The problem
69% of the S&P 500 have deployed AI. 2% track a metric over time.
These are the largest companies in the world, and they have AI teams. A firm of 500 people has none.
a16z, State of Markets II, 30 September 2026
Examples
Billing and credit control. Invoices go out the week a job finishes, and a reminder follows the day one goes overdue.
What we build
A model is not a system. In a system we run today, the model is about a tenth of the parts. We build the rest.
Company Harness
An AI operating system, built for one company.
The work
Every core process, redesigned to run on AI.
The agents
Built for the job, and run on your company's own live context.
The controls
Where your people run, review and steer all of it.

Who builds it
Built inside your company, by the people who keep it running.
Forward-deployed engineers.
Our engineers build it inside your company, beside the people who do the work today.
Operators beside your team.
They watch the exceptions, find the cases only your company has, and build each one in.
Deep in your stack.
It works inside the systems you already run, from the inbox to the bank. Nothing is replaced, and nothing has to be built on their side.
Yours
Every company can own its intelligence.
Owned outright.
The code, the context and the tests are yours, in your own cloud. If you leave, you keep everything.
Any model.
We test the models on your own work and run each process on the cheapest one that passes. When a better one arrives, it swaps in.
Corrections stay.
Every correction your team makes becomes a test, so the same mistake is not made twice.

Private
Your data stays yours.
Its own logins.
It signs in as itself, with the narrowest access that does the job.
Kept in the EU.
Your data does not leave it. If you need it kept in the UK, in your own cloud region or with a particular provider, we set it up that way.
Not used for training.
The model providers we use do not train on your data.
Open to audit.
Audit and access rights for you and your regulator, in a written contract.
Our work
Better than the most expensive AI, at a twentieth of the cost.
Joy is a German gift platform. Every gift redeemed becomes an order on one of thousands of shops' own websites, each built differently. A person used to place each one. Our CTO built the system that places them now.
Joy is orders. Yours may be invoices, claims or renewals.
Orders nobody had to touch
96.5%
Joy · real orders, payment included67.1%The five most expensive AI models, in Visa's testWhat the AI costs, per order
$0.96
A twentiethJoy · a low-cost model inside the system$19Our estimate for the same orders on a top AI model, at its published prices
- Over 16,000orders bought and confirmed since it started
- Nearly 2,000shops so far, with nothing built on their side
- 3 in 4,000purchases wrong. In Visa's test, nearly 1 in 4.
Every wrong purchase was flagged at once by an AI auditor and passed to a person to put right. A learning agent then worked out what went wrong and, with the team, wrote the fix into the system, so the same mistake does not happen again.
Visa tested AI models alone, in test shops, in November 2026, and stopped before checkout. Ours are real orders, payment included, in the 30 days to 11 November 2026.
How it starts
It starts with one process. Three weeks to the first job.
- Days 1 to 3
We come in.
Our engineers sit with the people who do the work and map one process, rare cases included.
- Weeks 1 to 3
We build it.
Inside your own tools, tested on a copy of your work. Then it goes live, with your team on the exceptions.
- After that
You own it.
Your team runs it. We keep it running, and start on the next.
For firms of 150 to 1,500 people · UK and Ireland · Insurance and recruitment first
Questions
Your people already have AI agents. Why is the work still on their desks?
An agent on a seat helps one person, under their login, when they ask. A process has to run for the whole company, with nobody watching and someone accountable. That takes a system, and these are the questions it has to answer.
Yes. Work comes in from your own systems, not from someone typing it in. Dozens of jobs run at once, the most urgent first, and no two ever touch the same account or record at the same moment. An agent on a seat waits for its person, and knows nothing of the other seats.
You decide which steps need a named person, and above what amount. Those limits are held in code and at the bank, where the model cannot change them. It signs in as itself, never as one of your staff, so the person who prepares and the person who approves stay separate.
No. Every job has one record. Before anything is tried again, the system checks what actually happened. A bill is not paid twice, a journal is not posted twice, and nobody is written to twice.
On a dashboard built for your company, not in a chat window. Each person sees and controls what their role needs: your team their exceptions, a manager every job, live or replayed step by step. One switch pauses everything. Another pauses a single client, supplier or campaign. The record is there for your auditors.
It stops with a named reason, and a rule decides what follows: try again, cancel, or alert the person responsible, once. If a model or a connection fails, it switches to another and carries on. In the morning you can see what stopped, why, and what had already gone out.
Every job carries its own cost. Each step runs on the cheapest model that passes your tests, and a new model goes live only after it passes them too. The code, the tests and everything it has learned from your team's corrections are yours.
Talk to us
