AI models

Self-hosted AI models: when the prompt should never leave the building

Public chat tools are fine for a first look. They are a poor place for case data, contracts or anything with a name on it.

17 September 2026 · 6 min

The actual question

Most teams do not have an AI strategy. They have a handful of people pasting text into a browser and hoping the terms of service are kind. For a public blog post that is harmless. For a patient record, a bid, or a staff file, it is a transfer of personal data to a company you do not contract with.

Self-hosting is not a religion. It is a decision about where the tokens go. If the model runs in your tenancy, on hardware you can name, you can write an access log, a retention rule and an incident plan. If it runs in a consumer product, you have a screenshot and a prayer.

What we actually host

A useful self-hosted setup is smaller than the slide decks suggest. One model that does one job: classify documents, draft against a private corpus, extract fields from a form. Retrieval over your own files, not the open web. Identity in front of it. No training on prompts unless you chose that, in writing.

The work is the plumbing: GPU or reserved CPU that finance can see, a private network, evaluation so you know when the answers drift, and a human who owns the runbook. That is the same discipline we use on any production model in Leamington Spa.

When not to self-host

If the data is already public, or the question is generic, a contracted API with a data-processing agreement can be enough. Do not spend six weeks standing up a cluster to rewrite meeting notes. Do spend it before you let a frontier model see casework.

Related service: AI & machine learning

Tell us the problem. We’ll tell you if we can help.

A short conversation is usually enough to know whether this is a two-week discovery or a longer piece of work. No pitch deck.

Book a conversation