Three Ways AI Replaced My Virtual Assistant

I used to have a virtual assistant. I do not any more, and it was not a decision so much as an erosion — three tasks moved across, and at some point there was nothing left in the middle.

None of the three are dramatic. That turns out to be the point, so let me take them one at a time.

One: scheduling

The old arrangement had a problem I never really made peace with. My VA needed access to my whole calendar. They could see everything — every client, every internal conversation, every thing I would rather not have broadcast, all of it visible because one part of the job required visibility into some of it.

That is not a criticism of the person. It is a criticism of the shape of the access. There was no way to say "you can see the shape of my week but not the contents", so the only option was all of it.

The AI only sees what it schedules.

Now I screenshot the conversation off my phone — a text thread, an email, a message where somebody and I have loosely agreed on a coffee — and it pulls out the time and the place and books it. Nothing else in the calendar is in scope.

Minutes down to about thirty seconds, and accuracy from roughly 90% to 98%.

The time saving is nice. The access change is the part I would actually defend. If you have ever had to think carefully about who can see what in your business, you will recognise this: the win is not that a machine did it faster, it is that the job could finally be scoped to the information it needed. That principle has a name in security work, and it is worth applying to your own diary.

Two: event research

Same job, three failure modes, every time:

• wrong events — technically relevant, useless to me
• slow shortlists — by the time the list arrived, half of it had sold out or passed
• a misunderstanding of why I was going — the deepest one, and the hardest to fix by giving feedback

That third one is the interesting failure. If I am going to an event to meet two specific kinds of people, then "is this event about my industry" is the wrong filter entirely. But explaining the real goal properly takes a conversation, and that conversation has to happen again with every new brief, and it degrades every time somebody new picks up the task.

With AI, I built that context in once. It knows what I am trying to do, so the filter is the actual filter rather than a proxy for it.

An hour at about 80% accuracy became five minutes at about 98% — with the reasoning for each pick attached.

The reasoning is not a nice-to-have. It is what makes the output reviewable. I can look at why something made the list and disagree with it in ten seconds, which is a completely different activity from re-doing the research myself to find out whether I trust it.

Three: formatting documents

The one nobody thinks about, and the one that ate the most hours.

Everyone knows AI can write it. That is not where the time went. The time went into making it look like ours — the structure, the headings, the order things go in, the bits of a proposal that are always in the same place because that is how we do proposals.

And there is the prose problem. Getting a document formatted without obvious AI tells leaking through still wants a human eye. You know the ones. "It's not just a blank, it's a blank." The triple. The paragraph that restates the previous paragraph with more adjectives. Nobody has to be told what that sounds like; you can feel it a sentence in.

But here is what changed. Give the AI the actual context for how our proposals should look — not "make it professional", but the real structure, with examples — and it can check itself against that. The tells are a known list. A model can be told to go and find them, which is a much better use of it than asking it to avoid them while it writes.

Two hours became about twenty minutes.

The twenty minutes is still mine, and it should be. The human eye at the end has not gone anywhere. It just does not start from a blank page any more.

The pattern

None of these are dramatic. There is no transformation programme here, no platform, no pilot. They are small interactions that happen a lot.

That is exactly why they added up to a role. A task that takes ten minutes and happens four times a day is not a ten-minute task — it is most of a person, hidden in the gaps between other work. The reason it never gets automated is that it never looks big enough to be worth a project.

It is the same maths that made outsourcing stop paying: the offshore hourly rate is not falling, the cost of capable AI is, and at some point the lines cross. It is also the same shape as the three places AI actually took cost out of Biz365 — none of the savings came from the big obvious candidate.

One caveat, because I have argued the other side of this and I still believe it. Everything above is AI doing, not AI thinking. I decide which events matter and why. I hold the model of what a good proposal is. The AI executes against a judgement I have already made — which is precisely the split I wrote about when I went back to a pen and paper. More AI do, less AI think.

Look at your own week for the ten-minute jobs that happen constantly. Not the annoying big thing. The small thing you have stopped noticing.

If you want to talk about where that applies in your business, come and have a chat with us — or see what this looks like running in production. Or get this kind of thinking weekly.