What to Ask Your AI Transcription Vendor

Law firms evaluate vendors for court reporting, e-discovery, and document review with standard due diligence processes. AI transcription deserves the same treatment — and the questions are different from what you'd ask a staffing agency or a software vendor.

The right questions for an AI transcription vendor focus on three areas: what happens to the data, how accurate is the output for legal audio, and what recourse do you have when quality falls short.

Quick Navigation

• Data Retention and Deletion
• AI Training Use of Your Audio
• Security and Encryption
• Accuracy for Legal Audio
• Speaker Identification Quality
• Pricing and Preview Policies
• BrassTranscripts on Each Question
• Frequently Asked Questions

Data Retention and Deletion

The first question is the most fundamental: How long do you retain uploaded audio files, and when are they deleted?

This question has a right answer for legal use. Audio retained indefinitely is audio that can be breached, subpoenaed, or accessed by the vendor's own employees indefinitely. Audio deleted within 24-48 hours of processing limits that window to nearly nothing.

Follow-up questions worth asking:

• Is deletion automatic, or does the attorney need to trigger it manually?
• Does deletion remove the file from backups as well, or only from the primary system?
• What is your data breach notification policy?

Vendors who can't answer these questions specifically haven't built their data handling with legal clients in mind.

Transcript retention is a separate question from audio retention. Some services retain transcript text longer than audio. A transcript of a privileged client meeting is itself a privileged document — it should be in your systems, not the vendor's.

AI Training Use of Your Audio

Do you use uploaded audio or transcripts to train AI models?

This is not a hypothetical concern. Several popular transcription services improve their underlying AI through user-uploaded data. The business logic is clear: more audio data produces better models. The legal implication is also clear: audio used for AI training is not solely in the vendor's hands for the purpose of your service. It becomes part of a dataset.

Ask for this policy in writing. Acceptable answers include:

• We never use customer audio or transcripts for AI model training.
• Customer data is used only to provide the transcription service; no training use of any kind.

Unacceptable or concerning answers include:

• Our models improve with use (without clarifying whether that means your data specifically)
• Vague references to "anonymization" without explanation of what that means
• Inability to produce a written policy

For attorney-client privileged recordings, a service that uses customer audio for AI training creates a meaningful ethical exposure. The audio has left the service relationship.

Security and Encryption

Transcription vendors handle audio files in transit and at rest. Standard due diligence questions:

In transit: Are files encrypted during upload and download? HTTPS is the minimum; end-to-end encryption for sensitive files is preferable.

At rest: Where are files stored, and is the storage encrypted? Cloud storage providers like AWS S3, Google Cloud Storage, and Cloudflare R2 all support server-side encryption. Ask which provider the vendor uses.

Access controls: Who at the vendor can access uploaded files? Is access logged? Is access to customer files limited to specific personnel for specific purposes?

Subprocessors: Does the vendor use third-party services that receive your audio? Transcription services sometimes pass audio to a different processing provider. Each subprocessor is an additional point of access.

These questions don't require technical depth to ask or evaluate. A vendor that can't answer them has likely not thought through their security posture for legal clients.

Accuracy for Legal Audio

What accuracy should I expect for legal recordings, and can I test before committing?

No standardized accuracy benchmark for AI transcription exists as of 2026. Accuracy claims without published methodology — "98% accuracy" with no citation — are marketing, not measurement. The most useful evaluation is a test with your own audio.

For legal transcription, the relevant accuracy dimensions are:

• Common legal terminology (deposition, plaintiff, subpoena, objection, voir dire, habeas corpus)
• Speaker identification across a deposition with attorney, witness, and questioning attorney
• Proper nouns — party names, expert witness names, geographic names relevant to the case
• Numbers and dates, which appear frequently in legal testimony

Request a trial transcription of a real but non-sensitive recording. An old deposition excerpt, a practice run with staff, or any recording that lets you evaluate real output against your expectations.

A vendor unwilling to offer a preview or trial period for legal evaluation is harder to evaluate accurately. BrassTranscripts offers a 30-word preview of every transcript before payment — enough to assess whether the first lines of the recording transcribed accurately.

The legal transcription accuracy considerations in the law firm guide covers specific review approaches for high-stakes uses like deposition work.

Speaker Identification Quality

For legal recordings, speaker identification is often more important than raw transcription accuracy.

A transcript that accurately captures words but doesn't distinguish the attorney's questions from the witness's answers requires substantial manual editing before it's useful for case preparation. A transcript with clean speaker separation — even with occasional word-level errors — lets you search for witness statements, attorney objections, and opposing counsel questions without reading the entire document.

Questions for vendor evaluation:

• How many distinct speakers can the system identify in a single recording?
• Does speaker identification work on phone call audio?
• How does the system handle speakers who sound similar, or speakers who interrupt each other?
• Are speaker labels consistent throughout a long recording (2+ hours)?

The practical test is a multi-speaker recording. Submit a deposition excerpt or a three-person meeting and evaluate whether speakers are correctly separated and consistently labeled. BrassTranscripts automatic speaker identification handles up to 6 distinct speakers per recording with consistent labeling across the full transcript.

Pricing and Preview Policies

AI transcription pricing models vary significantly. Common structures include:

Per-minute pricing — charged by the audio duration. Straightforward to calculate in advance. Common for services targeting high-volume users.

Per-file pricing — a flat rate per upload regardless of duration. Predictable for variable-length recordings; a 15-minute file and a 90-minute file cost the same. BrassTranscripts uses per-file pricing with two tiers based on duration.

Subscription pricing — a monthly fee for a set number of minutes. Efficient if your volume is consistent; wasteful if intake volume varies month to month.

For legal practice, per-file or per-minute flat rates with no subscription typically suit the workflow better than subscriptions. Case volume varies; subscription fees don't.

Preview policies matter. A vendor offering no preview before payment means you pay for a transcript before knowing whether it's usable. For legal audio with variable quality — phone calls, old recordings, noisy environments — a preview policy protects you from paying for transcripts that are too inaccurate to use.

BrassTranscripts shows the first 30 words of every transcript before payment. That's enough to see whether the recording was processed accurately at the beginning of the file. If the first 30 words are wrong, the full transcript is likely to have quality problems worth reviewing before committing.

BrassTranscripts on Each Question

For reference, direct answers to each due diligence question:

Question	BrassTranscripts Answer
Audio retention	Deleted 24 hours after upload (automatic)
Transcript retention	Deleted 48 hours after processing (automatic)
AI training use	Never — audio and transcripts not used for training
Storage	Cloudflare R2, encrypted in transit and at rest
Subprocessors	Cloudflare R2 storage (disclosed)
Pricing model	Per-file: $2.50 (1-15 min), $6.00 (16-120 min)
Preview	30-word preview before payment
Speaker identification	Up to 6 speakers, consistent labeling
Bulk pricing	$6.00/file (1-5) to $3.00/file (250+), no minimum
Satisfaction guarantee	100% money-back, no questions asked

This table covers the questions a legal client should ask any vendor. Compare the answers you receive with what you need for your practice's data handling requirements.

For the full legal use case coverage — deposition workflows, AI analysis prompts, and bulk case file processing — see the AI transcription for law firms guide.

Frequently Asked Questions

What Is the Most Important Data Question to Ask a Transcription Vendor?

Ask whether the vendor uses uploaded audio or transcripts to train AI models. Some services improve their models by processing user-uploaded recordings — which means client audio you upload may become training data for a commercial AI system. For legal practice, audio used in AI training is no longer solely in the vendor's possession for the purpose of your service; it becomes part of a larger dataset. Vendors with policies prohibiting training use of customer data provide a stronger privacy guarantee for attorney-client privileged material.

How Long Should a Transcription Vendor Retain Audio Files?

For legal practice, the shorter the better. Audio files retained indefinitely create an ongoing exposure — a vendor breach months later could expose audio that should have been deleted. Best practice is automatic deletion within 24-48 hours of processing. BrassTranscripts deletes audio 24 hours after upload and transcripts 48 hours after processing, with no manual step required from the attorney.

What Accuracy Claims Should I Be Skeptical Of?

Be skeptical of any accuracy percentage cited without a published source or methodology. Claims like "98% accuracy" or "highest accuracy in the industry" without a citation link to testing methodology, published research, or independent benchmark data are marketing assertions, not measurements. No standardized industry benchmark for transcription accuracy exists as of 2026. For legal use, the most useful question is not a percentage but: "Can I review the transcript before paying, and what is the refund policy if the quality is insufficient for my needs?"

Should I Ask About the Vendor's AI Model Specifically?

You can ask, but vendors may decline to disclose their specific AI stack for competitive reasons. What matters more than the model name is the output quality for your use case. Test with a real recording — a deposition excerpt, an intake interview, a phone call — and evaluate the result. Speaker identification accuracy, legal terminology handling, and timestamp precision are more useful evaluation criteria than model identity.