I founded NexSky on an AI operating layer. Today that means a boutique family office with no outsourced back office, where software agents do the work and people approve it. It is early days.
The aim is not to find out how few people a firm like this needs. It is to have the machinery take the administration, so the human hours go to judgment and to the client. That is the whole point, and it is worth saying before anything else, because "runs on AI" invites the opposite assumption.
I am not an engineer. I am the person orchestrating it, so that all of it points the same way. That is where this series is written from, and I think it is the useful angle, because most of what I have learned so far is not technical.
Here is the first thing.
Automating the work did not remove the cost of coordinating it. It moved it. And it changed how things break.
In most firms that cost is easy to see. Meetings, management layers, the hours spent making sure two people are not solving the same problem in two different ways. Building this way, I never took those hours on in the first place. What I took on instead is the work of keeping the systems straight. Last month that bill arrived four times.
Four separate incidents, in four unrelated parts of the office, with one cause between them. Every time, two different things had been allowed to write to the same place.
If you have ever had two people edit the same document at the same moment, and watched one person's changes vanish when the other hit save, you already understand the problem. That is all it is. The rest is scale.
The running log of what the office does was being written to by two separate routines. Both worked exactly as intended. Entries went missing anyway, because when two things write at the same instant, the file keeps whichever finished last, and nothing announces the loss.
One piece of our software existed in three copies instead of one. An update went out from an old copy, and took a live service down.
One shared file, two routines allowed to update it. Each was right on its own terms, and each quietly undid the other, so a correction made one week could be gone the next.
A routine that could create files and rename them but not delete them. Every time it finished successfully it left a marker behind, and that marker blocked the next run. It reported success and stopped the next one from working.
None of that was the AI going wrong. All of it was ordinary coordination going wrong, somewhere I could not see it. That was the first surprise.
The second one is how you find out.
When people fail to coordinate, it is loud. Someone misses a deadline. Two people do the same work twice. You find out because humans object.
When systems fail to coordinate, it is silent. It looks exactly like it worked. Everything reports success. The record says complete. The entries are gone and nobody says anything, because nothing in there is annoyed about it.
So you swap the friction of people coordinating for the job of proving the machines did. Nobody puts that on the slide with the productivity numbers.
Two rules came out of it.
One resource, one owner. Everything that gets written to has exactly one thing allowed to write to it. A second writer with different habits is not a backup, it is an incident waiting for a busy week. It took four goes to write that down, which is three more than it should have.
Check, do not assume. Nothing is done because something reported it done. A confirmation is a claim, and it is not true until someone has gone back and looked. So when one part of the operation hands work to another, what was actually checked is written down separately from what is merely assumed. That has caught two near misses so far. Both times, work arrived described as finished when it had only been planned, and the only thing that found it in time was going to look.
Is it better? Yes. Not for the reason I would have guessed.
The speed is real, and it is what a client notices first. But the part I would not give back is the record. Everything done on a family's behalf is written down, attributed and timestamped, including the small things nobody would have thought to minute. A back office this size run by people cannot do that. Not because people are sloppy. Because writing down what you did is the first thing to go in a busy week, and it leaves when they do.
That record does something I did not plan for. In most small firms the way things are done lives in somebody's head. Everyone knows it and nobody likes to say it, because it means the firm is one illness or one resignation away from a bad month. Here the way things are done is written down, and it runs. The office does not depend on anyone remembering how the office works.
The other thing I did not see coming is that it only moves in one direction. Every rule here exists because something went wrong first, and once a rule is written into the system the lesson stays learned. It never has to be learned twice. In a team, the same lesson usually gets learned once per person.
One rule has not changed and will not. Nothing leaves this firm without a person approving it. Not a payment, not a document, not an email to a counterparty, not a commitment of any kind.
This piece included. It was drafted on the same operating layer, and it did not go anywhere until I had read it, argued with it, and changed a good deal of it. It would be a strange article to publish any other way.
The technology prepares, a human releases. That is true today, and it stays true as the firm grows and there are more of us. It is not nervousness about the tools. It is the reason the rest of it can run fast.
Which brings me back to the point of doing any of this. Not a firm with nobody in it. A firm where the reconciliation, the chasing and the filing are done by machinery that never gets tired or bored, so that the people are spending their time on the client's actual problem. As we grow we will add people, and their hours will go to clients rather than to paperwork. The test is whether a family is better served, not whether the firm is more cheaply staffed.
If you are weighing this up for your own firm, the technology is the easy part and it is not where the cost is. The cost is the discipline that keeps the thing honest when nobody is watching it. There is no shortcut around that. I looked.
This is the first of a series. What works, what does not, and where it ends up better. The failures will be in here too. Those are the ones worth reading.
Christophe Schaillee, Founder and Managing Partner, NexSky