Nobody needs “AI”. Businesses need a particular repetitive job doing without a person doing it. Name that job and there is a project here. If you cannot, we will tell you so rather than sell you the word.
More AI money is wasted at this stage than at any other, and it is wasted quietly — on something that works in a demonstration and never gets used.
“Answer the forty questions a week our staff get about delivery times.” That is a project. “Use AI in the business” is not one yet, and no amount of money turns it into one.
AI earns its keep on repetition. Something done fifty times a week is worth automating; something done twice a month is usually cheaper left alone, and we will say so.
The honest one. These systems are right most of the time, not all of the time. Where a mistake means a re-read, that is fine. Where a mistake means a wrong dose, a wrong payment or a legal filing, a person stays in the loop — or we do not build it.
That third question is the one nobody asks you, and it is the one that decides whether a project is a success or an expensive embarrassment. We would rather have the awkward conversation at the start. If the answer is that the job must be right every single time with nobody checking, then what you want is ordinary software with rules in it — which is cheaper, faster and does exactly as it is told.
The questions your staff answer over and over — opening hours, delivery times, whether a thing is in stock, how a return works. A chatbot is worth building when you can already predict most of what it will be asked.
The failure mode is a bot that answers everything and is sometimes wrong. A good one answers less and is trusted more, and hands the rest to a human without the customer having to fight it.
An agent does not just answer — it acts. It reads the email, finds the order, updates the record, sends the reply. That is genuinely useful, and it is also the point at which the mistakes stop being conversations and start being consequences.
We build agents with the brakes on: a written boundary of what it may touch, and a record of everything it did. An agent nobody can audit is not a tool, it is a liability with good manners.
RAG means giving an AI your own documents to answer from, so it quotes your material instead of guessing. Your catalogue, your policies, your past quotations, your manuals. It is the difference between a clever stranger and somebody who has read your files.
This is the honest centre of most AI projects. The value is not the model, it is your material — and the reason it works is that the answer arrives with the page it came from, so a person can check it in seconds.
Paperwork arrives, somebody types it into a system. Invoices, delivery notes, purchase orders, forms, CVs. Reading them and pulling out the fields is the least glamorous AI project available and frequently the one that pays for itself first. NLP — natural language processing — is the same idea pointed at loose text rather than forms: sorting it, classifying it, pulling the facts out of it.
Measure this one honestly: how many documents came through without a person touching them. If that number is high the project has paid for itself, and if it is low we would rather find out in week two than at handover.
Software watching a camera and reporting what it sees — counting things, spotting things, noticing when something is where it should not be. This is a different discipline from the language work above and it is judged on different ground: it depends far more on where the camera is than on how clever the model is.
Be wary of anyone who quotes this from a description alone. The site visit is the project — a badly placed camera cannot be rescued by better software, and we would rather look first than promise first.
An LLM is a large language model — the thing behind ChatGPT and its rivals. Integration means wiring one into your systems so it does a job in the place the work happens, rather than in a chat window somebody has to remember to open. Since these words get used at you constantly, here is what they mean in practice.
The useful question is never “which model”. It is what it can see, what it may do, and who checks it — and those three answers belong in writing before anything is built.
Left to answer from general knowledge, yes, sometimes — confidently and in a complete sentence, which is what makes it dangerous. That is why the work we do points the model at your own documents and has it quote them, so an answer arrives with the page it came from and a person can check it in seconds. It is also why we will not put AI in front of a job where being wrong once is unacceptable.
No, and almost nobody should. Training a model from nothing costs more than most businesses spend on software in a decade, and it would still know less than the models that already exist. The work is in connecting a proven model to your material, your systems and your rules — that is where the value is and where the risk is.
Somewhere specific, and you are entitled to know exactly where before you commit. Most AI work sends text to a model run by somebody else, so before anything is built we tell you which provider, what leaves your building, what stays inside it, and what that provider's terms say about your data. If that answer is not acceptable to you, say so early — some jobs can be done with a model running on your own hardware instead, and it is far cheaper to decide that at the start.
There is an ongoing cost, and this is the difference between AI and ordinary software that surprises people most. A model charges for every question it answers, so the bill follows use rather than sitting still. We tell you the shape of it before you build, and part of designing the thing properly is making sure it does not ask the model questions it did not need to ask.
You measure it against the job it was given, which is another reason we insist on naming that job first. If it answers customer questions, the number that matters is how many it handled without a person. If it reads invoices, it is how many came through without correction. Anything that cannot be counted like this was never a well-defined project.
Fifteen minutes on a call, no charge. Sometimes the answer is a small piece of ordinary software, and you will hear that from us before you spend anything.