Artificial intelligence
Most businesses do not need an AI strategy. They need one repetitive, expensive task doing automatically, proved on their own data inside two weeks. That is where we start — and if the honest answer is that ordinary code does it cheaper, we will tell you that instead.
The short version
It behaves like any other part of a system: it has inputs, a cost per call, a failure mode and an accuracy you can measure. Treated that way it is extremely useful. Treated as a strategy it produces a pilot that never ships.
The businesses getting real value out of it are not the ones with an AI department. They are the ones who found the single task that was eating a day a week — reading delivery notes, sorting the inbox, writing the same quote for the ninetieth time — and removed it. Then found the next one.
So the useful question is not what could we do with AI. It is which piece of work here is repetitive, high volume, and currently done by someone we pay properly. Answer that and the project designs itself.
We build the answer into the software you already have, or into software we build for you, with the review steps and the audit trail that make it something you can actually put in front of a customer.
Where we start
A two-week proof of value on your own data
Not a slide deck and not a generic demo. Your documents, your tickets, your images — run through a real implementation, scored against known correct answers.
AI solutions
Every one of these is something a person is currently doing by hand in somebody’s business, and every one of them has a version that is measurable enough to trust.
Invoices, delivery notes, contracts, application forms, survey returns, handwritten job sheets. A model reads them, pulls out the fields you actually care about, and hands them to your system as structured data. The work that used to be somebody typing.
document AI · OCR · data extraction
Retrieval-augmented generation: your policies, manuals, price lists and past jobs become something staff or customers can just ask. Answers cite the source document, so nobody has to take the machine on trust.
RAG · knowledge base · semantic search
Incoming email, tickets, enquiries, defect reports. Sorted by what they are and who should have them, with a confidence score — and anything the model is unsure about goes to a human instead of being guessed at.
triage · classification · routing
Quotes from a set of line items, a fortnight of site notes into a client update, a long thread into three bullets before you open it. Drafting is where AI earns its keep first, because a person still signs it off.
drafting · summarisation
Photographs from site, from the line, from a van. Damage graded, stock counted, a part identified from a picture, a delivery proved. Vision models turn a phone camera into an input your software can act on.
computer vision · image classification
Calls, site walkarounds, meetings. Transcribed accurately, then turned into the records your system needs — a job note, a follow-up task, a CRM entry — instead of sitting in an audio file nobody opens.
speech to text · call analysis
Demand, stock, staffing, churn, cash. Ordinary statistical models rather than anything fashionable, most of the time — the value is in having the numbers in front of the decision, not in the technique.
forecasting · prediction
An agent that checks a supplier portal, reconciles it against your orders, flags the differences and writes the chase emails for approval. Narrow, supervised, and inside a workflow you defined — not a robot let loose on the business.
AI agents · workflow automation
AI integration services
Calling a model is an afternoon. Making it safe, measurable, affordable and part of how the business actually works is the engagement — and it is the part that decides whether anyone is still using it in six months.
The most common request and the least glamorous answer: your CRM, your job system, your portal, your finance package. We add AI where it removes work, through the APIs and databases you already have. Nothing gets replaced to make room for it.
When the product is new, AI is designed in rather than bolted on — the data model, the permissions and the audit trail account for it from day one. That is the difference between a feature and a demo.
Hosted models from Anthropic, OpenAI or Google, open-weight models on your own infrastructure, or a small classical model that costs nothing to run. We pick on accuracy, cost per call, latency and where the data is allowed to go — and we will tell you when the boring option wins.
Most AI projects fail on data, not on models. Getting your content out of the places it is stuck, cleaning it, chunking it, indexing it and keeping it in step as it changes is the majority of the work on a real system, and we quote it honestly.
We build a test set from your real cases and measure the thing against it, so "is it any good?" has a number rather than an opinion. Then it runs on every change, because a prompt edit can quietly break what worked last month.
Confidence thresholds, human approval on anything that leaves the building, full logs of what was asked and what came back, and a way to correct it. Anything that touches money, contracts or people gets a person in the loop by design.
Deployed on the same monitored infrastructure as everything else we host, with token and inference spend metered and visible. You should never find out what your AI costs from an invoice.
A tool nobody trusts gets worked around. We show the people who will use it what it is good at, what it is bad at, and how to tell — which does more for adoption than any amount of accuracy.
How it runs
Deliberately front-loaded, because the cheapest moment to abandon an AI idea is the second week.
Half a day walking through what your people actually do. We are looking for the task that is repetitive, high-volume and currently done by someone expensive — not for somewhere to put AI. If we cannot find one, we will say so.
A short paid proof of value, usually one to two weeks, run against your real documents, tickets or images. You get a measured accuracy figure, a cost per item and a working demonstration — and a clear recommendation, including "do not build this".
The model is the small part. The rest is the queue, the review screen, the permissions, the audit trail and the handling of everything it is not confident about. This is where an AI feature becomes AI software.
Live with monitoring on accuracy and spend, a dashboard your team can see, and corrections fed back so it improves. Models change under you; the evaluation suite is what tells you before your customers do.
How it changes what we deliver
We use it on our own builds, openly, and it has changed the shape of a project rather than just its speed. Budget moved off typing and onto the parts that decide whether software is any good — and several things that used to be the first casualty of a deadline are now simply included.
Scaffolding, migrations, test fixtures and the fifteenth variant of a form used to eat the budget. Less of the estimate now goes on typing, so more of it goes on the parts that decide whether the software is any good — the flow, the edge cases, the day-two problems.
Generated cases around the awkward inputs, property-based tests on the money paths, and every bug that ever reached staging turned into a permanent test. The tests are still read and still reviewed; there are simply more of them than a fixed budget used to buy.
Every change is read by a person and passed through automated review for the classes of defect people skim past — an unhandled null, a missing index, a permission check that is checked in the wrong order. It catches different things than we do, which is the point.
Handover notes, API references, runbooks and changelogs generated from the code as it is written and then edited by the person who wrote it. Documentation used to be the first thing cut when a deadline moved. Now it is not expensive enough to cut.
You can hold a working version of the awkward screen in the first week, while it is still cheap to change your mind. Deciding on a real thing rather than a wireframe is the single largest saving on any project.
Contrast, focus order, labels, image weight, query counts and bundle size are audited on every build rather than in a panic before launch. The standard goes up because keeping it there stopped being manual.
What has not changed: every line is read, tested and owned by a person here. Generated code that nobody understands is a liability with a shorter fuse than hand-written code that nobody understands, and we treat it that way.
Straight answers
Worth saying out loud, because the alternative is billing you to find out.
If a decision can be written as a policy, write it as a policy. Ordinary code is cheaper, faster, exactly repeatable and free to run. Reaching for a model because it is the current word for clever is how you end up paying per call for an IF statement.
If you cannot describe a correct answer well enough to check one, we cannot evaluate the system, which means we cannot tell you whether it works. That is a conversation about the process, not a project.
Statutory returns, payroll, safety-critical calculations. AI can draft and check these; it should not be the last thing to touch them. We will build the review step and be quite firm about keeping it.
Sometimes the answer is an open-weight model on your own hardware, and sometimes the honest answer is that the accuracy you would get there is not worth the project. We would rather tell you that in week one.
Questions
Including the two everybody asks last: what it costs, and whether it replaces people.
Connecting a model to the systems and data you already run, so a step that a person does by hand happens automatically. In practice it is an API call to a model wrapped in a lot of unglamorous engineering: getting the data out, giving the model the right context, checking what comes back, storing it against the right record, and showing a human anything the system is not confident about. The model is usually a few days of the work. The integration is the rest.
Almost certainly not. Training a model from scratch costs more than most businesses will ever recover, and the hosted frontier models are better than anything you would train. What you may want is your own data attached to a general model through retrieval, or a small fine-tune for a narrow classification task. We will tell you which of the three you need before you spend anything.
On the commercial API tiers we use, your data is not used to train the providers’ models and is retained only briefly for abuse monitoring. We can also run open-weight models on infrastructure in the UK or EU so nothing leaves at all. Which route we take is a decision we make with you at the start, in writing, and it is usually driven by your contracts rather than by the technology.
Because we measure it. We build an evaluation set from your real cases with known correct answers, and report accuracy as a number before anything goes live. Then the suite runs on every change, so a prompt edit or a provider update cannot quietly degrade it. If the number is not good enough for the task, the answer is a human review step, not a hopeful launch.
A proof of value on your own data is typically £1,500–£4,000 and takes one to two weeks. A production AI feature inside existing software is usually £6,000–£20,000 depending on how much data plumbing it needs. Running costs are metered — most document and triage workloads land between a fraction of a penny and a few pence per item, which we measure during the proof so the figure is yours, not an industry average.
Usually, yes. If it has an API, a database we can reach, or an export we can schedule, there is a way in. We audit it first and tell you honestly whether the integration is straightforward, awkward but worthwhile, or a sign that the underlying system is the actual problem.
The projects that work remove the part of a job nobody wanted — the retyping, the sorting, the chasing — and give the same people more of the work that needs judgement. The projects that fail are usually the ones that started as a headcount exercise, because they get built without the people who know where the exceptions are, and the exceptions are the whole game.
Working proof on your data inside two weeks. That is deliberate: it is the fastest way to find out whether the idea holds up, and it is far cheaper to stop after a fortnight than after a quarter.
Yes, openly, and it is part of why our estimates look the way they do. It writes scaffolding, tests, migrations and documentation under review, which moves budget away from typing and towards design and edge cases. Every line is still read, tested and owned by a person here — the accountability does not move.
Describe it in a couple of lines. You will get a straight answer from a developer about whether AI removes it, roughly what a proof of value would cost, and what it would take to put into production.
