Ask an HR director why people leave and you get an answer in about three seconds. Compensation. Career path. Their manager.
Ask the same person what leavers said in the third quarter specifically, and whether it was different from what they said in the first, and the answer takes much longer. Often it does not arrive at all.
That second question is the one worth answering, and the reason it goes unanswered has nothing to do with effort. Your team ran the interviews. Somebody sat in every one of them. The information existed, out loud, in a room. It just never survived the trip from the conversation to the file.
What survives an exit interview is a category, not a sentence
Here is what actually happens in the twenty minutes after someone tells you why they are leaving.
The interviewer, who has been half listening and half typing, opens the form. The form has a dropdown. The dropdown has eight options. The interviewer picks the closest one, writes two lines in the free text box, and moves on to the next thing in a day that has eleven other things in it.
Everything the person said gets compressed into a word.
By the time it reaches the annual attrition review, that word has been counted alongside forty other words and turned into a bar chart. The chart says compensation is the leading reason. It said that last year. It will say it again next year, because the compression happens before the analysis, and no amount of analysis downstream can decompress it.
What you have is not a dataset. It is a tally of what one interviewer thought each conversation was mostly about, recorded while they were distracted.
The part you lose is the part that was actionable
Compensation is not a finding. It is the category that anything vaguely money adjacent falls into, and it covers at least four different problems that require four different responses.
Someone who left because a competitor paid 30% more has told you about the market. Someone who left because their promotion was approved and then quietly deferred twice has told you about a specific manager. Someone who left because the bonus formula changed in a way nobody explained has told you about internal communication. Someone who left because they were doing two jobs after a colleague resigned and the pay never reflected it has told you about headcount planning, which is a finance problem wearing a compensation costume.
All four get logged as compensation. All four produce the same recommendation, which is to benchmark salaries, which fixes one of them.
The distinction lived in the sentences people used. It never made it to the form.
There is a second thing you lose, and it is subtler. Nobody decides to leave during the exit interview. They decided months earlier, and if you listen properly they will usually tell you when. There was a meeting. A project got reassigned. A new person was hired above them without a conversation. That moment is the most operationally useful thing in the entire interview, because it is early enough that you could have done something, and it almost never fits in a dropdown.
Before you record anything, get the legal and cultural part right
None of what follows works if you skip this section, and the failure mode is not a bad analysis. It is a complaint.
Recording an exit interview is not the same as recording a client call. The person in the room is in an unequal position, is often upset, and is about to lose their access badge. Consent that is technically obtained but practically coerced is worthless to you both legally and analytically, because a person who feels recorded against their will tells you nothing true.
The rules vary more than most HR teams assume. Under GDPR you need a lawful basis for processing, and consent from an employee is a weak one precisely because of that power imbalance, so most European employers rely on legitimate interests and document the assessment. In Germany, monitoring measures that touch employee behaviour generally engage the works council’s co-determination rights, which means the Betriebsrat is involved before the first recording, not after. In the United States, recording consent rules differ state by state, and in several states the requirement covers confidential in person conversations rather than only phone calls, which is exactly the situation an exit interview is.
The practical version, which works nearly everywhere:
Ask at the start, out loud, on the recording. Explain what it is for, which is improving the process rather than evaluating the person. Say who will hear it and how long you keep it. Make declining genuinely free by having a written note fallback ready that you use without visible irritation. Delete the audio on a fixed schedule once the text exists.
Then check whether your tooling supports the promises you just made. Encryption in transit and at rest, no data sharing with third parties, deletion on request, and a vendor that documents GDPR compliance rather than gesturing at it. Vomo covers those and can be deleted per recording, which matters when someone withdraws consent two weeks later and you need to actually honour it rather than explain why you cannot.
Expect a meaningful decline rate in the first year. That is a healthy sign. A 100% acceptance rate means people did not feel able to say no.
Turning a year of recordings into something readable
The mechanical part is genuinely boring, which is the point. It should not be the reason the project dies.
An exit interview runs 30 to 50 minutes. Twelve months of them at a mid sized company is somewhere between 20 and 60 recordings, sitting as M4A files from a phone or MP4 exports from a video call. That is 20 to 40 hours of audio. Nobody is going to listen to it twice, and the second listen is where the analysis actually happens.
An audio transcription tool turns that pile into something you can read in an afternoon and search in a second. The flow is four steps and takes about as long as making coffee.
Upload the file. M4A, MP3, WAV and FLAC all work, along with the usual video formats if your calls were recorded on Zoom or Teams. There is no cap on the length of a single file, which matters when an interview runs long and you do not want to be splitting audio.
Confirm the language, or let it detect. Roughly 50 languages are supported, which is the difference between running this globally and running it in your headquarters country only.
Let it process. You get punctuated text with speakers labelled automatically, plus a summary and the points it picked out. Accuracy sits around 95% on clean audio, lower if the room was noisy or people talked over each other, which they will.
Then read, correct and export. Speaker names can be fixed by hand in the editor, and the text comes out as TXT, DOCX, PDF or Markdown depending on where your analysis lives.
One thing worth setting up on day one: folders per quarter, not per person. You are going to want to read them as a batch, and file structure quietly determines whether you do.
Four things to read for, and nothing else
Reading transcripts looking for insight in general will produce nothing and convince you the method does not work. Read for four specific artefacts and write each one into a separate document.
The decision moment. The event, with a date if they give one, after which they started looking. Not the reason they cite. The trigger.
The verbatim sentence about their manager. Not your summary of it. The words, quoted, because “he never gave me anything to do that was hard” and “she took credit for the migration” become the same bullet point in a summary and require completely different interventions.
The named process. Every leaver names at least one internal system or ritual that wasted their time. The performance cycle. The approval chain for anything over a threshold. The Monday meeting. These are cheap to fix and almost never surface anywhere else, because current employees have stopped noticing them.
The comparison. What the new role offers that yours did not, in their words. This is your only free source of competitive intelligence on employment terms, and it is usually more specific than anything a compensation survey will tell you.
Once the transcripts are text, you can also just ask. Vomo’s chat feature lets you query a transcript directly, so “what did they say about the promotion process” returns the relevant passage instead of you scrolling. For a single interview that is a convenience. Across a quarter, it is what makes the exercise finish in an afternoon.
Read across the year, not within the conversation
The single interview is not the unit of analysis. It never was.
One person saying the review process is opaque is an opinion. Nine people saying it, in nine different phrasings, across three departments, over eleven months, is a finding, and it is a finding you cannot see if each conversation was compressed to a dropdown value at the moment it happened.
The pattern that tends to surprise people is temporal. Reasons cluster around events. A reorganisation in March shows up in resignations in July, described by people who will tell you it was about compensation, because compensation is the socially acceptable reason and reorganisation trauma is not. You only catch that by reading the decision moments in sequence and noticing they share a month.
The instrument is biased, and you should say so in the report
Exit interviews are among the worst survey instruments in common use, and pretending otherwise will get your conclusions dismissed by the first sceptical executive who reads them.
People are managing a reference. Nobody torches a manager they might need a recommendation from in eighteen months, so criticism arrives heavily softened and sometimes not at all.
The population is self selected in a specific direction. You are hearing from people who left. You are not hearing from the person who considered leaving in April, stayed, and is now quietly disengaged, and that person is more expensive than the leaver.
Timing distorts everything. The interview happens after the decision, which means the person has already built a coherent story about why, and coherent stories are edited.
And the audio itself has a failure mode that matters here more than usual. Transcription errors concentrate on proper nouns, which in this context means the names of managers, teams and internal systems. Before any quote goes into a document that another human will read, check it against the recording at that point. A misattributed sentence about a named manager is not a data quality issue. It is an HR incident.
Write these limits into the top of the report. Use the transcripts for language, patterns and specificity. Do not use them to measure satisfaction, and do not put a percentage on anything.
A rhythm that survives a real HR calendar
The reason this fails is never the analysis. It is that it gets scheduled annually, lands in the same week as compensation planning, and gets postponed.
Do it quarterly, in one sitting, with a small enough scope that it cannot be deferred. Transcribe as you go rather than in a batch at year end, because the ten minutes after the interview is the only time anyone will bother to correct the speaker labels. Read the quarter’s transcripts in one block. Pull the four artefacts. Write one page.
The one page has a single job, which is to name one thing that came up more than twice and did not come up last quarter. That is the whole output. Not a dashboard.
On cost, because it is the objection that arrives fastest: the free tier covers 30 minutes of transcription a week, which is roughly one interview, and is fine for testing whether your team will actually do this. Unlimited runs $1.92 a week. A quarter of transcription costs less than the catering for the meeting where you present it.
What changes
You stop reporting that people leave for compensation, growth and management, which everyone already knew and which no one can act on.
You start reporting that four of this quarter’s seven leavers named the same approval chain, that three of them started looking within six weeks of the same reorganisation, and that two used almost identical wording about a promotion that was approved and then deferred.
None of that is new information. All of it was said out loud, to someone on your team, in a room. It was there the whole time, in the one format nobody could read.
