Nobody approved this.
It happened anyway.

The part I keep coming back to is that nobody signed anything. There was no meeting where a company weighed the risk, took a view and decided to start running its work through a language model. It began one deadline at a time, and by the time anyone thought to ask, it was already how the work got done.

You see it in small things. A note-taker that joined the call before anyone established whose account it belonged to. A supplier’s product that grew an assistant tab over a weekend. A draft contract pasted into a chat window near midnight because it was due in the morning. None of that is reckless. It is people trying to finish.

The standard response is a document. An acceptable-use policy goes round, people scroll to the end and tick the box, and a request form appears for anything new. Give it a month and the form has a queue, the queue has a workaround, and the workaround is somebody’s personal account on their own laptop.

I do not think that is a discipline problem, and I do not think a stricter policy fixes it. A policy asks people to recall a rule at exactly the moment they are least able to — late, tired, one job left before they can stop. And when it fails, it fails quietly. Nobody finds out until somebody asks a question that nobody can answer, and then reconstructing the answer costs more than the original job ever did.

Most tools in this area are sold to the people who write the policy. This one is built for the people who have to live with it. That is a smaller market. I think it is the right one.

So the question I started with was not how to stop people. It was narrower.

What if the tool wrote things down by itself?

Not a monitoring layer bolted on beside the work — people switch those off, or route around them, and honestly, they are right to. The same software you would want to use anyway, keeping its own record because that is simply how it works.

That is what Lumen is. It writes the day up from your meetings, your notes and the jobs you handed over, so Tuesday still makes sense on Friday. When it answers a question it names the file and the page it read, so you can check it rather than trust it. When you give an agent a set of files, that set is the boundary and it cannot read around the edge of it. And when the work should not leave the building, the model runs on the machine in front of you — through Ollama, LM Studio, GPUStack or Intel’s OpenVINO server, whichever you already have — with the network switched off.

Mesh is the same idea applied to a whole company rather than one desk: the same rules, the same record, and an account of the logins that belong to software rather than to people. It is in early access, and not yet something you can run yourself. I would rather write that here than let a demo imply otherwise.

It is also worth saying who this is not for. If nobody at your company is ever going to be asked what the AI touched, Niello is overhead you do not need. If what you want is one assistant inside one tool, there are cheaper and better things to buy. And if you need a certificate on the wall this quarter, we do not have one to give you.

Which is the first of the things we have decided not to do, each of which costs us something. We do not put a compliance badge on this site, so we lose the easy tick on procurement questionnaires — our evidence is mapped to the clauses those frameworks ask about, and mapped is not certified. Preview features are labelled preview in the product as well as on this site, even where a softer word would keep a conversation moving. And when something is half-built we say so, including in the meeting where it would be far easier not to.

We will get things wrong. When we do, the fix is the one the product argues for everywhere else: write it down, in public, where somebody can check it.

None of this is a claim about the future of work. It is a smaller bet — that the record should be a by-product of doing the job, rather than a project someone runs afterwards with a spreadsheet and a deadline. If that sounds obvious, I am glad. I would rather be obvious than clever.

If you want to judge it, judge the mechanism: the platform page has the whole thing, six layers and one request end to end. If you would rather use it than read about it, Lumen is free on one computer. And if you only do one of these, do the smallest one — tell me which part of this page you think is wrong.