You can generate a podcast episode in under 20 minutes. The question is whether anyone will listen to the second one.
The “AI podcasting is instant” pitch is everywhere right now — and it’s half true. The generation part is fast. The part where listeners decide you’re worth their commute? That’s where the workflow either holds or collapses.
Caveat: This isn’t for zero-budget readers. If you can’t spend £30–£80/month on tools and aren’t willing to put human thinking into the narrative layer, close this tab.
AI podcasting done properly isn’t free, and done cheaply it’s worse than not doing it at all. Thou shalt not AI slop.
Descript Just Repriced — And the AI Credit Model Is Burning Creators
For the uninitiated, Descript is an AI-powered audio and video editor that streamlines podcasting by allowing users to edit recordings by editing a text transcript. It features AI tools (Underlord) for removing filler words, enhancing audio quality with “Studio Sound,” generating show notes/titles, and creating social media clips.
Before you commit to any stack, you need to know what happened to Descript’s pricing in early 2026.
Users in the Descript subreddit are blunt about it: “AI credits burn too fast; pricing makes Descript unusable” — that’s a direct quote from a March 2026 user thread. This isn’t a fringe complaint. The credit-metering model means your editing costs scale with usage in ways that aren’t obvious when you sign up.
The practical consequence: one user reported on Trustpilot that “I’ve edited only three podcast episodes and I’m already out of AI credits.” Three episodes. That’s a pilot run, not a show.
Meanwhile, the interest in AI podcast generation is real and growing. A Hacker News thread from late 2024 documents a user who created an example podcast from a 38-page PDF using NotebookLM — generating a full audio overview with no recording equipment. That capability exists. The demand is there.
The problem is the gap between “this is possible” and “this is sustainable.” Pick the wrong platform or ignore recent pricing changes and you’re looking at surprise bills or broken workflows costing £100–£400/year. That’s not a rounding error for a side project.
Why Full-AI Generation Produces Content Nobody Shares
Here’s the mechanism that most “AI podcast tutorials” skip entirely.
NotebookLM can generate a podcast from a 38-page PDF. The audio overview feature produces two synthetic voices discussing the content. It’s technically impressive. It’s also immediately recognisable as artificial — and listeners know it.
The Reddit verdict on AI-only podcasts is not kind. One commenter put it plainly: “There is no value in them.” (source) That’s not a minority position. It reflects something real about how humans process conversation — we’re wired to detect emotional authenticity, and full-AI generation doesn’t pass that test at scale.
AI pushes complexity into the generation layer. Human narrative pushes complexity into the editing layer. The shows that work do both.
The workflow that survives is not “AI does everything.” It’s AI handling the structural and production load — script drafting, voice synthesis, audio cleanup — while a human controls the narrative arc, the opinion layer, and the moments that make someone share an episode with a friend.
Rely on full-AI generation and you lose emotional engagement. Audience evaporates. Word-of-mouth doesn’t happen. You’ve built a show no one listens to — and spent real money doing it.
The Workflow: How to Build an AI Podcast Episode Without Burning Credits or Losing Listeners
This is the step-by-step. Follow it in order. The credit management step is not optional.
Step 1: Source material and narrative brief
Start with a document — a PDF, a research paper, a blog post, a client brief. This is your raw input. Before you touch any AI tool, write a 3–5 sentence narrative brief: what’s the opinion of this episode? What do you want the listener to think differently about by the end? AI can’t generate this for you. This is the human layer.
Step 2: Script generation via NotebookLM or equivalent
NotebookLM is currently the cleanest free option for turning source documents into podcast-style audio overviews. A user documented the process on Hacker News — upload the PDF, generate the audio overview, export the transcript. Use the transcript as a structural draft, not a final script. Edit it against your narrative brief from Step 1.
Step 3: Voice synthesis — credit-aware
This is where you need to be careful. ElevenLabs costs from £5/month on the Starter plan and is the current standard for voice quality in AI podcast production. I’ve used ElevenLabs in production for 4–5 months — voice quality holds across long-form content, but you need to monitor your character count against your plan limits before you generate, not after.
Set a hard rule: calculate your estimated character count from the script before hitting generate. A 2,000-word script runs approximately 12,000–14,000 characters. Know your plan ceiling before you start, not halfway through a generation queue.
Step 4: Audio editing — credit-aware (again)
If you’re using Descript for editing, the AI credit situation is live and material. Descript plans start around £12/month, but given the March 2026 pricing complaints, audit which editing features consume AI credits versus which are standard. Transcription-based editing (cutting by text) is the core value. Overdub and AI cleanup features are where credits disappear fast.
Alternative Option: use Descript for transcript editing only, and handle audio cleanup in Audacity or a DAW you already own.
Step 5: Distribution and monetisation verification
Before you publish, check the platform’s AI content policy. Spotify, Apple Podcasts, and YouTube all have evolving positions on AI-generated audio. Declare AI involvement where required. Skipping this step doesn’t save time — it risks account flags that cost you the entire back catalogue.
Miss any of these steps and you’re looking at 2–6 hours of remediation and £20–£60 wasted per pilot episode.
“Minutes” Is the Demo. Here’s What Production Actually Takes
Every AI podcasting demo shows a finished episode in under 10 minutes. Production shows something different.
The NotebookLM demo is real — a 38-page PDF becomes a podcast in minutes. That part is accurate. What the demo doesn’t show: the narrative brief you need to write first, the script editing pass, the voice synthesis queue time, the audio review, the show notes, the thumbnail, the distribution upload, the metadata.
Realistic timeline for a properly produced 20-minute AI-assisted episode, first time through:
- Narrative brief + source prep: 30–45 minutes
- Script generation + editing: 45–60 minutes
- Voice synthesis + review: 20–30 minutes
- Audio editing + cleanup: 30–45 minutes
- Distribution + metadata: 20–30 minutes
Total: 2.5–3.5 hours. Not minutes. The tools have their own learning curve so don’t expect these results in month 1.
By episode 10, with a repeatable workflow, you can compress this to 60–90 minutes. That’s the realistic efficiency gain. Believe “minutes” without planning and you’ll spend extra hours fixing pacing errors and billing disputes — delaying launch and increasing cost.
The Beginner Shortcut That Kills Shows Before Episode 5
The most common mistake: publish AI-only content and call it a podcast.
Listeners are not fooled. “They feel like AI — not a real conversation” — that’s a direct quote from a Reddit thread on AI podcast quality. And the bluntest version: “There is no value in them.” (source) These aren’t edge cases. They’re the majority listener response to full-AI generation with no human narrative layer.
The second mistake is voice cloning without verification. ElevenLabs is the market leader for AI podcast voice synthesis, but cloning accuracy is not guaranteed. One Trustpilot reviewer documented “Poor Voice Cloning Accuracy — the cloned voice had only ~50–60% resemblance” to the original. (source) Publishing a cloned voice that sounds 50% like you — or 50% like your client — is a brand problem, not a production shortcut. Worth it if you’re doing consistent, high-volume production and have time to test clones before publishing. Skip it if you need a voice clone to be ready on first attempt without a verification pass.
Further, you have a lot of parameters you can tweak to tune your output. At a minimum, the top 2 to be aware of:
- The AI model you are using. Eleven Multilingual v2 is the recommended default but there is a cheaper Flash or expensive V3 model. Try V2 or V3.
- Tweak speed. Make it balanced, somewhere on 0.81 was my best result.
My advice: Record one sentence (from text to speech) and try some variations.
I don’t recommend publishing AI-only podcast content without a clear disclosure and a human editorial layer. Not because of regulation (though that’s coming), but because it’s a bad product. Listeners churn after a few episodes. The show dies. You’ve spent money and time building something with no audience.
The shortcut costs more than the proper workflow. Publish AI-only content and you damage your brand — listeners churn after a few episodes, and voice-clone failures hurt reputation in ways that are hard to recover from.
Your Workflow
Use AI for what it’s good at: structural drafting, voice synthesis, audio cleanup, transcript editing. Use your brain for what AI can’t do: opinion, narrative arc, the reason someone should care.
The stack that works right now:
- NotebookLM — free, for source-to-structure generation
- ElevenLabs — for voice synthesis, with character-count discipline. This is the main voice tool I recommend for AI podcast production. Worth it if you’re producing 2+ episodes per month and have a consistent source format. Skip it if you’re producing one episode per quarter — the subscription cost doesn’t justify it at that volume.
- Descript — for transcript-based editing only, with AI credits rationed tightly. Worth it for the text-based editing workflow alone. Skip the AI cleanup features until you’ve mapped exactly how fast they drain your monthly credit allowance.
- Human editorial layer — non-negotiable, not a tool
Start with one episode. Run the full workflow. Time every step. Then decide if the economics work for your use case.
Pick this workflow if you’re producing 2+ episodes per month and willing to invest 2.5–3.5 hours per episode up front. The per-episode cost drops sharply by episode 10, and the quality gap over full-AI output is the difference between a show that grows and one that flatlines. Skip it if you’re experimenting once per quarter — use the free tier of NotebookLM and a basic DAW instead, and revisit when your volume justifies the tooling cost.
Build episode one this week. Time every stage. By episode three you’ll know exactly where your workflow leaks time and money — and that’s the version worth scaling.