Models and tools

We only build on well-known, hosted AI model providers; we don't use experimental or self-hosted open-weight models for client work. In practice, that means Claude (from Anthropic), OpenAI's GPT models, and Google's Gemini models. Which one we use for a given task depends on the task itself (cost, speed, and how well a model handles a particular kind of reasoning or content all vary) not on a fixed preference for one provider over another.

Alongside the model itself, most automations also need something to connect your existing tools together: your inbox, calendar, CRM, invoicing software, and so on. For that, we typically use established automation platforms (n8n, Make, or Zapier) rather than writing bespoke integration code for every connection, because these platforms are widely used, well maintained, and easier for another developer to pick up later if needed.

We choose per task rather than standardising on a single model for everything. A straightforward categorisation task (sorting enquiries into a handful of types) often works well on a smaller, cheaper model; a task that needs to read and reason over a longer document, or hold a more open-ended conversation, usually needs a more capable one. We test against real examples from your business before committing to a particular model for a given workflow, rather than assuming the most expensive option is automatically the best fit.

On data: using a hosted model provider means a request (a message, a document extract, whatever the task needs) is sent to that provider's service to be processed. We check each provider's current terms with you before anything goes live, rather than relying on an assumption about data retention or training use that may be out of date, because these terms do change over time and vary by provider and plan. We also minimise what's sent in the first place: only the information a given task actually needs, not a wholesale export of your systems.

This guide is deliberately about naming the categories and how we choose, rather than pinning ourselves to specific model versions, which change frequently. If you want the specifics for your own project (which model, which automation platform, and what that means for your data in practice), that's exactly what the discovery call covers.