AI integrations that do something. Not decoration.
UK-based AI integration developer. Cut support load, automate the admin nobody wants to do, and add AI features customers actually use. Chatbots, document analysis, and vision assistants connected properly and secured server-side — shipped into production, not demoed once and abandoned.
Three patterns that earn their keep.
Chatbots grounded in your content
Support assistants that answer from your documentation, pricing, and policies rather than inventing an answer. Retrieval means the model quotes what you actually published, and questions it cannot answer get handed to a human instead of guessed at.
Document analysis and admin automation
Uploads read, summarised, classified, and turned into structured data your systems can use. The work that quietly consumes a day a week — reading forms, extracting details, filing them somewhere — handled in seconds with a human check at the end.
Vision features inside your product
Image and photo understanding built into a real user flow. The AI Professor inside Instrumental Teachers Hub reads a photo of handwritten sheet music and gives feedback on it — a live subscription feature, not a proof of concept.
Typical stack:
Server-side, always.
A surprising number of AI features shipped in the last two years call the model directly from the browser with the API key sitting in the front-end code. Anyone who opens developer tools can take that key and spend your budget. It is the single most common mistake in this space and it is entirely avoidable.
Every integration I build runs through your own server. The key never reaches the browser. Requests are authenticated against your user accounts, rate-limited per user so one person cannot run up a bill, and logged so you can see what was asked and what it cost.
Cost control is part of the build rather than something you discover on the first invoice. Sensible model selection for the job, caching for repeated questions, and hard limits that fail politely instead of silently draining an account.
The same applies to data. What gets sent to a model provider is decided deliberately and documented, so you can answer the question honestly when a customer asks where their information goes. If the sensible answer is to keep something out of the prompt entirely, that is the design.
From idea to shipped feature.
Find the job
Start with the task that costs you time or loses you customers. If AI is not the right tool for it, I will say so before you spend anything.
Scope and quote
Fixed-price quote with the feature, the integration points, and the expected running cost written down. 50% deposit to start.
Build and tune
Built against your real content and tested on real questions, not a demo dataset. Prompts and retrieval tuned until the answers hold up.
Ship and watch
Live with logging and limits in place. You keep the accounts and the keys. 50% on delivery, retainer optional.
From £950. Fixed before we start.
AI features start at £950, quoted as a fixed price after a short brief, so the number you agree is the number you pay. Most projects land between £950 and £2,500. Where a build sits in that range depends on how much of your content needs indexing and how many systems the feature has to reach into. Provider usage is billed to your own account at cost and is never marked up.
AI questions
Can you add AI to a site I already have?
Usually yes. Most AI features sit alongside an existing site rather than replacing it, whether that site is WordPress, Next.js, or something else. I will look at what you have and tell you what is realistic before quoting.
Who pays for the API usage?
You do, directly to the provider, on accounts in your name. I do not resell usage or mark it up. Expected monthly cost is estimated during scoping and limits are built in so it cannot run away from you.
What stops the AI making things up?
Grounding it in your own content and constraining what it is allowed to answer. Retrieval pulls the relevant source material into every request, and anything outside that scope routes to a human rather than being improvised.
How do you communicate during a project?
Email and written messages throughout. Regular progress updates, you review and reply when convenient. No weekly calls unless you specifically want them.
AI inside your workflows.
Not every AI job needs to be a feature in your product. Sometimes the right home for it is a step inside a workflow that already runs — an OpenAI node that drafts the content or pulls structured fields out of an email before the next step files it. That is the automation side of the work: self-hosted n8n, running on your own infrastructure, with no per-task fees.
Got a job AI could actually take off you?
Send a short note about the task, the volume, and where it happens today. I reply within one working day. No calls required.
hello@cbwstudio.studio