Working together
We start with your problem, not our product.
We take on a small number of engagements each year. The first week contains no solution and no deck — just a working session to find where your work actually piles up. Everything after that is built for your organisation specifically, deployed narrow, and measured against what your people approve.
How we think about it
Nothing here is a package.
Most AI consulting starts with a solution and goes looking for somewhere to put it. We think that is backwards, and it is why so many pilots end with a working demo and an unchanged process. The interesting part of your company is the part that does not generalise: the exception your team makes every March, the approval that always stalls in the same place, the report three people rebuild by hand because the system cannot express it.
So we begin by understanding that, in a room, with your people. If what you need turns out to be smaller than you thought, we will say so. If it is not something we are the right team for, we will say that too, and point you somewhere better.
Where we tend to be useful
Four problems we recognise quickly.
These are not service lines. They are the shapes of problem we have spent the most time inside, which mostly means we waste less of your time working out what is going on.
Your company's context is trapped in people and inboxes.
The decisions, precedents, and unwritten rules that make your organisation work are scattered across mail, documents, spreadsheets, and the heads of your most senior people. We build the model that holds it, connected to the systems you already run, so nothing has to be replaced or migrated.
You need agents in production, not in a demo.
Getting an agent to work once is a weekend. Keeping it working when the upstream data shifts, the model is deprecated, and a regulator asks what happened on the 14th is the actual job. We build for that second thing, with the evals, guardrails, and audit trail it requires.
Your hardest problem is shaped like a network.
Customers and products, supply chains, internal knowledge graphs. We have spent years on problems of this shape, and we would rather tell you it is a graph problem in week one than after you have paid for six months of something else.
You want the runtime run for you.
Some teams want the agents without owning the monitoring, routing, and model layer underneath them. We host on Platos and operate it, and you can take the whole thing in-house whenever you want it, because it is open source and you already have the code.
How an engagement runs
Understand, deploy narrow, measure, expand.
Twelve weeks from first session to a decision backed by numbers. Your team pairs with ours the whole way, so the knowledge stays in your building instead of leaving with us.
Understand
A working session with your team, with nothing to sell in it. We map where work actually piles up, which approvals become bottlenecks, and which of your processes depend on knowledge that only lives in someone's head. Most of what we need is already in the room.
Deploy narrow
One team, one workflow, on your stack and inside your approval rules. Narrow is deliberate: it is the only honest way to find out whether the thing works before it is load-bearing anywhere that matters.
Measure
Approved-work volume, time returned to your people, error rates. You watch the same dashboard we do, which means we cannot tell you a story about the numbers that the numbers do not support.
Expand
We widen to the bottleneck we found together, and only once the measurements make the case for it. If they do not make the case, we say so. That is a real outcome of this process, not a failure of it.
How we operate
Governance is not a feature we added later.
We build for organisations where an unaudited action is an unacceptable action. That constraint shapes the architecture rather than sitting on top of it.
Maker-checker, for AI.
Your organisation already runs on approvals. So does everything we build. Every AI action is a proposal until a person approves it, with multi-level sign-off where you need it and a full audit trail underneath — every action logged, timestamped, and replayable.
The model is yours, and it is exportable.
The context we build with you does not become our asset or our lock-in. It is customer-owned and portable, encrypted at rest and in transit, with single-tenant deployment available. We will have the conversation your risk team needs to have.
You pay for approved work.
We meter the work your people actually approve, and that is what gets billed. Not seats that sit unused. A cost per unit of work is a number your organisation already knows how to evaluate. If the work is not approved, it is not billed.
Where to start
Forty-five minutes. One workflow.
Bring the workflow that hurts most. We will map it in the session and show you honestly what a governed AI worker would do with it, including the parts it would not help with. No deck, and nothing to sign at the end of it.