I let my VA use my AI tools
# I let my VA use my AI tools
The most popular prediction about virtual assistants right now says AI will replace them. You have probably heard a version of it. Why pay a person to do research, draft an email, or clean up a document when a model does all of that in seconds for almost nothing? It is a tidy little story, and I think it is wrong.
Rather than wait around to find out, I ran the opposite experiment. I handed my assistant my AI tools. A real person on my team got access to the same models I use every day, plus actual training on how to use them, and then I paid close attention to what happened.
The short version: a capable person with a good set of AI tools is far stronger than either one alone. Each side quietly covers exactly what the other is bad at. So let me walk through what I gave her, how the combination holds up on real work, where it still fails, and the hard lines I keep around all of it.
Why I ran it backwards
The framing that made me want to try this: AI plus a capable human is a force multiplier. People constantly confuse that with replacement, and the two ideas point in opposite directions. Replacement removes the person and hopes the output survives. A multiplier keeps the person and makes everything they already do faster and lighter.
The multiplier wins for a simple reason. The human brings judgment, context about my business and customers, and accountability when something goes wrong, which no model carries. The AI brings one huge thing in return: it deletes the slow first draft grind. The blank page, the hour of typing before there is anything to react to, all of that goes away. Combine them and you get a person who spends her time deciding and checking instead of typing and waiting. That shift, from producing to deciding, was the whole prize I was chasing.
The three tools I handed over
I kept the list deliberately small.
First, a drafting assistant. A chat model she can hand a rough brief to and get a first version of an email, a document or an outline back in seconds.
Second, transcription. A tool that turns a long recording or a messy meeting into clean text she can work with.
Third, a research helper. A model she can ask to gather and summarise what is known about a topic before she digs in herself.
That is genuinely it. Nobody got a sprawling pile of twenty tools on day one with a hearty good luck. Three tools, each solving a clear problem she already had. A small kit a person actually learns will always beat a huge one they are quietly afraid to touch.
The guardrails went in on the same day, because this part matters more than the tools. She uses her own logins, never mine, so none of it touches accounts connected to money or anything sensitive. The tools can read and draft. They cannot send, publish, or pay for anything on their own. And we agreed one rule out loud before she touched a single model: the AI produces a draft, and a person decides. Nothing a model writes goes out into the world until a human has read it and actively chosen to send it. I said that rule to her face instead of burying it in a document nobody opens, because a rule you can repeat from memory is a rule that survives a busy afternoon.
First draft, then real judgment
Here is the combination working on an ordinary writing task. She needs a longer reply, a short guide, or a clear description of something. That used to start with a blank page and twenty minutes of staring. Now she hands the model a rough brief and gets a full first draft back in seconds.
The part that matters comes next. She edits with real judgment before anything moves. The sentences that sound generic and hollow get cut. The two small facts the model fumbled get fixed. The specific detail only a human who knows the situation would think to include gets added. The AI got her to a rough version fast, and she turned the rough version into something worth sending. Neither half does the job alone.
Long recordings become short summaries
The second example saves a silly amount of time. I record a lot: calls, quick voice notes, long rambling explanations of how I want some task done. A person used to sit through the whole thing at real speed just to extract the useful parts. Now the recording runs through transcription, and the model hands back a summary with the key points.
She reads that summary back against the actual transcript before trusting it, and this is the important half. AI summaries drop things. They will confidently omit the one detail that mattered, or soften a hard instruction into a vague suggestion. I have watched a summary turn a firm deadline into a soft maybe, which is exactly the small slip that breaks a project a week later. The model does the first pass and saves the hour of listening. The person does the check that makes the result trustworthy.
Research fast, then fact checked
Say we need to understand a new tool, a small market, or how some unfamiliar process works. The model pulls a starting picture together far faster than a person clicking through twenty open tabs, sketching the rough shape of a topic in a couple of minutes.
A model will also state something flatly wrong with total confidence, and invent a clean sounding detail that looks exactly as solid as a true one. So she treats the output as a rough map. The parts that actually matter get checked against a real source, and the rest gets organised into something usable. The AI made the gathering fast. The human made it correct. Reverse that order and you have built a machine for publishing confident nonsense at speed.
Routine questions from a template
The most everyday combination of all. The same routine questions arrive over and over, and each one used to need a fresh reply written from scratch. Now a model drafts an answer based on the pattern of how we have answered that exact question before, and she reviews it before anything goes out.
Her review is more than a rubber stamp. She checks the answer fits this specific person and situation, because the model has no idea that this same customer had a bad experience last week and a cheerful generic reply would land badly. One wrong tone on a routine reply can undo weeks of goodwill, and the model cannot feel that risk sitting under the words. The template gets ninety percent of the way there in seconds. She makes sure the final ten percent is right, and nothing reaches a real person until she has said yes.
The clean split
Step back from those four examples and a clear division shows up. The AI is good at speed, structure and first drafts. It is fast, it never tires, and it will produce a decent starting point at three in the morning without complaint. The human is essential for a different set of things entirely: judgment about what matters in this exact case, catching the confident mistake stated so smoothly you would believe it, context about my business and history, and anything touching a real relationship where tone and trust are the entire game.
Speed on one side. Judgment on the other. The trick was never choosing between them. The trick is knowing which is which.
The honest results
Two things clearly went up after living with this for a while. Output rose, and turnaround got faster. Work that used to eat a whole morning now finishes before lunch.
I want to be careful here, because this is where people start overselling. Quality still depends entirely on the person doing the checking, and that part changed by exactly zero. The AI made an already good assistant noticeably more productive. It would have done nothing to rescue a careless one. Give sloppy checkers these tools and they produce sloppy work faster, which is worse. The multiplier multiplies whatever was already there and holds no opinion about whether it was any good. A bad process running faster is still a bad process, just arriving at the wrong answer sooner.
She kept her job through every bit of this, and the setup never came close to threatening it. Every example above ends with a human decision the tools cannot make, and honestly one I would never want them to make. The win was taking the slow, dull first draft work off her plate so the sharp judgment work got more of her attention. Less typing, more deciding. Anyone selling you the version where the person quietly disappears is selling a nice story that would survive about one week of contact with real customers.
The lines I hold
Some things stay outside the reach of the tools and the assistant without me in the loop.
Final publishing to a real audience is first. Anything going out under my name to many people at once gets my own eyes on it. The cost of one confident mistake at that scale is simply too high, whatever my trust in the person.
Credentials and money come second. The tools cannot reach either, by design, and the person handles them only through the normal careful channels, never through a model.
Anything where a confident wrong answer would be expensive stays supervised too. The rule under all of it never changes: the more a mistake would cost, the more certain I am that a human checkpoint sits in the path. Cheap mistakes are allowed to move fast. Expensive ones wait for a person.
Who this is for
If you already work with an assistant or a small remote team, giving them good AI tools and some real training is one of the highest leverage moves available to you this year. You take someone already capable and make them meaningfully faster at the exact parts that dragged them down. Training is the quiet key. These tools do very little for a person who has never been shown where they lie and where they shine.
If you have no assistant yet, that hire is your first step, well before any of the AI part. The tools multiply a person. They cannot conjure one.
The place I found mine is OnlineJobs.ph. For more honest breakdowns of running a lean operation with AI and a small team, no hype and no doom, everything lives at [xavierfok.com](/).
Get new guides and videos first — join the Telegram channel.