Private AI for life sciences

Research teams have more documents than anyone can read, and most of them cannot go near a public AI tool. A private platform lets your scientists search and summarise them inside your own estate, with collaborators seeing only what their agreements allow.

Data scope still open? Start with an AI readiness assessment
One approved collection and a use case? Start with a proof of concept

What a research platform reads

The useful material is mostly text: published literature, internal reports, assay and study reports, protocols, electronic lab notebook exports, regulatory drafts, standard operating procedures and the minutes of a hundred project meetings. A private platform indexes the collections you choose, and answers questions with references back to the source documents.

Indexing makes copies, so the design lists them. The source documents stay where they are. Text extracted from them, the embeddings and the vector index are new copies, stored in your tenancy or data centre. Prompts, answers, logs and backups are kept there too, for the periods you set, when the model runs on compute you control. With a managed model deployed in your own subscription and UK region, prompts and the passages sent with them are processed by the provider's model in that region, under your agreement: still a crossing to assess before any research document goes in.

It does not need raw instrument data or patient records to be useful, and in most first projects it should not have them.

Three kinds of data, three boundaries

Your research documents

Indexed into collections that follow your existing access groups, so a project team searches its own work and nobody else's.

Collaborator material

Kept in its own collection, on the terms of the agreement it came under, and readable only by the people that agreement names. When the collaboration ends, the source files, extracted text, embeddings and index entries for that collection are removed; logs and backups follow the retention you set.

Personal and patient level data

We propose leaving it out of scope, in its own secure environment, with the platform receiving approved aggregated outputs at most. That is a decision for your scope. If it is brought in, pseudonymised data is still personal data under UK GDPR. Many UK research environments apply the Five Safes (safe people, projects, settings, data and outputs) to the same question.

Collaborators without oversharing

Partners sign in through your identity provider as guest accounts, so you can see who they are and switch them off. Each guest is scoped to named collections. Their questions and the sources behind each answer can be logged in your tools, which matters when a collaboration agreement asks you to show how shared material was used. When the work ends, their access is removed and their collection is removed on the terms above.

How identity and logging work in detail is on our assurance page.

Before you file

Patent attorneys generally advise keeping an invention confidential until a filing is made, because a public disclosure before the filing date can count against novelty. Whether typing an invention into a public AI tool counts as disclosure depends on that tool's terms, and is a question for your patent attorney, not for us.

What a private platform changes is who handles the text. With a model on compute you control, invention notes, draft claims and the data behind them are searched and summarised on infrastructure you run, under your own access rules. With a managed model service, prompts and passages go to that provider under your agreement, so decide whether invention material may use it before any goes in.

When AI touches regulated records

A platform that searches literature and internal reports is a very different system from one whose output feeds a good practice (GxP) record or decision. For the second, the MHRA's data integrity guidance expects the system to be validated for its intended use, with records that are attributable, legible, contemporaneous, original and accurate, and an audit trail to match.

So the first decision is which of the two you are building, and it belongs to your quality team: the intended use, whether validation is needed and to what standard, and which audit records the system must keep. Expectations differ across GxP activities; for good laboratory practice, for example, the MHRA notes that OECD guidance takes precedence. We document the intended use and its limits as your quality team defines them.

What auditors and partners will ask

Where did this answer come from?

Answers can point back to the passages they drew on, so a scientist can open the source document they are allowed to see.

Which model, which version?

The model version in use is recorded, and a new version goes through your change process and is tested against the same questions.

Who could see what?

Collections, groups and guests are listed, and access reviews are part of running the platform.

Where is the data?

The illustrative architecture shows where documents, embeddings, the model, logs and backups sit, and every way out.

Start with your documents.

Tell us which collections matter and which data must never move. We will come back with a scope.

Thank you. Your request is on its way. We will reply from info@tlgsolutions.co.uk to arrange a first discussion.

That did not send. Please email info@tlgsolutions.co.uk instead, and we will reply from there.