About
We build the thing we run on.
SteelWorks Intelligence Systems designs and operates structural AI systems for businesses with operational depth and regulatory weight. Founded by Matthew Steger. Every claim on this site is measured by the system it describes.
Why it exists
The problem was never capability. It was ownership.
The company started the way most AI projects do: adopt a tool, get a result, adopt another. What became obvious quickly is that the failure mode was not the models. It was that no part of the work had an owner. Nobody could say which system had touched a record, what it had been allowed to touch, or who was accountable when it was wrong.
Businesses solved that problem a century ago and called the answer an org chart. Departments with decision rights. Escalation for the calls that need a person. A security function that reports separately from the people it reviews. None of that was invented for AI, and all of it applies.
So we stopped assembling tools and started building an organisation that happens to be made of agents — and then ran our own company on it, which is the only honest way to find out whether it works.
The platform
SteelWorks Intelligence is the system underneath
SteelWorks Intelligence is the platform we built and operate: the org runtime that holds the units and their decision rights, the recruiting pipeline that gives an agent a role card and an exam before it gets a post, the router that selects a model per task on measured score, the gates that refuse work below standard, and the security layer that screens what agents read before they act on it.
It runs this company. The scheduled operations behind this site — the audits, the content checks, the security scoring, the deploys — are SteelWorks Intelligence jobs. When we deploy a system for a client, it is a governed slice of the same platform on hardware they own, with their chart, their gates and their documentation. We evaluate outside frameworks in an isolated sandbox rather than in production, so a comparison never risks the operation.
How we work
Measured, gated, and willing to say no.
Every figure on this site is read from live system state when the site builds rather than typed in, and the claim audit re-reads the live pages after each deploy. The security scorecard is published in full, including the controls where we score zero. A model only wins a task after it out-scores the others enough times to mean something.
That machinery exists for one reason: it is very easy for an autonomous system to produce confident work that is wrong, and the only defence is to make the standard mechanical rather than a matter of good intentions.
What we will not do
Claim a result we did not measure.
We do not describe a vertical as delivered until a system in it has been assessed and handed over. We do not advertise a trained model we have not built. If a check could not be run, the report says so and why, and if we think you do not need what we sell, the reply says that instead.
The reason to hold that line is not modesty. It is that the entire product is a system you are being asked to let make decisions inside your business, and a company that shades its own numbers has already told you how that will go.
Talk to us
A person reads every one of these.
Leave your company and the operation that costs you the most time. You get a written reply with the shape of the system that would carry it — or an honest note that you do not need one.