How I Split Work Between A VA And An Automation
People ask me this a lot once they find out I run a channel on AI automation and also pay a VA every month: why do you still need a person if you build agents for a living? The honest answer is that the va versus automation question isn't really a competition. It's a division of labor problem, and getting the split wrong costs more time than either option saves.
I'll walk through how I actually draw that line, with real examples from running this channel, not a theoretical framework.
The real question isn't va or automation, it's who owns judgment
Every task I hand off falls into one of two buckets: does it need judgment, or does it need consistency?
Judgment means the task has ambiguous inputs and the right answer depends on context that isn't written down anywhere. Consistency means the task has a clear input, a clear output, and the same steps every time, even if there are a lot of steps.
Automation is excellent at consistency and bad at judgment. It will do the exact same thing on the 500th run as the first, which is great until the 500th case is slightly different and nobody's watching. A VA is the opposite. She's slower per task but she notices when something looks off, and she can make a call I'd actually agree with if she asked me.
Once I started sorting tasks by that question instead of by "could AI theoretically do this," the split got a lot clearer.
What actually goes to automation
The things I've automated are all tasks where the steps don't change:
- Transcribing raw footage. I run this locally with a Whisper model rather than uploading footage to a cloud API, mostly because the files are large and I don't want unreleased footage sitting on someone else's servers before the video is public.
- Renaming and organizing exported clips and thumbnails into the folder structure my editing project expects.
- Pulling out timestamp candidates from the transcript for chapter markers, which I still review before they go live.
- Formatting video metadata (title casing, tag lists, description templates) into the shape my upload checklist wants.
- Checking new comments against a list of keywords so the ones asking about specific tools or errors get flagged for a reply, instead of me or my VA scrolling the whole thread.
None of this involves a decision that would embarrass me if it were wrong in a slightly different way each time. If the transcript automation mislabels a speaker, that's a fixable annoyance, not a judgment call. That's exactly the kind of task that should never touch a person's time.
What stays with my VA
Everything with ambiguity or a relationship attached to it stays human:
- Replying to comments that need actual judgment. A keyword filter can flag a comment; it can't decide whether the tone calls for a short answer or a longer one, or whether someone's frustrated and needs a different kind of response than someone who's just curious.
- Reviewing automated output before it goes anywhere public. My VA checks chapter markers, metadata, and thumbnails the automation produced. She's the last human eyes on anything before it's live.
- Research where source quality matters. If I need to know how a tool's licensing actually works before I say something about it on camera, I want a person cross-checking sources, not a script pulling the first search result.
- One-off requests. Sponsors, collaborators, viewers with a specific ask that doesn't fit a template. These are exactly the cases where a rigid pipeline breaks and a person adapting on the fly doesn't.
I want to be clear about something here: none of this is about a VA being "not worth automating yet." It's that these tasks are structurally different. An automation that tries to guess tone in a reply, or vet a source's credibility, isn't doing a smaller version of the same job. It's doing a different job badly.
Where the split breaks, and what happens then
Automations fail at their edges, always. My chapter marker script sometimes puts a marker in the middle of a sentence because it's working off transcript pauses, not actual topic changes. My comment filter catches keyword matches, but sarcasm and rhetorical questions slip through in both directions, flagged when they shouldn't be, ignored when they should've been flagged.
That's not a defect I'm trying to eliminate. It's the reason the review step with my VA exists at all. The automation's job is to get 90% of the repetitive work off the table so a person only has to touch the 10% that actually needs a brain. If I tried to build an automation that handled 100% of any of these tasks, I'd spend more time hardening edge cases than I saved, and I'd still get it wrong on cases nobody thought to test for.
Local versus cloud changes where the split falls, too
Part of why I run transcription locally instead of through a cloud API isn't about capability, it's about what I'm comfortable sending off my machine before it's public. Raw footage, unreleased scripts, and anything under a sponsor NDA stay local by default. That's a constraint that shapes the automation side of the split as much as judgment does. A task might be perfectly automatable in principle, but if doing it well requires a cloud model and the input is something I don't want leaving my machine, I either run a local model that's good enough for that narrow job, or I leave the task with my VA, who I do trust with sensitive material.
This is worth saying plainly: local models are not a drop-in replacement for the larger cloud models on every task. For narrow, well-defined jobs like transcription or formatting, a local model does fine. For open-ended research or drafting, I still lean on cloud tools when the data isn't sensitive. The split between VA and automation and the split between local and cloud are two separate decisions that happen to interact.
Agents don't remove the VA, they change what she reviews
Lately more of my automation runs as agents wired into real tools through MCP, meaning a model can call my file system, my transcript store, or my upload checklist directly instead of me gluing scripts together by hand. That's made the automation side handle more steps in a single run. It has not changed who signs off before anything goes live.
If anything, agentic pipelines need a clearer review step, not a lighter one. When an agent chains several tool calls together, a small misjudgment early on compounds by the time it reaches the last step. My VA's review of the final output matters more, not less, because she's now checking the result of a longer chain of automated decisions instead of one simple script.
The framework I actually use
When a new recurring task shows up, I ask two questions. First, are the steps and the correct output the same every time, or does the right answer depend on context that changes? If it's the former, I automate it, and I run it locally if the input is sensitive. If it's the latter, it goes to my VA.
Second, even for automated tasks, does the output touch anything public before a person checks it? If yes, my VA reviews it regardless of how good the automation has been so far.
That's the whole system. No part of it involves automation "taking over" what my VA does. It's a division of labor based on what each side is actually reliable at, and a review layer that catches the automation's edge cases before they become anyone's problem.
If you want to see how I actually build these pipelines, including the local transcription setup and how I wire agents into real tools, I cover it on the channel and on the site at [xavierfok.com](/).
Get new guides and videos first — join the Telegram channel.