Write to us Write to us 
Where AI earns its place in an operating business — and where it does not. Part of our management consultancy.
What this activity does
Most AI work dies between the proof of concept and the Monday morning it is supposed to run. We advise operators on the unglamorous half — the evaluation harness, the golden dataset, the human-in-the-loop step, the real cost at production volume — the half that decides whether anyone is still using it in six months.
Proven internally first
Misclassified cloth is expensive twice — in duty paid and in time lost at the border. The classifier was built for our own import desk, run alongside the human process rather than instead of it, and held there until its disagreements with the broker were the interesting ones. It ships to clients only because it survived our own volume first.
How an engagement works
We establish the question the work has to answer, and write it down before anything is bought or built.
Delivery and measurement go live together, reviewed on a stated cadence with the model open to you.
Effort shifts from rented results to owned assets, so what we build outlasts the engagement.
Where it is deployed
Common questions
Almost always the latter. The work that decides whether an AI feature survives contact with an operating team is retrieval, evaluation and the handling of the cases the model gets wrong — not training a model from scratch. We build that layer.
An evaluation harness with a golden set built from real examples the operating team has already judged. It runs before deployment and on every change afterwards, so a regression is caught by the harness rather than by a customer.
It is designed for in advance. Every deployment has a defined path for low-confidence cases — usually routing to a person, always logged — because a workflow with no answer for the wrong case is a workflow that gets switched off in month two.
Yes, though we deploy internally first. Anything offered to a client has run against our own freight, textile or e-commerce volume, which is why the recommendations come with an opinion about what will not work.
Inference economics are part of the design, not a surprise afterwards. Model choice, caching and routing are set against a stated cost per transaction, because a workflow that is cheaper done by hand should be done by hand.
Begin
If AI is the wrong tool for it, that will be the reply.
