Count verified CRM records, not audio minutes

Local call transcription is often compared by the cost of processing one minute of audio. For a sales manager, that is an incomplete metric. Until someone has checked the customer's name, amount, deadline and promised action, a transcript has not become a useful record. Automation pays off when it reduces the time spent per *accepted* CRM record without increasing the cost of mistakes.

This is about batch processing completed calls, not live assistance during a conversation. First establish what actually happens today: does an employee listen to the full recording, write a note from memory, or leave the outcome undocumented? Labor savings can only be measured against an existing process. If nobody previously reviewed recordings, automation creates new information; it does not save minutes that were never spent.

A small on-premises workflow

The phone system exports a completed call with its identifier and timestamp. The file enters restricted storage; a background worker checks the format, detects speech, runs automatic speech recognition (ASR), and stores the transcript with time references. A note template then asks for the topic, agreement, deadline and unresolved question. Fields without reliable evidence remain blank. An employee checks important details against the audio, and only then is the record saved to the CRM. Reprocessing the same call ID must not create a second contact or task.

Open Whisper models and the faster-whisper implementation can support a local prototype. Whisper's documentation lists model variants and approximate video-memory needs: around 2 GB for small, 5 GB for medium and 6 GB for turbo. These are planning guides, not promises of speed on a particular server. faster-whisper supports batch processing and voice activity detection (VAD); test the latter on short utterances so that a crucial “yes” or “no” is not lost. A local language model for drafting the note is optional. If a transcript and manual field entry suffice, do not add another processing stage and its cost.

The workflow needs access to call exports, storage with access and retention controls, a workstation or server for ASR, a job queue, CRM integration and a log of approvals. Check the basis for processing recordings and which staff may access them. Local execution avoids sending audio to an external API, but it does not by itself solve access control, backup or deletion.

A modeled calculation: where payback comes from

The following is a *modeled scenario*, not a measured company result or a supplier price. Assumptions: 800 recorded calls a month, averaging six minutes each. Today an employee spends four minutes after each call checking and filing a record. Their fully loaded cost is ₽700 per hour. With the new workflow, 70% of calls take 1.5 minutes to verify; the remaining 30%, because of noise or disputed details, still take four minutes. A human ultimately approves all 800 records. The share of an existing server, storage and maintenance allocated to this process is ₽6,000 a month; setting up the integration costs ₽60,000 once.

  • Current labor: 800 × 4 / 60 × 700 = about ₽37,300 per month, or ₽46.7 per verified record.
  • After automation: 560 × 1.5 + 240 × 4 = 1,800 review minutes, costing ₽21,000 in labor. Add ₽6,000 of infrastructure for ₽27,000 monthly, or ₽33.75 per record.
  • Modeled savings are about ₽10,300 a month. At unchanged volume, the ₽60,000 setup cost would pay back in roughly 5.8 months.

The result is most sensitive to the proportion of calls that really become quicker to review. If only 40% do, monthly savings fall to roughly ₽3,300 and payback stretches to 18 months. Below approximately 26%, the ₽6,000 of monthly running costs are no longer covered by labor savings. If the original filing task took two minutes rather than four, all figures must be recalculated. Do not count freed hours as a financial benefit unless the team can use them for other valuable work.

What can erase the gain

Transcription quality varies by language, accent, noise and specialist vocabulary. Whisper's developers explicitly warn that the model may insert words never spoken in the audio. Independent research has also investigated such errors on non-speech segments. In a commercial call, a wrong amount or promised date can cost far more than a few minutes of checking. Do not send critical fields to the CRM without human confirmation.

The calculation should include not only machine time but correction of bad notes, storage, backups, queue monitoring, model updates and integration maintenance. During a call-volume peak, the server may finish processing after the business deadline, making the result less useful. For a low-volume business, the present manual procedure may be cheaper than a dedicated local system. For a business required to record calls and keep audio off external services, on-premises processing may be a separate requirement, but its cost must still be visible in the budget.

A next step without buying a server

Take 100–200 completed calls of different kinds: short, noisy, and containing product names, amounts and commitments. Measure the actual time spent filing them today. Then run local ASR on an already available machine in a test environment, without automatically writing anything to the CRM. Staff should record how many notes could be approved quickly, how many required full replay, and where dangerous errors appeared. Replace the modeled 70% and 30% with those observed proportions. Decide on a server only after calculating the cost of a verified record and confirming that processing finishes within the working day.

Sources: [Whisper repository and model table](https://github.com/openai/whisper), [Whisper model card and limitations](https://github.com/openai/whisper/blob/main/model-card.md), [faster-whisper documentation](https://github.com/SYSTRAN/faster-whisper), [study of Whisper errors on non-speech audio](https://arxiv.org/abs/2501.11378).