Three lines of work
What each one is, who it is for, and what usually brings them.
They differ by what has to be true before the work can start, and by who inside the company owns the problem.
We build B2B systems from zero — for processes no off-the-shelf product fits, in environments where the data is dirty, the integrations are legacy, the setting is regulated, and mistakes cost money.
AI goes in where it earns its place, and the data work is part of the build. We build the system a company runs on, and we keep running it afterwards.
Systems tailor-built to last.
The system a company runs on, designed around a process that cannot be bent to fit software.
A business whose main process does not fit any product on the market, and which runs that process in spreadsheets beside an ERP built for a different one. Usually 50 to 500 people. Felt first by whoever owns the process.
The trigger: off-the-shelf did not fit, or the ERP went in and the work is still being done in Excel, or the integrator will not take it on.
Agents in operations.
An agent that reaches a verdict is a different engineering problem from an agent that offers suggestions. It has to be right, it has to say why, and it has to know when to stop and hand over to a person.
A business where people spend the day reading documents in order to decide something — approving, checking, comparing against a policy — and where being wrong costs more than a redo.
The trigger is one of three: an agent is running and nobody can tell where its answers come from; an agent is about to ship and nobody will sign off on it; an agent answers, and people redo the work by hand anyway.
The data layer underneath.
Matching, master data, document intake, the exception tail. Where an order stops when the data is not clean enough to act on.
A business that takes in a large volume of incoming documents in formats it does not control: orders, requests and specifications arriving as email, PDF and free text, against a catalogue or master file that grew for years.
The trigger: an open vacancy for order entry, a bad review of the vendor already in place, or a catalogue nobody can search.
The third one sits under the other two
A system built from zero usually contains agents, and both stand on a layer of matching, master data and document intake that had to be built first. Where that layer is already sound, the work starts higher up. Where it is not, starting higher up is what makes the project take twice as long as it was quoted for.
Airports
Airports are the first two lines applied to a terminal, sold to a commercial director who counts revenue per passenger. It is the one place where we are alone in the niche.
Is this not just low-code?
Most of what people call automation, low-code already does, and where that is the case we will say so. Two questions decide which you have: whether the rule can be drawn, and what a wrong answer costs.
Five things this work is actually made of
AODB and FIDS. Reading an airport's operational database and its flight displays, and turning a departure that slipped forty minutes into an hour someone can be sold something in.
A vessel's link is narrow, and it drops. Software written on the assumption that it will not is software that works in the office. Working over Starlink without replacing the hardware already on board is a design constraint, not a footnote.
M8×40 A2-70 carries eight parameters and states five. The pitch, the point the length is measured from and the whole block of head form and standard are supplied by a buyer who does not notice he is supplying them — and by nothing else between the inbox and the warehouse.
Two hundred times. What a document corpus has to be compressed by before an agent can hold enough of it to answer a question and say where the answer came from.
Six agents, on microservices, with boundaries of responsibility that still hold when one of the six is wrong. Orchestration is the easy part. Deciding what happens when a component is confidently incorrect is the work.
Forty-five minutes. Bring the process, the documents, or the agent that is getting things wrong — we will tell you which of the three you are actually in.