Most “virtual assistant” projects end up building something that can answer questions about the work. The useful version produces the work.
That difference sounds small and it decides the whole outcome. An assistant that answers questions is a search box with better manners. Someone still has to go and do the thing. An assistant that produces work hands you a drafted document, an updated board, a filled-in tracker — and you check it instead of building it.
Getting to the second kind is not a prompting problem. It is a structure problem. You need a factory.
What a factory actually is
Four parts, and they only work together.
- A model does the reasoning.
- A database holds what is true — every decision, every task, every finding, each carrying a source, a date and an owner.
- Role-bound agents do the judging. One lane each, and a duty to stay in it.
- Connectors do the acting: writing the card, drafting the reply, updating the record.
Take any one away and the other three stop being useful. A model with no memory is an amnesiac consultant — brilliant every morning, starting over every morning. A database with no model is a filing cabinet. Agents with no connectors produce opinions nobody asked for. Connectors with no agents are a very fast way to make a mess.
Put the four together and you get something that can be handed a department’s actual workload.
Start with one department
The temptation is to build one assistant for the whole company. It does not work, and the reason is not technical.
Every department has its own vocabulary, its own idea of what “done” means, and its own set of things that are obvious to everyone in the room and written down nowhere. An assistant that serves everybody knows none of that, so it stays generic, so nobody trusts its output, so it goes unused. One department, on the other hand, is a small enough world to actually learn.
Pick the department that produces the same thing over and over. That repetition is not boring. It is the training ground.
Look for the work that runs on a cycle — the weekly report, the monthly review, the intake that arrives in the same shape every time. Those are the jobs where the assistant can produce a first draft that is genuinely close, because there are a hundred previous versions to learn the shape from.
What it does all day
A departmental assistant that is earning its place is doing unglamorous things continuously. It reads the channels the team already talks in. It keeps the board honest — the card that says “in progress” but has not moved in eleven days gets surfaced, not silently tolerated. It drafts the recurring document before anyone has to ask. It notices the same question being answered for the fourth time and writes it down properly.
None of that is impressive in a demo. All of it is the work that quietly eats a department’s week.
Speed is the method, not the goal
A factory processes the work and watches how the work is done, at the same time. The second half is the part people miss, and it is the reason to run things fast.
A job that takes three weeks teaches you almost nothing, because by the time it lands nobody can remember which step cost the time. The same job run in an afternoon shows you the shape of the process while it is still fresh — where it stalls, what gets retyped, which question always has to be asked twice. Run it seven times and the pattern stops being someone’s opinion.
This is also why the good findings never come from asking people what is broken. They come from watching the work. The things that cost the most time are usually invisible from the inside, because they are simply how it has always been done.
What we are not claiming
An assistant like this does not replace the department. It is not left to run on its own, and it does not decide anything that matters. A person approves the work before it goes anywhere, every time. The judgment calls — what to prioritize, what to say to a customer, when something is not ready — stay where they belong.
What changes is how much of the week goes into assembling things by hand.
We would rather be specific about what is finished than impressive about what is planned. If you have a department producing the same work over and over and you want to see whether this applies to it, that is a straightforward conversation to have.