First we lay the foundations, then we build the agents.
For the company that handed everyone an AI account, ran a year of experiments, and cannot see an improvement on the bottom line. Faster individuals do not automatically make a more productive company.
Faster people, same company.
A company built now, or redesigned on purpose, runs orders of magnitude faster than one that bolted a model onto the process it already had. That is the gap, and it widens every quarter you leave it.
Architects had steel for years before anyone saw that it meant the skyscraper. Industrialists had electricity before anyone saw that it meant a new factory floor rather than a faster one. The same lag is running now.
The impact on the bottom line follows once a company is reorganised around what the model can already do.
How we think an AI-native company is built
They moved the motor and left the floor plan alone.
A context layer, a harness, and the workflows that run on them.
What we build depends on the problem. Sometimes it is the context layer, sometimes a custom harness, most often both.
Stop buying copilots, chatbots and transformation decks. Build a durable operating system around the work you already do.
The model is a commodity, and it gets cheaper and better without you. What lasts is the context you feed it and the harness that turns tokens into bounded, auditable production work.
One workflow, and it has to matter
We start on a single high-stakes workflow. Invoicing, claims handling, foundational processes that were never written down. High stakes is the point. People correct a system that matters, and ignore one that does not.
The context layer
Before a model sees any of your data, we write down the unwritten rules of that workflow. Who may approve a discount above the threshold is the kind of rule that lives in one person's head and nowhere else.
It lives as versioned files in a repository, next to the schema definitions and the known exceptions. Agents draft the first pass. A senior engineer corrects it, sitting alongside your most trusted colleagues. That is where the value comes from.
This is a different object from ordinary documentation. Somebody draws a diagram once, but most of the time nobody ever opens it afterwards. These files get read every day, because an agent reads them every day, and a document that is read is a document that gets corrected.
The harness
On its own, a model is a brain without hands. The harness is the software around it. It hands the model tools and memory, decides what it may see and do, and records what it did.
- Deterministic work stays in code. Arithmetic, policy lookups, validation. A model is not asked to do what a script does better.
- Evaluation sets and quality gates that fail closed, so a bad answer stops rather than ships.
- A sandboxed runtime, inside your own cloud.
- A model gateway, so changing vendor is a configuration change rather than a rebuild.
- Observability, access control, and cost traces a CFO can read without a translator.
The harness is the part that compounds. Once it stands, an engineer who joins next year joins a system that already absorbs AI, rather than a team still arguing about whether to.
What you keep
We send everything to a repository you own. The specs, the commands, the evaluation sets, the knowledge graph. And we measure the work by what it returns per token spent, rather than by how many people logged into a tool.
Things we build in the same way
-
A daily inventory-financing report for a trading company. The numbers come out of the ERP each morning, get checked against the rule, and go out.
-
A market engine for a sales team. It watches every company they could sell to, judges fit, and proposes the next touch. A person approves each one.
-
A knowledge base for an agency that runs work for competing clients. Each client's material is walled off by the structure, not by a promise about behaviour.
-
The company brain at Voxdale. Thousands of Confluence pages retired in eight days, and colleagues now build their own small tools on top of it.
Read the Voxdale story
I love that Machinegeist helped us push hard and take risks, while bringing structure and keeping control over data, privacy and quality.
People, process, tech. In that order.
People
We find the people who are good at this already, in leadership and in each department. They become the hub, and the rest of the company learns from them rather than from a training video. Your own staff end up operating the system, so there is no AI department to hire.
Process
A company does not get more efficient by making each person more productive. We redraw the handoffs first. Then we sort every step into one of three: plain software, an agent, or a person who decides with the evidence already prepared.
Tech
The brain, the harness the agents run in, and where the model runs. On its own, a model is a brain without hands. The harness is the software around it. It hands the model tools and memory, decides what it may see and do, and records afterwards what it did. We stay model-agnostic by design, so a provider can be swapped without a rebuild.
We sell AI, so take this from us. Most problems are waiting for better people or better software. In the systems we build the model is a component, rarely the biggest part. Hand a model a task a script would have done and you pay three times over. In tokens, in maintainability, and in predictability. The rest of the system we keep deterministic on purpose.
Three phases. Every one adds value on its own.
-
Audit and prioritize
under 2 weeksWe learn how you decide and how you work. We find the processes that constrain your company's output the most and suggest how to fix them. Followed by a go or no-go.
-
Build and deploy
days per workflowWe encode the operating model first. Once the brain stands, each workflow is built and shipped in days, on data you already have and with your team in the room. The second one is faster than the first.
-
Train and maintain
Your team learns to run and extend the system. Run it without us, or keep us on to keep it sharp. No lock-in either way.
Everything we build stays with you.
Map the bottleneck in a week. Deploy the first workflow in days.