Quality
How we get to 99%
Accuracy is a measurement, not a slogan. This page sets out how we produce it, how we check it, and where the limits are.
Where AI stops
AI drafts. People decide.
On our AI-drafted tier, a machine produces the first draft and nothing else. It does not review its own output, it does not approve anything, and it never determines what reaches you. Every subsequent pass is performed by a person.
On our human-from-first-pass tier, there is no machine draft at all. Either way, the review and quality assurance that make a transcript usable are human work, in multiple passes, every time.
Method
Every transcript goes through four passes
01
Draft
Either a transcriptionist types from the audio, or a machine produces a first draft. This is the only point at which software touches the words.
02
Verification against audio
A transcriptionist plays the recording against the draft end to end, correcting words, restoring what was missed, and marking anything genuinely unclear with a timestamp.
03
Structure and house style
Speaker attribution is checked at every handover. Question and answer structure, exhibit marks, headings, certificate pages and your template conventions are applied.
04
Final quality assurance
A second reviewer checks the finished document against the brief: format, completeness, flagged passages, and the specific details that matter in your field — figures, drug names, claim numbers, names of people.
Measurement
What 99% actually means
Accuracy claims are meaningless without a method. Ours is word-level: we sample completed transcripts, compare them against the audio, and count every substitution, omission and insertion as an error. Ninety-nine per cent means no more than one error in every hundred words on a verified transcript.
Punctuation and formatting judgement calls are not counted as word errors, but they are checked separately in pass three. Passages we could not resolve are not counted as errors either, because we do not silently guess at them — they are marked in the document with a timestamp so you can make the call.
Limits
Where we stop
We do not guess at inaudible passages. They are timestamped and flagged.
We do not use your audio to train any model, ever.
We do not accept a deadline we do not believe we can meet.
We do not let anything reach you without a human reviewing it first.
FAQ
Common questions
Is any part of the review automated?
No. Automated checks assist our transcriptionists with mechanical formatting, but every review and quality-assurance decision is made by a person. Nothing is approved or delivered on a machine's judgement.
What happens if I find an error?
Tell us and we will correct it. One round of revisions is included with every transcript, and we will explain what went wrong.
Do you guarantee 99% on any file?
We quote it on verified tiers for audio of usable quality. If a recording is bad enough that we cannot reach it, we tell you before you commit rather than after we deliver.