A lot of company knowledge is trapped in videos that nobody wants to scrub at 1.5x for the twentieth time.
Training recordings. Founder rants that accidentally contain policy. Conference talks your team swears by. Screen shares where the real SOP is spoken while the cursor wanders.
Do not ask people to “watch the playlist.” Run a pipeline that turns video dumps into a searchable playbook humans and agents can use.
Why video alone fails as a knowledge system
Video is great for first learning. It is terrible for lookup.
Nobody wants to rewatch a 47-minute Zoom to find the one sentence about weekend admissions. Timestamps help a little. Searchable cards help a lot.
Also: spoken SOPs drift. If the only source is a video from last spring, you are institutionalizing a ghost.
The pipeline at a glance
- Collect URLs or files with a clear title and owner
- Transcribe
- Clean and segment
- Extract playbook cards
- Human review and conflict check
- Publish to the KB with metadata
- Schedule re-review
If you stop at step 4, you built an AI fanfic generator about your company.
Transcription is necessary, not sufficient
Good transcripts still contain jokes, ums, tangents, and “wait, ignore what I said two minutes ago.”
Keep the raw transcript for audit. Then produce a cleaned version that removes filler and marks uncertain sections. The cleaner text is what extraction sees.
For multi-speaker training calls, speaker labels matter. Who said it is sometimes as important as what was said.
Extraction into cards, not essays
Ask the model to emit structured cards:
- topic title
- applies when
- steps or rules
- exceptions
- warnings
- suggested escalation
- source video and timestamp range
- confidence (high/medium/low)
Low confidence cards go to a human queue. They do not go straight to the agent.
This format matches how agents retrieve and how staff skim.
Human review is the product
Reviewers check:
- Is this still true?
- Does it conflict with a written SOP?
- Did the speaker state a preference or a policy?
- Are we leaking anything that should not be searchable?
Preference vs policy is a frequent landmine. A manager venting “I usually just text them” is not a playbook entry.
Dedup and conflict handling
Videos overlap. People re-explain the same process with different numbers.
Cluster near-duplicate cards and force a merge decision. Conflicts get a ticket, not a silent average. The KB should reflect the decision after the argument, not during it.
Make it searchable for humans first
If staff will not use the playbook, your agent will be the only consumer, and staff will distrust the agent.
Focus on titles that match how people ask, a simple search UI, and links back to the moment in the video for skeptics. Skeptics are healthy. Give them a timestamp.
Permissions and common sense
Not every video should enter a general KB.
HR topics, private personnel issues, and anything regulated need a different lane or a hard exclude list. The pipeline should check a source allowlist, not vacuum every link in Slack.
Maintenance cadence
Every card gets an owner and a review date.
When call reviews show the agent citing a stale video-derived rule, fix the card and often retire the video as authoritative. Video becomes background. The card becomes law.
That shift matters. Teams stop pointing at folklore clips and start pointing at the playbook.
What this opens up for agents
Once cards exist, voice and chat agents can retrieve the same truths. Onboarding gets shorter. After-hours answers get less improvisational.
The win is not “we used AI on YouTube.” The win is that tribal knowledge became operable.
Bottom line
A YouTube dump is not a knowledge base. It is a pile of evidence that knowledge exists somewhere.
Run transcripts to cards, force humans to bless them, and give the result a home with owners and dates. Then your agents stop guessing, and your team stops rewatching the same monologue for a single rule.
