What we build
AI that has actually read your data.
Ask most AI tools about your business and they will give you a confident, articulate answer that is entirely made up, because they have never seen your data. Connect them properly and something quite different becomes possible: a plain-English question, answered from live company records, with the source traceable.
“Which customers ordered less this quarter than last, and what did we quote them?”
That question currently takes somebody most of a morning, three exports and a pivot table. It should take eleven seconds.
Why the chatbot on your website can't do this
A general AI model is trained on the public internet. It knows an enormous amount about the world and precisely nothing about your business. It has never seen your order book, your CRM, or your margins.
Ask it a question about them and it will answer anyway. That's what these models do. And the answer will be fluent, plausible, and invented.
To be useful, it has to be connected to the systems where your actual data lives. That's the work.
The bit that has to happen first
Here's what nobody selling AI wants to explain: you cannot point a language model at raw business data and get reliable answers.
Every system stores things differently. A deal's stage might be StageName in one platform, dealstage in the database and stage_id in the billing system, three names, three formats, for one idea. Dates differ. Statuses differ. The same customer exists three times with three spellings.
Feed that mess straight to an AI and the noise drowns the signal. The model has to guess which field means what, and it guesses confidently and sometimes wrongly.
So the real work, the boring, valuable, unglamorous work, is building the layer underneath: one consistent schema, aligned names, standardised formats, relationships made explicit. Get that right and the AI part is almost straightforward. Skip it and no amount of clever prompting will save you.
This is the same unification work we'd be doing anyway. AI is what it unlocks, not a separate project.
RAG, not fine-tuning
There are two ways to give a model your knowledge, and one of them is nearly always wrong for a business your size.
Fine-tuning retrains the model on your data. It's expensive, it's static, and it goes stale the moment somebody adds a customer. It also doesn't stop the model inventing numbers, which rather defeats the purpose.
Retrieval-augmented generation is an open-book exam. The question comes in, the system finds the genuinely relevant records right now, hands them to the model, and the model answers from what's in front of it. Live data. No retraining. And every answer traces back to a source you can go and check.
For a business that wants to trust the answer, that traceability is the whole game.
Which model goes in
We've built against Claude, ChatGPT and Gemini, and ElevenLabs for voice. Which model goes into your system is a decision we make on the job, based on what it needs to do, what it costs to run and where your data is allowed to go, not on which one is fashionable that month.
The Model Context Protocol
Until recently, connecting an AI tool to a database meant custom middleware, a bespoke bridge for every combination of tool and system. Ten tools and ten systems meant a hundred bridges.
MCP is an open standard, released by Anthropic and now governed by the Linux Foundation, that fixes this. It's the USB-C of AI integration: one standard connector, so any compatible AI client can talk to any system that exposes an MCP server. A hundred bridges becomes twenty.
An MCP server exposes three things:
- Resources, what the AI is allowed to read. Database rows, files, records.
- Tools, what it's allowed to do. Run a query. Send an email. Update a record.
- Prompts, templates for the specific jobs your business does repeatedly.
Crucially, access runs through your existing permissions. The AI can only see what the person asking is already allowed to see. Your sales manager doesn't get the payroll. Secured properly, with OAuth and role-based access control, this is a serious, auditable piece of infrastructure rather than a party trick.
What it actually gets you
- Reporting that takes minutes, not mornings. The weekly report that eats eight hours becomes a conversation.
- Questions nobody currently asks, because asking them costs too much. Those are usually the valuable ones.
- Answers you can check. Every response points at the record it came from.
- Non-technical people, self-serving. No SQL, no export, no waiting for the one person who understands the database.
The honest caveat
This only works on data that's been unified first. If your customer exists in three systems with three spellings, an AI will tell you three different things with total confidence, and you'll trust it, and that is worse than not having it at all.
Do not let anyone sell you the AI layer without the foundation. It's the expensive way to get confidently wrong answers faster.
Where this stands
MCP is genuinely new, genuinely open, and genuinely useful. It is also, right now, mostly being used by developers rather than businesses, which means there is a window in which a small company can have something its competitors don't understand yet.
We build these. We also build with them daily, which matters more than it sounds, because we know exactly how they fail. Controlled trials show AI coding tools make even professional developers 19% slower on complex work while feeling faster, and add 23.7% more security vulnerabilities. The tool will hand you something that looks right and quietly isn't. Knowing which answers to keep is the entire job, and it's not one you can subscribe to.
Worth a conversation?
Forty-five minutes, free. Tell us what question you can't currently answer about your own business, and we'll tell you honestly what it would take.
Where to start
A free call to work out what's possible
Forty-five minutes, no charge, no deck. Tell us what you are trying to achieve and where technology feels like it is holding you back. We will tell you honestly what is possible, what it would take, and roughly what it is worth. If there is nothing worth building, we will say so.
The scoping call is a short version of Analyse. If you decide to go ahead, we harden it up into the full thing, a detailed project summary rather than a first impression. If you don't, you still leave knowing what's possible and roughly what it's worth.