Transcription is an input problem, not a transcription problem

A transcript with excellent word accuracy can still be almost useless to an agent, and the reason is structural rather than acoustic.

Consider a one-hour meeting rendered as a single unbroken block of correct words. Every term is right. Nothing indicates who spoke, where one topic ended, which sentence was a decision and which was someone thinking out loud. An agent asked to extract the action items has to reconstruct all of that from prose, and it will do so with confidence and occasional invention.

What an agent actually needs

• Speaker attribution. "We should postpone" means something different depending on who said it, and an agent cannot infer authority from text alone.
• Segment boundaries. Topic shifts are where retrieval boundaries belong. Without them the whole hour is one chunk, and one chunk is not searchable.
• Timestamps that survive processing. When an agent asserts that a decision was made, the ability to jump to the moment it was made is the difference between a claim and a citation.
• Confidence signalling. A passage the system found difficult should be marked as such rather than smoothed into fluent text that reads as certain.

These are not refinements on top of accuracy. They are what makes accuracy usable.

The confident-smoothing failure

The failure mode we see most often is a system that resolves ambiguity silently. Audio was unclear, and the output is a clean sentence with no indication that it was a guess.

Downstream, an agent treats that sentence as it treats every other sentence, and a reconstruction becomes a fact in a summary that someone acts on. The error did not originate with the agent, and the agent had no way to know.

A transcript that marks its own uncertainty is less pleasant to read and considerably safer to build on.

Where this fits in an agent system

Transcription is the start of a chain: audio becomes text, text becomes chunks, chunks become retrieved context, context becomes an answer someone relies on. A defect introduced at the first step is invisible by the last one.

This is why we treat transcript quality as part of the evaluation loop rather than a separate procurement decision. If your evals only score the final answer, a retrieval failure and a transcription failure look the same, and you will spend weeks tuning the wrong layer.

Practical advice

Evaluate on your own audio, not a clean sample. Real recordings have crosstalk, accents, background noise and the specific vocabulary of your organisation, and those are exactly the conditions where systems diverge.

Then evaluate the thing you actually want: not word error rate but whether an agent asked a real question about the recording retrieves the right passage and answers correctly. That measures the chain rather than a link in it.

Next

The transcription pages cover language coverage and deployment options, including running inside your own infrastructure where recordings cannot leave it. The Agent Orchestration System pages cover the evaluation layer that keeps the whole chain measurable.