks to the quote check that runs before any reject.

What to tell a candidate whose call dropped

Most candidates whose call dropped assume they failed. Many will not ask. A short message after you reset the booking changes that, and it costs you a minute. Something like this works:

“Hi Priya, your screening call for the Customer Support role ended before all the questions were covered, which looks like a connection problem on the line. That is not counted against you. We have opened a fresh slot for you. Please pick a time that suits you from the link, and if you can, use Wi-Fi and a quiet room.”

Three things make it work. It says plainly that the drop does not count against them, which is true and is the thing they are most worried about. It gives a concrete next step. And it offers one practical tip without implying the candidate did anything wrong. Avoid promising a specific outcome, and avoid mentioning scores; an Incomplete has no meaningful score to share.

If the transcript shows the candidate chose to leave, the message is different. Thank them, ask once whether they would like a new slot or a conversation with a person instead, and respect a no. Some people do not want to speak to an AI, and a drive that forces them to loses good candidates for no gain.

When Incomplete calls point at your setup, not your candidates

If Incompletes cluster, study the cluster before you judge the people in it.

Many on one slot. Our capacity guard admits a candidate only while live interviews are under the ceiling, 25 concurrent calls on production, and turns extras away with a busy message rather than a broken call. But if your invites all land on the same few minutes, more people wait and more give up. Spreading slots helps, and the no-show guide covers the rest of the drop-off chain.

Many in the evening, on phones. When drops cluster at night on mobile, patchy data is a common suspect. Tell candidates in the invite to use Wi-Fi if they can and to close other apps first.

Many in the first minute with no knockout. Usually mic permissions or a browser issue. A clear pre-call checklist removes most of these.

Many that are really poor audio, not drops. A completed call with a bad recording is a different case: the call counts, but communication is then judged from the words alone, without the audio side. The poor audio guide covers that case.

Our pilots now total more than 1,099 interviews, and the Incomplete rows that caused trouble were rarely technical mysteries. They were rows nobody opened, sitting quietly between the verdicts that did get actioned.

Clear your Incomplete pile today

Log in to HireQwik, pick your most recent job, go to the Screened tab and look for the Incomplete pill. For each row, read the last thirty seconds of the transcript and match it to a row in the table above. Most take under a minute. If you see a pattern you cannot explain, send it to us and we will dig into it together.