We delivered an AI podcast producer for a North Charleston producer running nine client shows and fourteen episodes a month — one that transcribes every microphone separately, proposes each cut as a reviewable list rather than making it, drafts the notes, clips, titles and newsletter in each show's own voice, and turned eight hundred back-catalogue episodes into an archive she can search by describing what she remembers.

She spent nine years in commercial radio at a station off Rivers Avenue before the format changed and her job went away, and she started producing podcasts out of a spare bedroom in Park Circle with two microphones and a borrowed interface. That was six years ago. She now produces nine shows — a law firm's monthly interview program, a Daniel Island real estate team's weekly market update, a Lowcountry food show that outgrew its own ambitions, a church's sermon-and-conversation feed, a health system's patient-story series, and four others — which comes to roughly fourteen episodes a month, every month, indefinitely.

Her clients think they hired someone to make their audio sound good. What they actually hired was someone doing nineteen separate jobs per episode, most of which have nothing to do with sound. A ninety-minute raw recording arrives as a multitrack file: one channel per microphone, plus whatever the remote guest's connection produced that day. From there she cuts the false starts, the throat-clears, the four-second silences where someone is reading a text message, and the two minutes where the host and the guest talk over each other about parking.

She writes the show notes. She finds the timestamps. She pulls three vertical clips for social, which means finding three moments that are genuinely self-contained — a moment that requires context from eight minutes earlier is useless as a clip, and knowing the difference requires having listened to the whole thing. She writes five candidate titles in the voice of that particular show, because the law firm's audience and the food show's audience do not respond to the same sentence.

She matches the loudness to the platform target. She places the sponsor reads at breaks that do not feel like interruptions. She writes the newsletter. She uploads, schedules, and then answers the client's email asking where the episode is.

She had measured it once, out of curiosity and then regret: eleven hours of work for a forty-seven-minute episode. Fourteen episodes a month at eleven hours is a hundred and fifty-four hours, which is most of a full-time job before she has recorded anything, sold anything, or thought about a single show creatively. When she came to us she had six finished interviews sitting on a drive, recorded, paid for in gas and lunches and an afternoon each, and never released — not because they were bad, but because she could not face the eleven hours each of them represented. The second problem was quieter and, she assumed, unsolvable.

Across six years and nine shows she had produced more than eight hundred episodes. Every one of them contained things she remembered — a shrimper on the food show describing what Shem Creek looked like the morning after Hugo, an estate attorney explaining heirs' property in the clearest four minutes she had ever heard on the subject, a surgeon who went quiet for six full seconds before answering a question about a patient he lost. She could not find any of them. No transcripts, no index, no search.

Eight hundred episodes of the best material she had ever collected sat in a folder structure organized by date, which is to say it sat nowhere. We built her a production system that starts the moment her recorder writes the file. The multitrack is ingested automatically and each channel is transcribed separately, with the speaker attached to the microphone rather than guessed from the mix, which is why the transcripts stay accurate through the crosstalk that defeated every consumer tool she had tried. The system then produces an edit decision list — not an edit, a list.

Every proposed cut carries a timecode, a category, and a reason: filler word, false start, dead air over two seconds, crosstalk, off-topic tangent, the guest's phone buzzing against the table at nineteen minutes. She reviews it as a list and accepts or rejects each one, which takes about twelve minutes and leaves the editorial judgment entirely with her, because a four-second silence after a hard question is sometimes the best moment in the episode and no system should be allowed to decide that on its own. Once she approves the list the cuts are applied, the loudness is normalized to the target her hosting platform expects, and the episode is rendered.

Everything downstream drafts in parallel. Show notes with real timestamps that point at the actual moment rather than approximately near it. Three clip candidates, each scored on whether it stands alone without prior context, whether it has an arc rather than just one good sentence, and whether it lands between thirty and ninety seconds without being cut mid-thought. Five titles per episode, written in the voice of that specific show and learned from that show's own back catalogue rather than from podcast titles in general.

Sponsor read placements at the natural breaks the transcript reveals. A newsletter draft in the client's established voice. And a per-show status line the clients can see for themselves, which ended the where-is-my-episode emails entirely. The archive turned out to be the part she cared about most.

All eight hundred and change back-catalogue episodes were transcribed and indexed, and every new episode now joins them. She can find the shrimper describing Shem Creek in about four seconds, by describing it. When a client asks whether they have ever covered a topic before, she knows, with timestamps, across six years. In March a producer at a national outlet called looking for someone who could speak to heirs' property on Johns Island; she found the attorney, the episode, and the four-minute passage before the call ended, and her client got the placement.

Eleven hours became a little under two. The six-episode backlog went out over eleven days. She has taken on two additional shows without adding a person, which is the first real margin the business has ever had, and she has started recording a show of her own — the thing she meant to do six years ago and never once had a free evening for. "I did not lose eleven hours an episode to the craft," she said.

"I lost it to the tasks around the craft. Nobody has ever thanked me for a clean transcript or a well-placed ad break. They thank me for the conversation. Now the conversation is the only part I spend my time on."