The pressure to do something with AI is real. It also pushes teams into pilots that impress for a quarter and then get abandoned. We work the other way: start with the problem, choose the smallest useful AI, test it, then harden it.
Honest advice, including “not yet”
Our first job is to tell you the truth, including when AI is not the answer. AI is worth doing when it has a real process, decent data and a clear owner. Where those are missing, we will point at what to fix first. Sometimes a rule, form or database wins. We cover that test in when not to use AI.
Most failed AI efforts begin with a tool looking for a use. We name the problem, the people affected and the result that would prove value, then decide whether AI belongs. That makes the model choice much easier, which we cover in choosing an AI model for your business. If you want a quick read first, the free AI readiness assessment scores data, security, process and people.
Grounded in your data, not the open internet
A general model is impressive and ungrounded: it knows the internet and nothing about your contracts, your policies or your history. Useful AI is the opposite. It answers from your data, shows its working, and is tested before anyone leans on it.
We connect language models, open or commercial, to your documents and systems of record using retrieval-augmented generation (RAG) and vector search. In plain terms, RAG gives the model an open-book exam instead of asking it to recall everything from memory: when a question comes in, the system finds the most relevant passages in your own material, hands those to the model, and asks it to answer using only what it found, with links back to the source. That keeps answers current as your documents change, makes them traceable, and means the model isn’t inventing facts to fill a gap.
What we build
- Knowledge assistants and custom ChatGPT-style data assistants that answer staff and customer questions from your own documents.
- Document intelligence that reads contracts, forms and reports and extracts the fields that matter.
- Copilots and agents embedded in the workflows your team already uses.
- Custom and fine-tuned models for the tasks where a general model keeps getting your domain wrong.
Our Minutes Radar product shows the pattern end to end: ingest messy source documents, summarise them, and make them searchable by topic, place and decision. Most of this value is locked in the unstructured data in your shared drive, the contracts, emails, PDFs and notes that no report can reach today.
When a general model isn’t enough
Off-the-shelf models get you a long way, then stop short. They don’t know your domain’s language and they miss the distinctions that matter to you. Most projects should still start with a general model and good retrieval; it’s faster and cheaper, and we’ll tell you when that’s the right answer. But there are clear cases for training your own: when accuracy on your specific task is the difference between useful and unusable, when your domain has edge cases a general model keeps getting wrong, or when a smaller specialised model would run more cheaply and privately than a giant general one.
When we do train, we define the task and the success metric up front, curate and label the data, then fine-tune and iterate against a held-out benchmark you’ve agreed to. Nothing goes to production on vibes, and the benchmark stays with you so the model can be re-tested whenever the work or the data changes.
Tested, grounded, and safe to rely on
AI can be wrong with enormous confidence. So we add evaluation, guardrails and logging before launch, ground every answer in a real source, and keep a human in the loop where the stakes warrant it. We build the evaluation set from real questions so “accurate enough” is measured, not assumed, and we monitor it once it’s live.
The same discipline applies to governance. Two failure modes dominate: no policy at all, with staff pasting sensitive data into whatever tool is open, or a rulebook so heavy nobody reads it. There’s a sensible middle, a one-page policy people actually follow, which we set out in the minimum viable AI policy.
Sequence beats scale
Adoption works when each step is small enough to finish, useful enough to notice, and solid enough to build on. We lay out that order of operations in a practical AI adoption roadmap for regional businesses, and you don’t need a data-science team to follow it. One automation that saves a team a few hours a week earns more trust than a platform that promises to transform everything next year.
None of this depends on a capital-city postcode, the case we make in why serious engineering doesn’t need one. Where your data can’t leave your control, we run the whole pipeline on private, on-prem AI instead, which is worth understanding properly: what “private AI” actually means. And the AI is usually only one part of the system; the workflow automation, integration and custom software around it are what let people rely on it.