AI and machine learning applied to actual business processes

Not a strategy deck. Working systems that read documents, classify records, forecast demand or answer questions over your own data — with accuracy measured against a real baseline before anything goes live.

What we build

Where it genuinely pays off.

AI is worth applying where a process is repetitive, high-volume and currently done by people reading things. That is a narrower set of problems than the market suggests — and these are the ones that work in practice.

/ 01

Document and data extraction

Invoices, delivery notes, contracts, forms. Pulling structured fields out of documents that arrive as PDFs and photographs with no consistent layout. Of the three, this one has the clearest payback: it replaces measurable hours of typing.

/ 02

Classification and routing

Sorting incoming email, tickets, transactions or records into categories, and flagging the ones that do not fit. Anomaly detection belongs here too — fraud patterns, sensor drift, stock discrepancies that a threshold rule would miss.

/ 03

Assistants over your own data

LLM-backed search and question answering across internal documentation, product data or case history, with citations back to the source. Built retrieval-first, so answers are grounded in your documents rather than produced from a model's general impression.

How we work

Measured against a baseline.

Every AI project we take on starts with the same question: what is the process doing today, and how often is it right. Without that number, there is no way to tell whether a model is an improvement, and a demo that looks impressive on ten hand-picked examples tells you nothing about the eleven thousand that follow.

How we approach it

  • A representative sample of your real data first — including the messy cases — that is where accuracy claims usually collapse
  • A measured baseline of current performance, so the improvement is a number rather than an impression
  • The simplest method that clears the bar: a well-built rule set or classical model often beats a large model on cost, latency and explainability
  • A human review path for low-confidence cases — the system should escalate what it is unsure about rather than guess confidently
  • Monitoring after launch, because input data drifts and accuracy measured once is a number you can no longer vouch for

Where your data goes

This is usually the first question from anyone with a compliance obligation, and it should be. Models can run on our infrastructure or on yours; where a third-party API is involved we tell you exactly which one, what is sent and what its retention terms are. Where data cannot leave your environment at all, we work with models that run locally — with an honest account of what that costs in capability.

When we say no

We will talk you out of an AI project if it does not stand up. If the volume is too low to justify the cost of building it, if the data needed does not exist in usable form, or if the real problem is a broken process rather than a missing model, the honest answer saves everyone a year. After twenty years, the difference between what demonstrates well and what survives contact with production is fairly obvious to us.

Questions

Things clients ask before starting.

How much data do we need?
It depends on the task. Document extraction with a modern model can work from a few dozen examples; a custom classifier usually wants hundreds to thousands of labelled records. We assess your actual data before proposing an approach, rather than quoting a number that sounds reassuring.
Will our data be used to train someone else's model?
Not without your explicit agreement. By default we choose providers and configurations whose terms exclude your data from model training, and we document which service processes what. If that is not acceptable, we build with models that run entirely on infrastructure you control.
What accuracy can you promise?
None before we have seen your data — and treat any number quoted before that as a guess. We establish a baseline, run a proof of concept on your real records and report the measured result. If it does not clear the bar that makes the project worthwhile, we say so at that point, not after the full build.
Can this integrate with our existing systems?
That is normally the larger half of the work. The model is rarely the hard part; the hard part is getting data in and results back out into your ERP, CRM or document system reliably. That integration work is what we have been doing since 2004, which is why the AI work tends to land.
Next step

Tell us which process eats the hours.

Describe the task, the volume and what the data looks like. We will come back with a proof-of-concept plan: what to test first, how much of your data to test it on, and what result would make it worth building.