What cheap automated transcription does well
On a single clear voice in a quiet room, modern recognition is very accurate, nearly instant, and costs cents or nothing. For a podcast draft, a lecture you attended, a meeting you want a record of, or a note you will read and correct yourself, paying for review would buy nothing.
It also scales without negotiation. Hundreds of hours can be processed without a contract, a procurement conversation, or a minimum commitment, which is a real advantage for anyone whose volume is irregular.
What the price excludes, specifically
Nobody checks it. The errors that reach you are the fluent ones - a similar-sounding name, a dose normalised to a more usual value, a dropped negation - and they read correctly, which means a quick scan does not catch them.
There is usually no BAA on a free tier, which makes the service unusable for protected health information regardless of how good the transcription is. Training use is common and often buried in terms: if your audio trains a model, it has been used for a purpose your patient or client did not authorise. Retention is the provider's decision, and deletion-on-request with confirmation is rarely offered.
And true verbatim is rarely available, because cleaning the text is what makes automated output look good. For research or legal work where a hesitation or self-correction is the point, silent cleaning removes material you needed and does not tell you it happened.
- No review stage, so fluent errors arrive unflagged
- No BAA, which rules out protected health information
- Audio often used for model training by default
- Retention set by the provider; deletion rarely confirmed
- Verbatim fidelity usually unavailable
- No record of who checked what, because nobody did
Four cases where it is the wrong tool
Anything containing protected health information, where the absence of a BAA is dispositive rather than a trade-off. Anything that will be signed into a record, because the signature asserts review and an unread signed document misrepresents what happened. Anything where a figure matters - a dose, a date, an amount - since those are exactly what fluent errors corrupt. And anything you may later have to prove was checked, because no cheap tier produces that record.
Outside those four, cheap automated transcription is usually the right answer, and a service charging dollars per minute for a podcast transcript is charging for something the buyer does not need.
Our position on this honestly
We sell both, at different prices, chosen per job. Unreviewed AI output starts at $0.45 per minute with no order minimum and returns in minutes; human-verified work starts at $1.75 with a $25 minimum. The finished document records which mode produced it, so an AI-only transcript says so plainly rather than staying silent about whether anyone looked at it.
We are not the cheapest in the market and would not claim to be - a free tier will undercut us on anything that does not need review. What we would say is that the decision belongs per file rather than per account, and a provider forcing one mode on all your material is selling you the wrong thing half the time.
Reading a free tier's terms for the three things that matter
Terms of service are long and mostly boilerplate, and three clauses carry nearly all the consequence. Search the document for them rather than reading it end to end.
Training use: look for "improve our services", "machine learning" or "derived data". If your audio can be used to improve the model, it has been used for a purpose nobody you recorded agreed to. Retention: look for a period, and note whether deletion is a right you hold or a request they may honour. Sub-processors: look for a list, and for whether they commit to notifying you when it changes.
What is usually absent is any statement about who may access your files internally. Its absence is itself informative - a provider with access controls worth describing tends to describe them.
- Training use: "improve our services", "machine learning", "derived data"
- Retention: a stated period, and whether deletion is a right or a request
- Sub-processors: a named list, and notice when it changes
- Internal access: usually unstated, which is itself an answer
A workable middle position
Most organisations do not need to choose one provider for everything, and the ones that do usually chose on price for their largest category and then used the same tool for their most sensitive. Running a cheap tier for internal and low-stakes material and a reviewed service for anything consequential costs less in total than either applied universally.
The thing to avoid is the drift: a free tool adopted for meeting notes that gradually gets used for a clinical dictation because it is already open in a tab. That is how protected health information reaches a provider with no BAA, and nobody decides it - it just happens. Deciding explicitly which material goes where is cheap and prevents exactly that.
Sources
- 45 C.F.R. § 164.308(b)(1) (business associate agreement before disclosure of PHI)
- 42 C.F.R. § 482.24(c)(1) (authentication of record entries by the responsible clinician)
Verified 19 September 2026.
The regulatory information on this page is general background compiled from public primary sources, not legal or compliance advice. Requirements change and vary by jurisdiction and by court. Verify current rules with the relevant authority or your own counsel before relying on them.