We pair with Claude for AI-assisted engineering, and build Claude-powered assistants grounded in your own data via RAG and MCP.
We use Claude where language, reasoning or document work is the product — assistants, summarisation, classification and content generation grounded in your own material. It is our default model for engineering assistance and for customer-facing assistants that must stay on-brand and refuse gracefully.
Why we build on Claude
Claude is our default large language model for both delivery work and the products we ship. The reasons are practical rather than tribal: strong instruction-following on long, messy real-world documents; reliable refusal behaviour when a request falls outside scope; and enterprise terms under which your prompts and content are not used to train public models. For clients in regulated sectors, that last point frequently decides the project.
Long context matters more than benchmark scores in the work we do. Real business inputs are a forty-page tender, an inconsistent product catalogue, a decade of support tickets. Models that hold a large amount of that material coherently produce answers that need less correction, and correction cost is what determines whether an AI feature survives its first month.
How we use it in production
In products, Claude sits behind retrieval rather than in front of it. We index your approved content, retrieve the relevant passages at query time, and require the model to answer from those with citations back to source. That architecture converts a plausible-sounding assistant into a checkable one, and means content updates take effect immediately because you edit documents rather than retrain anything.
Where an assistant needs to act rather than answer, we connect it to governed tools through Model Context Protocol — a CRM lookup, an inventory check, a booking system — each with explicit permissions and full logging. Scope, tone and refusal behaviour are agreed with your risk owners and implemented as enforced guardrails, then covered by automated evaluations so a model or prompt change cannot silently degrade quality.
Where we will not use it
We do not put Claude, or any model, in a position to make a final decision about someone's money, eligibility or account without human approval. We do not generate political messaging automatically. We do not present model output as clinical, legal or financial advice.
And when a deterministic rule would do the job more cheaply, faster and with a complete audit trail, we recommend the rule. Reaching for a language model where logic suffices is one of the more expensive mistakes we see in 2026, and it is usually an agency decision rather than a client one.