Shipping thirty caption tracks without losing your glossary

Your VP said “thirty languages” on a Wednesday. Finance still uses one vendor for AI. Legal wants one DPA. You need files you can diff, not a hundred mystery exports from browser extensions.

Here is what trips most people up

Translators fix wording but forget numbers and units. A line that read “$9.99” becomes “9,99 €” in a locale that should have stayed in USD. Here is what trips most people up: nobody assigned a spot-check list for currency lines.

BYOK vs bundled credits

If procurement already pays OpenAI or another compatible host, route the job through a server-managed provider profile so prompts and short audio windows stay inside the contract you already signed while usage remains auditable. If you are mid-pilot and only need predictable invoices, bundled credits from Clipr beat anonymous SaaS add-ons—compare monthly pools on pricing.

Step-by-step

1. Import the source captions with speaker labels if YouTube exposed them.
2. Lock the glossary: drug names, competitors, banned phrases.
3. Generate target locales in batches small enough that a human can scan diffs the same day.

Timing tools and overlap warnings are summarized under features.

What if it does not work?

• Overlapping lines after MT: run overlap detection before merging WebVTT; split crowded lines even if it costs an extra subtitle index.
• Credits vanish mid-week: check whether BYOK jobs failed open—flip heavy locales to the server-managed provider route before Monday’s deadline.
• Brand rejects a locale: keep model version + glossary revision next to the exported file so you can answer “who approved line 412?” without guessing.

Closing

Automation should delete typing, not accountability. Pair batching with visible limits and a tier that admits how many devices you really use—again, see pricing.