Enterprise GPT.

An enterprise GPT is a large language model deployed inside your own cloud tenancy or data centre, grounded in your documents and governed by your access controls. A public chatbot knows nothing about your business, honours none of your permissions and leaves you no audit record. The underlying model technology is the same class; the governance around it is not.

Your content stays inside your boundary and is not used to train a foundation model. Deployment runs in your VPC or on-premise, so documents and prompts do not leave your environment. Every session is logged for audit, data-leak prevention and redaction rules apply on both input and output, and retention is configured to match GDPR and India's DPDP obligations.

Retrieval keeps knowledge outside the model, so updating a policy means re-indexing one document rather than retraining. It also lets every answer carry citations, which is what makes output checkable. Fine-tuning changes tone, format and task behaviour effectively, but it does not reliably teach facts and it cannot enforce which user is allowed to see which document.

Permissions are enforced at retrieval, not in the prompt. Each indexed chunk carries its source document's access-control list, the user authenticates through your existing SSO, and the search filters to documents that identity can already open. Group membership syncs from your directory, so revoking access in the source system removes it from the assistant on the next query.

Four things: tokens consumed per question, which retrieved context size and conversation length dominate; embedding and re-indexing load as your content changes; GPU or hosted-inference capacity if you run open-weight models yourself; and integration and support effort. Routing simple questions to a smaller model and capping retrieved context are the two changes with the largest cost effect.

With an evaluation set of real questions and approved answers, scored on groundedness — whether every claim traces to a retrieved source — relevance, and correct refusal when the answer is not in the corpus. Answers cite sources so users can verify them. Retrieval failures, not model failures, cause most wrong answers, so chunking, hybrid search and reranking get tuned first.