On-Premise RAG for UK Regulated Firms: A Buyer's Checklist

Retrieval-augmented generation, usually shortened to RAG, is the pattern behind most useful AI search: find the passages in your own documents that answer a question, then have a language model write an answer grounded in those passages, with references back to the source. It is the difference between an assistant that guesses and one that cites.

For a regulated organisation in the UK, the pattern is attractive and the risks are specific. The documents you most want to search are exactly the ones you are least able to send anywhere: client files, patient correspondence, case notes, contracts, tender responses, security documentation. This checklist is written for the people who have to sign the purchase off: information governance leads, data protection officers, partners, heads of IT and security managers.

1. Where does processing actually happen?

Ask the vendor to draw the data flow on one page. Every arrow matters: indexing, embedding, storage, the model that writes answers, logging, monitoring and support access. "On-premise" should mean all of those steps run on infrastructure you control, whether that is your own servers or a cloud subscription in your own name. If any step calls out to a third-party API, the product is not on-premise, however it is marketed.

With Athena, every step runs on your hardware or inside your own Azure subscription. There are no external calls and no telemetry.

2. Controller, processor, or neither?

Under UK GDPR you remain the controller of personal data in your documents. The question is whether the AI vendor becomes a processor. If documents or prompts are sent to a hosted service, it almost certainly does, and you need a processing agreement, a view on sub-processors and, often, an assessment of international transfers. If the software runs entirely inside your environment and the vendor has no access to the content, the picture is much simpler. Get the vendor's answer in writing, including what support staff can and cannot see.

3. Have you scoped the DPIA?

Large-scale processing of sensitive documents with new technology is a strong signal that a data protection impact assessment is needed. A good vendor will help you complete one rather than leave it to you. Useful inputs to ask for: the full data flow, retention behaviour for queries and answers, how personal data is handled at index time, how access is controlled, and how a data subject access request would be answered.

4. Does it respect the permissions you already have?

This is the most common failure in AI search. A system that indexes a share and then lets anyone ask questions about it has quietly flattened your access controls. Check that results are filtered by each user's existing permissions, at query time, using your identity provider. Athena integrates with Microsoft Entra ID or your existing OIDC single sign-on and only answers from documents the person asking can already open.

Then test it. Ask the vendor to show a junior account failing to see a restricted document that a senior account can see.

5. What happens to personal data at index time?

Many questions do not need names, addresses or identifiers to be answered well. Ask whether personal data is stored in the search index at all. Athena anonymises personal data during indexing, so the index holds the meaning needed to find answers without holding the identifiers. That reduces what is at stake if the index itself were ever exposed.

6. Can it run disconnected?

Some environments are air-gapped by policy: defence suppliers, parts of healthcare, some legal and public sector teams. Ask whether the product works with no internet connection once models are downloaded, and whether updates can be delivered offline. Athena's Enterprise plan includes offline, air-gapped deployment and on-site installation.

7. Is there an audit trail you can actually use?

You will be asked who searched for what, which documents informed an answer, and when. Make sure every query, access event and answer is logged in a form you can export and review, and that logs stay inside your environment. Athena records every query and access event to support your compliance reporting.

8. How are answers grounded, and how is that measured?

Ask every vendor two questions: how do you stop the model inventing sources, and how do you know? A credible answer includes a measurement on a defined question set, a clear description of how answers were judged, and the conditions of the test. Treat any figure without that context with caution.

We publish ours. In our latest measurements, Athena's answers were judged correct 90.1% of the time on a 312-question demonstration set, and a separate release measurement found no invented citations. The measurement note on the Athena page explains exactly how that was done and what it does and does not tell you. Your own documents are the real test, which is why we recommend a pilot on a small, representative collection.

9. What is the cost model as usage grows?

Per-seat and per-query pricing look modest in a pilot and grow quickly when a tool becomes part of daily work. For public bodies and partnerships in particular, unpredictable bills are hard to budget. Athena is licensed per organisation at a flat monthly fee, with no per-seat or per-query charges, so adoption does not change the price.

10. Who owns the model choice?

Ask which models the product uses and whether you can change them. A product locked to one proprietary model ties your roadmap to someone else's. Athena runs open-weight model families locally and lets you swap or upgrade them.

Questions to ask any vendor

Take this list into every demonstration:

1. Show me the full data flow. Which steps, if any, leave our environment?
2. Will you be a processor of our personal data? Who are your sub-processors?
3. What can your support staff see, and how is that access controlled and logged?
4. How are our existing permissions enforced at query time?
5. Is personal data stored in the index? If so, why?
6. Can it run with no internet connection? How are updates delivered?
7. What is logged, where, and for how long? Can we export it?
8. How do you measure answer quality and invented citations? Show me the method.
9. What will it cost when every member of staff uses it daily?
10. Can we change the model without changing vendor?

The short version

On-premise RAG is the right shape for UK regulated organisations because it keeps the most sensitive part of the problem, your documents, where your obligations already apply. The details still matter. Insist on a one-page data flow, permission filtering you have seen work, an audit trail you can export, and quality figures that come with their method.

If you would like to see how Athena answers each of these questions, book a demo or ask us about a pilot on your own documents.