How to Automate HR Interview Transcription with Make.com and AI

Interview transcription is one of the clearest examples of a high-volume, rules-based HR task that should never be done by hand. Every recorded interview follows the same path: audio file lands somewhere, someone has to turn it into text, someone has to extract meaning from that text, and that meaning has to reach a candidate record. That path is fully automatable — and once you automate it, you free recruiters to do the work that actually requires human judgment. This guide walks through exactly how to build that workflow using Make.com™, a transcription API, and GPT-powered summarization, as part of the broader smart AI workflows for HR and recruiting with Make.com™ strategy. Structure first. Intelligence second. Always.

Before You Start

This workflow requires the following before you open the scenario builder:

• A governed cloud storage location where interview recordings land consistently — one bucket, one folder convention, owned by HR or IT. Do not build on ad hoc storage.
• An API key for a transcription service — Google Cloud Speech-to-Text, Azure Cognitive Services Speech, or OpenAI Whisper. You need one provisioned and tested before building.
• An OpenAI API key for GPT-powered summarization.
• Confirmed ATS write access — verify your ATS supports inbound API writes and that you have the field mapping documented. This is the step most teams skip and regret.
• Data governance sign-off — confirm which data stores are approved for audio and transcript retention. Do not improvise compliance mid-build. For the full compliance architecture, see the guide on securing Make.com™ AI HR workflows for data and compliance.
• Time budget: 4–8 hours for a working proof-of-concept; 2–3 days for a production-ready workflow with error handling.
• Risk acknowledgment: AI-generated transcripts can contain errors, especially with heavy accents, crosstalk, or poor audio quality. Plan a human spot-check step from day one. Do not remove it until you have confirmed error rates against your own audio library.

Step 1 — Map Your Data Flow Before Touching Make.com™

Draw the complete recording-to-ATS handoff on a whiteboard before you open the scenario builder. Skipping this step is the single most common reason these builds stall.

Answer four questions in writing:

1. Where does every interview recording land? Name the specific folder or bucket, the file naming convention, and who controls write access.
2. What transcription service will you use? Confirm the API accepts your audio format (MP3, WAV, M4A) and the file size limits involved.
3. What information do you need in the ATS candidate record? Full transcript? GPT summary only? Extracted skills list? Decide this now — it drives every downstream prompt.
4. Who reviews before the ATS write? Define the threshold for human review. A confidence score below 0.80 from the transcription API is a reasonable starting rule.

Document the answers. Share them with HR, recruiting, and IT before Step 2. The scenario build is fast. The alignment conversation takes longer and is more valuable.

Asana research finds that knowledge workers spend a disproportionate share of their week on low-value coordination and documentation tasks rather than skilled work — interview transcription is a textbook example of that drag, and it belongs in the automation layer, not on a recruiter’s task list.

Step 2 — Configure the Cloud Storage Watch Trigger

The trigger is the front door of your workflow. When a new recording appears in your designated folder, Make.com™ wakes up and starts the chain.

• In Make.com™, create a new scenario and add a Watch Files module for your storage provider (Google Drive, OneDrive, or an S3-compatible bucket via HTTP).
• Point the module at the specific folder where recordings land. Do not watch a root drive — scope it to the exact path.
• Set the polling interval. For most HR teams, every 15 minutes is sufficient. Real-time processing is rarely necessary and consumes operations unnecessarily.
• Add a File Size Filter using a Make.com™ router or filter step — reject files under 500KB (likely empty or corrupted) and files over your transcription service’s maximum file size limit.
• Add a File Type Filter — accept only your approved audio/video formats. Reject anything else with a log entry to a designated error sheet.

Test the trigger by uploading a sample recording. Confirm Make.com™ detects it within one polling cycle before moving to Step 3.

Step 3 — Route Audio to the Transcription API

This step passes the audio file to your chosen AI transcription service and receives structured text back.

• Add an HTTP Module (or the native connector if available) configured to call your transcription service endpoint.
• Pass the file as a binary upload or a signed URL, depending on what your transcription service requires. Signed URLs are preferable — they avoid moving raw audio through Make.com™ data stores.
• Enable speaker diarization in the API request parameters if your service supports it. This labels transcript segments by speaker (Speaker A, Speaker B) and is essential for separating candidate responses from interviewer questions in the GPT step that follows.
• Request a confidence score in the API response. Store it as a Make.com™ variable for the routing decision in Step 4.
• Map the transcript text output to a Make.com™ data store or pass it directly to the next module. Do not write raw transcripts to shared HR systems at this stage — that happens after GPT processing in Step 5.

Parseur’s research on manual data entry costs documents that a single manual re-keying error in a professional workflow can cost an organization more than $28,500 per affected employee annually when compounded across downstream systems. Routing audio through an automated transcription API with a confidence-score gate is a direct control against that cost vector.

Step 4 — Gate on Confidence Score (Human Review Router)

This step is not optional. It is the quality gate that makes the rest of the workflow trustworthy.

• Add a Router module in Make.com™ immediately after the transcription API response.
• Route A (High Confidence): If the returned confidence score meets your threshold (start at 0.80 and calibrate after 30 days of production data), pass the transcript to Step 5 for GPT summarization.
• Route B (Low Confidence): If the confidence score is below threshold, send a notification — via email or a Slack message — to a designated reviewer with a link to the original audio file and the raw transcript. Log the event to your error tracking sheet. Do not proceed to summarization or ATS write until a human clears the transcript.
• Build a Resume Path: the human reviewer edits or approves the transcript and triggers a webhook or form submission back to Make.com™ that resumes the scenario at Step 5.

In typical HR audio environments, expect 10–20% of recordings to fall below an 0.80 confidence threshold initially. That number drops as you improve recording practices — dedicated interview rooms, headsets, and muting background noise. Track your weekly rate; it is a leading indicator of audio infrastructure quality.

Step 5 — Run GPT Summarization and Structured Extraction

This is where automation hands off to AI. The transcript enters; structured, actionable candidate data exits.

• Add an OpenAI — Create Completion module (or equivalent HTTP call to the OpenAI Chat Completions endpoint).
• Write a system prompt that instructs GPT to act as an objective HR documentation assistant. The prompt should specify: output format (JSON), fields to extract (candidate name, interviewer name, date, skills mentioned, years of experience stated, behavioral examples provided, one-paragraph summary, any unresolved questions for follow-up), and explicit instruction to base output only on the transcript text — no inference beyond what was said.
• Include the diarized transcript — with Speaker A/Candidate and Speaker B/Interviewer labels resolved — in the user message. If diarization was not available, instruct GPT to infer speaker roles from context and flag low-confidence role assignments.
• Set a max token limit appropriate to your ATS field sizes. Oversized outputs cause write failures in Step 6.
• Parse the JSON response using Make.com™’s built-in JSON Parse module. Map each field to a named variable.

McKinsey Global Institute research consistently finds that AI applied to well-structured data inputs produces significantly more reliable outputs than AI applied to unstructured or ambiguous inputs. A clean, diarized transcript is the structured input that makes GPT summarization reliable. Garbage in, garbage out — the quality gate in Step 4 exists precisely to protect this step.

This is the same AI-plus-automation architecture that powers AI candidate screening workflows with Make.com™ and GPT and automating HR meeting note summarization with AI — the pattern is reusable across HR document types once you have the pipeline wired.

Step 6 — Write Structured Output to Your ATS

Structured candidate data now flows into your system of record.

• Add the native Make.com™ connector for your ATS (Greenhouse, Lever, Workday, BambooHR, or similar) or an HTTP module if no native connector exists.
• Map each GPT-extracted variable to the corresponding ATS field. Use the field mapping document you created in Step 1. Any field without a confirmed ATS destination should be written to a staging sheet rather than discarded — decisions about field mapping change, and you want the data preserved.
• Add an error handler on this module. ATS API rate limits and field validation errors are common. Configure the handler to log failures with the candidate ID, error message, and timestamp, then alert the HR ops owner.
• Write the one-paragraph GPT summary to a candidate notes field. Write extracted skills to a structured skills tag field if your ATS supports it. Write the full transcript to a document attachment or a linked secure document store — not to a free-text notes field.
• Do not write an automated score or ranking to the ATS. Extract structured information; leave ranking to humans. This is the distinction between information surfacing and automated decision-making, and it is where building ethical AI workflows for HR and recruiting begins.

Step 7 — Notify the Hiring Team

Close the loop with a lightweight notification that drives action without creating noise.

• Add a Send Email or Post Message module (Slack, Teams, or your collaboration tool of choice).
• The notification should include: candidate name, interview date, a direct link to the ATS record, the one-paragraph GPT summary, and any unresolved follow-up questions extracted in Step 5.
• Do not dump the full transcript into the notification. Link to it. Recruiters need a decision trigger, not a wall of text.
• For team environments, route the notification to the specific hiring manager tied to the open role — not a generic HR channel. Make.com™ can look up the hiring manager assignment from the ATS before sending.

Step 8 — Add Logging and Monitoring

A workflow without logging is a black box. Build observability in from day one.

• At the start of every scenario run, write a log entry — scenario ID, trigger timestamp, audio file name, file size — to a dedicated Google Sheet or Airtable base.
• At the end of every successful run, write the completion timestamp, confidence score, and ATS write status.
• On any error path, write the error type, module name, and candidate/file identifier.
• Set a Make.com™ alert to notify the HR ops owner if more than three consecutive runs fail. Three failures in a row indicates a systemic issue — API key expiry, storage permission change, or ATS field validation change — not a one-off error.
• Review the log weekly for the first 60 days. After that, monthly review is sufficient if error rates are stable.

How to Know It Worked

Measure these four signals at 30 and 60 days post-launch:

1. Processing time per interview: Average time from recording upload to ATS write-back. Target under 5 minutes for high-confidence transcripts. Manual baseline for comparison should be documented before launch.
2. Recruiter time freed per week: Survey the team. If your team processed 30 interviews per week at 45 minutes of manual transcription and summarization each, a 70% automation rate frees roughly 16 recruiter hours weekly.
3. Human review rate: Percentage of transcripts routed to the Step 4 human gate. Track the trend. A declining rate means audio quality is improving. A rising rate means something in the recording environment changed.
4. ATS write error rate: Percentage of scenario runs that fail at Step 6. Target under 2%. Above 5% indicates a field mapping or API stability issue that requires immediate attention.

Gartner research on HR technology adoption consistently finds that time-to-value for automation projects is compressed when clear success metrics are defined before launch — teams that define “done” in advance reach stable production faster than those that iterate without a target.

Common Mistakes and Troubleshooting

Mistake 1 — Building before confirming ATS write access

The transcription and summarization steps work. Then the ATS module fails because the API isn’t enabled on your tier, or the field names in documentation don’t match the actual API payload. Confirm write access with a test call before building Steps 3–7.

Mistake 2 — Skipping the confidence-score gate

Teams that skip Step 4 to save build time discover the problem the first time a poor-quality recording produces a hallucinated or garbled GPT summary that reaches the ATS. The gate is cheap insurance. Build it first.

Mistake 3 — Writing raw audio or full transcripts to shared HR systems

Raw audio and verbatim transcripts carry significant data sensitivity. Route them to governed, access-controlled storage — not to shared drives, team channels, or ATS free-text fields. Write links, not files, to shared systems.

Mistake 4 — Using GPT output for candidate scoring without human review

Extracting structured data is appropriate. Generating a numeric score that influences whether a candidate advances is a different act with legal and ethical implications. Keep scoring in human hands. Surface information; do not make decisions.

Mistake 5 — Not logging errors with enough context to diagnose them

A log entry that says “error” without the module name, candidate ID, and error code is useless. Write structured log entries from the beginning. Diagnosing a production error at 11 PM is dramatically easier when your log tells you exactly which step failed and why.

Scaling This Workflow Across the Full HR Pipeline

Once interview transcription is running in production, the same pipeline pattern — trigger, API call, GPT extraction, structured write, notification — applies across the HR document stack. Automating performance review summaries with Make.com™ and AI follows the same architecture. So does AI-powered resume analysis. The pattern is the asset — not any individual workflow.

SHRM research on cost-per-hire and time-to-fill consistently finds that organizations with faster, more accurate candidate documentation pipelines make better hiring decisions at lower cost. The transcription workflow you build here is one module of a larger HR automation stack. For the full module inventory, see essential Make.com™ modules for HR AI automation, and for the business case behind the investment, see the ROI case for Make.com™ AI workflows in HR.

Microsoft Work Trend Index research documents that workers spend a significant portion of their week on tasks that could be automated — documentation, transcription, and data re-entry consistently rank at the top of that list. The workflow in this guide eliminates all three for every interview your team conducts.

Structure the automation first. Let AI handle the judgment layer. That sequence — applied consistently across HR operations — is the entire thesis behind smart AI workflows for HR and recruiting with Make.com™.