“Train AI on our company” can mean several different projects. We first identify the task: finding current policy, producing a consistent document, classifying records or assisting an application workflow. That choice determines the data, infrastructure, evaluation and budget.

Retrieval: supply the right information at answer time

Retrieval-augmented generation, or RAG, searches approved information and places relevant material into the model's input. It does not change the model's weights. This is usually our first approach for questions about company manuals, procedures and reference material because sources can be refreshed and cited. NVIDIA's RAG explanation

We define which sources are permitted, how documents refresh, how citations work and how permissions follow the requesting user. The evaluation includes questions the sources cannot answer. The system should acknowledge missing information instead of presenting an invented policy as fact.

Fine-tuning: change behavior using examples

Fine-tuning adapts a pretrained model with training examples. Parameter-efficient methods update a smaller set of parameters. We consider this when a measured baseline has a recurring task or formatting problem that better retrieval and instructions do not resolve. Training a foundation model from scratch is a different, much larger research scope. NVIDIA customization guidance

Training examples need appropriate rights, quality review and a held-out evaluation set. A model that memorizes the examples has not demonstrated useful generalization. Fine-tuning also does not guarantee current facts, confidentiality or correct business decisions. It can be combined with retrieval when both needs exist.

Integrations turn an assistant into a business tool

An internal interface can search approved documents, draft a service summary or classify an incoming request. An integration that changes records adds another responsibility. Our default design starts with read access and introduces narrowly scoped actions with explicit permissions, validation and approval where the workflow requires it.

A proposal should identify each connector, its owner, supported interface, refresh frequency and failure behavior. The test should include an unavailable system, duplicate request and an unauthorized operation, not just a successful demonstration.

Private deployment needs explicit controls

On-site hardware does not automatically make an AI service private. Missing retrieval permissions, unsafe actions and unapproved external content can expose information. We define identity, source permissions, network egress, logging, retention, backup and support access. Testing includes attempts to access another user's documents and to redirect the assistant through malicious instructions in retrieved content. NVIDIA AI Red Team guidance

Model licenses are checked for the exact model and version. An open-weight label is not a universal commercial-use permission. For example, the publisher marks Qwen3-32B as Apache-2.0; that is a statement about that model, not every model with a similar name. Qwen3-32B model card

Measure the business result

Before production, agree on representative questions, citation quality, access controls, simultaneous requests and response-time targets. Compare the result with the current process. Keep a human review path for consequential output and a clear way to report errors.

See private AI packages and planning budgets. Technical references reviewed September 6, 2026.