The recorder stops at 14:32. Deadline is 18:00. The interview was good, the kind where the source said three things you did not see coming and one thing they may regret saying. You have ninety minutes of audio. You have three and a half hours of clock. The last thing you want to do is listen to the recording back at 1.5 speed.

This is the workflow that gets the draft out by 18:00 without losing the parts of the interview that mattered.

A meeting AI used by journalists is one that returns verbatim quotes, attributes them correctly, and gives you enough variety in framing that you can pick the one that fits your angle. Multi-model summarization does this in one upload.

Step 1, Upload once, run two models

Start by uploading the recording to a single session. Run Claude Opus 4.7 for the careful exhaustive read, it preserves the moments where the source qualified, hedged, or backed off. Run ChatGPT 5.4 in parallel for the structured pull, it lays out the quotes cleanly and matches the rhythm of an editor's first pass.

You now have two summaries on the same audio. Read them side by side. Where they agree, the source said exactly that. Where they diverge, you know where to listen back to the recording, usually a 30-second segment, not the full 90 minutes.

This is the move. Two models on one upload, side-by-side reading. Cost: a few minutes of processing. Saving: an hour of re-listening.

Step 2, Use the free-text instructions to steer

The free-text instruction box is small but disproportionately useful for interview work. Tell the tool what artifact you want.

Three prompts that work for a long interview:

"Pull every direct quote longer than two sentences. Attribute to the speaker. Order chronologically by minute mark."
"Identify any moment the speaker hedged, qualified, or asked to go off the record. Quote both the question and the response."
"List every named person, organization, dollar amount, or date. One line of context each."

These do not replace the summary. They sit on top of it. The summary gives you the shape; the instructions give you the artifacts you need for the draft.

Step 3, Pull the verbatim quotes you'll actually use

A typical 90-minute interview yields somewhere between four and seven quotes that earn a place in the published piece. The rest is context, color, and the texture of the conversation. The summary will surface twenty candidates; you will keep five.

What you're looking for	What the summary should give you
The headline quote	One sentence the source said cleanly, that captures the angle without context required
The contradiction	A quote that sits against the conventional wisdom on the topic
The number / date / specific	A concrete claim that anchors the piece in fact, not opinion
The hedge	A moment where the source qualified or backed off, that signals nuance
The tone-setter	The line that sets the room, angry, weary, defiant, careful, without saying so

If the summary delivers all five categories, you have a draft. If it delivers two of five, the source did not say enough; that is on the interview, not the tool.

Step 4, From summary to outline to draft

The path from a multi-model summary to a publishable draft, by the clock:

14:32  Recording ends, you upload
14:34  Two models running in parallel
14:42  Summaries land. Read both side-by-side
15:00  Pull the five quotes that earn their place
15:15  Decide the angle, write the lede
15:45  Write the body around the quote scaffold
17:00  Self-edit pass; verify each quote against the recording
17:45  Send to editor
18:00  Done

The bottleneck is no longer transcription or note-taking. It is the editorial work of deciding what the story is. Summary tools do not write stories, they take the manual prep off the clock so the writing fits in the time you actually have.

A note on accuracy

The summary tells you what the source said. Before you publish a quote, verify against the recording. Not because the model is unreliable; because journalism's standard for direct quotes is "I heard this exact phrase", not "the model summarized this phrase". A 30-second listen-back per quote is twenty minutes for five quotes. Worth it.

For interviews under embargo, holding pattern, or off-the-record portions: drop those segments before upload, or use the summary on the full file but instruct the model to flag the off-record portion. Then you do not accidentally lift from a section the source asked you to keep out.

A note on retention

EnClair stores audio and summaries for 24 hours, then deletes both. We do not train models on user inputs or outputs. For a deadline-bound interview workflow, this matters: the recording does not sit in someone else's training corpus, and the summary you generated does not become part of a future model's training distribution. The full posture is documented on the security page.

What to take from this

Ninety minutes of interview audio is not the bottleneck anymore. The deadline is the bottleneck, and the decision of what the story is. Multi-model summarization gives you the prep, quotes, attributions, hedges, structure, in the time it takes to make a coffee. Use the rest of the afternoon to write.