What is an AI knowledge system?
An AI knowledge system is a private retrieval-augmented generation deployment over an organisation's proprietary corpus, reports, case files, archives, telemetry, that lets staff query the institutional knowledge directly. At OpenKit we build these as integrated systems on top of the client's existing storage and identity stack, with permissioning honoured at the source.
How is this different from ChatGPT or Microsoft Copilot?
Public AI tools answer from their training data. They do not know your archive. They cite generic sources when the actual answer is in your 2019 report. A private knowledge system retrieves from your corpus only and cites back to the document. Confidence is grounded in your data, not synthesised from someone else.
Where does our data live?
Your data stays where it is. The retrieval layer reads from your existing storage (SharePoint, on-prem document stores, S3, your case management system). The LLM call goes to a configurable backend: a UK or EU region of a major provider, a private deployment on your cloud tenant, or on-prem inference if your obligations require it. We do not require migration to a new platform.
How do you handle access controls and permissioning?
The system respects whatever access controls are configured in your storage and identity systems. If a user cannot see a document in SharePoint, they cannot retrieve it through the AI layer either. Permissioning is checked at retrieval time against your existing identity provider, not pre-baked into the embeddings.
What about hallucinations?
Every answer includes a citation back to the source document or excerpt. If the system cannot find a relevant source in your corpus, it says so rather than making one up. The retrieval-grounded answer is the only answer the system returns. Hallucination rates on grounded answers measured in single-digit percentages on engagements where we benchmarked.
How is this different from your retrieval-augmented generation page?
The technical RAG page describes the engineering: embeddings, retrieval, evaluation, deployment patterns. This page describes the engagement from the buyer's side: what we build for an insights agency, a regulatory consultancy, a specialist law firm, a manufacturer. Both apply; this one is the audience-led version.
How is this different from a productised private AI setup?
Productised private AI offerings are configured wrappers over a generic LLM deployment. They work well when the client wants AI access without the engineering depth. Our knowledge-systems work is custom-built against the client's proprietary corpus and integrated with their existing storage, identity, and audit-logging stack. Different scope, different price point. The audit identifies which is the right fit for your business.
How long does a knowledge-systems engagement take?
A typical sequence starts with the AI Audit, which scopes the build. The build itself runs from a few weeks to a few months depending on the corpus size, the integration surface, and the regulated-controls work. We can usually demonstrate the retrieval layer running against a sample of your corpus within the first month.
How much does it cost?
Scoped per engagement after the audit, since the cost depends on corpus size, integration surface, and regulatory requirements. The audit that scopes the build is a fixed-fee engagement; build engagements are quoted after we have seen the corpus and the integration surface.
Do you work with regulated industries?
Yes. We have built knowledge systems for insurance media, manufacturing, cultural insights, and regulated professional services. ISO 27001 audit logging is on as standard. Sector-specific obligations (UK GDPR, FCA SYSC, NHS DSP Toolkit, SRA) sit at the integration layer per engagement.
Can we keep the system on-prem?
Yes. We deploy retrieval, embedding generation, and LLM inference on-prem when sovereignty or contract obligations require it. Open-weights models running on your GPU infrastructure are the most common pattern for fully on-prem deployments.