AI & intelligent systems
AI that does a specific job
Agents, assistants and workflows built on your own content and connected to your own systems. Not a chatbot bolted to a website — something that answers a real question or completes a real task.
We will tell you honestly whether AI is the right tool for it. Sometimes the answer is a database query.
What we build
Each of these is a project we can scope. Most start smaller than people expect.
An agent that answers your customers
It reads your product documentation, your policies and your past answers, and replies in your own words. When it does not know, it says so and hands the conversation to a person instead of guessing.
An assistant that reads documents
Invoices, delivery notes, contracts, forms and scanned paper go in; structured rows come out, into your system, with the uncertain ones flagged for a person rather than filed silently.
An internal assistant for your team
Ask it what the warranty policy says, or what was agreed with a supplier last year, and get the answer with the document it came from — so it can be checked rather than believed.
A sales assistant that does the first pass
It reads an incoming enquiry, works out what is being asked for, drafts a reply with the right prices and attachments, and puts it in front of your team to send.
A workflow with several steps
A request arrives, the system checks stock in one place and the customer's history in another, decides what to do, and either completes the task or asks a person to approve it first.
Search that understands the question
Your catalogue, your archive or your case history, searchable by what someone means rather than by the exact words they typed — and returning the source every time.
Voice, where it genuinely helps
Calls answered, transcribed and summarised into your system. Voice is harder to get right than text and is worth it where people cannot type — on the road, on site, with their hands full.
AI inside what you already run
Not another tab. A summary on the record your staff already open, a suggested reply in the inbox they already use, a draft in the form they already fill in.
What you should know before buying AI
Five things that are true of every AI project, including ours. Better said here than discovered later.
It can be confidently wrong
A language model produces plausible text, not verified fact. That is a property of the technology, not a bug we have not fixed yet. Anything that matters gets checked — by a rule, by a source citation, or by a person — and we design that in rather than bolt it on.
It depends on what it can reach
An assistant is only as good as the systems, permissions and data it is given. Half of any AI project is the integration and the access rules, not the model — and where your data is not reachable or not clean, that is the work.
The model is somebody else's
We build on models operated by third parties. They change them, they price them, and occasionally they retire them; their terms apply to the data your system sends. We name the providers in the proposal so you can read those terms before anything is built.
Usage costs money, and it varies
Providers bill per use, so a busy month costs more than a quiet one. We tell you where the exposure is and build limits where the design allows for them — an unbounded loop is the expensive failure in this field.
It is not professional advice
AI output must not be presented to anyone as legal, medical, financial or other regulated advice, and it should not decide anything about a person's employment, credit or health without human review. Where a project touches that ground we will say so at scoping.
What is underneath
For the reader who wants the technical answer.
- Large language models from established providers, chosen per project rather than fixed in advance
- Retrieval over your own content (RAG), with embeddings and a vector index, so answers cite a source
- Tool and function calling, so an agent can read and write in your systems under explicit permissions
- Structured output with schema validation, so a workflow gets rows rather than prose
- Evaluation sets and logging, so a change to a prompt or a model can be measured instead of guessed at
- Approval steps and rate limits on anything that spends money or contacts a customer
What a project actually depends on
Scope, cost and feasibility depend on which systems we can reach, what permissions you can grant, which provider the work runs on, how clean your data is, and what the consequence of a wrong answer would be. We work those out with you before anything is built, and we will say so if the honest answer is that it cannot be done well.