AI that does a specific job, properly
Most businesses do not need "an AI strategy". They need one repetitive task handled reliably so that a person can stop doing it.
We build AI into systems where it removes real work: answering the same customer questions, reading documents, drafting quotes, sorting enquiries, summarising long threads. Narrow jobs, done consistently.
We are deliberately cautious about where AI is allowed to act on its own. Anything touching money, contracts or customer commitments gets a human approval step. That is a design decision, not a limitation.
Where it genuinely pays
- Support assistants that answer from your actual documents, not guesswork
- Enquiry sorting and routing, so the urgent ones surface first
- Reading invoices, forms and PDFs into structured data
- Drafting quotes and replies for a person to check and send
- Summarising long email threads or call notes
Where we will tell you not to bother
If a task happens rarely, changes constantly, or has to be right every single time with no one checking, AI is usually the wrong tool and we will say so. Automating something that only takes ten minutes a month is a cost, not a saving.
Running costs
AI features have an ongoing usage cost from the provider on top of the build. We put spending limits in place from day one so a mistake or a burst of traffic cannot produce a surprise bill, and we show you the actual numbers before launch.
How we stop it inventing things
A language model will produce a confident answer whether or not it knows one. That is the central risk in this whole category, and it is an engineering problem rather than something you solve by writing a better instruction.
The main defence is grounding: the assistant answers only from documents you have given it, and every answer can be traced to the passage it came from. Where the documents do not cover a question, it says so and offers a person instead of improvising. We narrow what it is allowed to discuss, so an assistant built to answer questions about your services cannot be talked into discussing anything else.
Before launch we test it against a written list of real questions, including the awkward ones and the ones designed to trip it up, and you see those results. After launch the conversations are reviewable, so a wrong answer is something you can find and fix rather than something you hear about from a customer.
A realistic picture of the saving
The businesses that get value from this share a pattern: a specific, high-volume, repetitive task where being ninety per cent automated with a human checking the rest is a genuine win.
A support inbox where the same twenty questions arrive constantly is a good candidate. A stack of supplier invoices that somebody re-types into accounting software is a good candidate. Drafting the first version of a quote from an enquiry is a good candidate. In each case the work does not vanish, it shrinks, a person moves from producing to checking.
What we will not tell you is that it replaces a member of staff. In our experience it removes the part of a job everybody dislikes and gives that time back to the work that actually needs a person.
Your obligations, plainly
Using AI with customer data does not exempt you from UK GDPR, and a few things follow from that. People should be told when they are talking to an automated assistant rather than a person. Personal data sent to a provider needs a lawful basis and should be minimised to what the task requires. You should be able to say where that data goes and how long it is kept. We configure the setup so these are answerable, and we will put the specifics of your configuration in writing, but the responsibility sits with you as the data controller, and anyone telling you otherwise is selling.
Common questions
Will it make things up?
That is the main risk with these tools, and it is why we ground answers in your own documents and restrict what the assistant is allowed to talk about. Where being wrong would be costly, we route to a human instead of guessing.
Does our data get used to train someone else's model?
Not with the setups we use. We configure providers so your content is not retained for training, and we will confirm that in writing for your specific setup.
Can it be switched off?
Yes. Every AI feature we build has an off switch that leaves the rest of the system working normally.
Which AI models do you use?
Whichever fits the job, and we are not tied to one supplier. Some tasks need a large frontier model; plenty are handled by a smaller, cheaper one at a fraction of the running cost, and a few are better solved with ordinary code and no AI at all. We build so the model can be swapped later, because this field changes quickly and today's best choice will not be next year's.
How much does it cost to run each month?
It depends on volume, and we will estimate it from your actual numbers before you commit. A support assistant handling a few hundred conversations a month is typically a small fixed cost rather than a large one. The important part is the spending cap we set, which means the worst case is a feature that stops rather than a bill that does not.
Will it replace my staff?
Not in our experience, and we would rather not sell it that way. It removes repetitive work from people who have better things to do. If your goal is specifically a headcount reduction, be sceptical of anybody who promises it, the checking and exception-handling still need someone competent.
What do you need from us to build it?
Whatever documents currently hold the answers: service pages, price lists, policies, past email replies, an FAQ. Messy is fine; we structure it. If that knowledge only exists in somebody's head, the first useful step is getting it written down, which is worth doing regardless of whether you build anything.
Related
Tell us what you need
Describe the project in a couple of sentences and we will come back with a fixed price and a realistic timescale. No obligation, and no sales sequence.
Start a build