AI and machine learning

Questions people ask before they commit.

How is this different from the automation we already run?

Automation carries out a rule you can write down in advance — when a form is submitted, send the message. This category is for the work you can only explain by example. Pulling the supplier, the amount and the due date out of an invoice is not a rule, because every supplier's invoice looks different. Almost every worthwhile system ends up as both, with a judgement step in the middle and ordinary automation moving things around it.

We want AI trained on our company data. Is that what we need?

Usually what that phrase means is answer questions using our data, and that is retrieval rather than training — faster to build, cheaper to run, updatable when a document changes, and able to cite where each answer came from. Actual training changes the way a model writes, not what it knows. The question that separates them is whether you want it to answer from your documents or to write in your voice.

Which of these services do we actually need?

It is decided by the shape of the answer you want. If the answer lives in written pages, that is a knowledge assistant. If it lives in database rows, that is a system that calls your systems. If it is a number about next month, that is forecasting rather than a language model. And if there is no single answer because you want a whole process carried out, that is an agent. Getting this wrong on day one is the most expensive mistake available here.

How do we know it is working, rather than just looking impressive?

Through an evaluation set — real questions from your business with the correct answers written by your own people, scored before launch and re-scored after every change. It turns quality into a number you can watch instead of an impression that shifts with whoever used it last. It is also the thing that tells you months later whether performance has drifted.

What will you not promise us?

That anything is a hundred percent accurate, that a system will never produce a wrong answer, that a chatbot replaces a support team, that the system improves by itself over time, or that anything here is legally compliant in your jurisdiction. All five are things buyers hear regularly and none of them are true. What we do commit to is showing you where an answer came from, and being able to explain any failure.

Where does our data go, and what does it cost to run?

Data handling is written down before the build — which provider, what their terms actually say about training, whether anything needs to stay in a particular jurisdiction, and what personal information is sent at all, since the answer should usually be as little as possible. Running cost is separate from build cost, driven almost entirely by volume, and we show you the arithmetic in the proposal rather than after the first month.