Next-Generation Autonomous Intelligence & Knowledge Retrieval.

• 01What is RAG, and how is it different from an AI agent?

  RAG — retrieval-augmented generation — makes a language model answer from your content rather than from its training data: your documents are indexed, relevant passages are retrieved per question, and the model is constrained to answer only from those passages with citations. An AI agent goes further and acts — creating a ticket, updating a record, sending a response — through tools it is permitted to call. Answering is a retrieval problem; acting is a permissions, integration and accountability problem, and it needs an audit trail plus a human in the loop wherever a mistake would be expensive.

• 02Can an agent see documents a user should not have access to?

  Not if it is built correctly, and this is the single most common reason security refuses to sign off. Permissions must be carried through ingestion and applied as a filter during retrieval, so the index physically cannot return a passage from a document the requesting user could not already open. Filtering after retrieval is a common shortcut and an unsafe one, because the content has already entered the pipeline. We also apply sensitivity labels and test the boundary adversarially rather than assuming it holds.

• 03Should we fine-tune a model on our data?

  Almost certainly not, and this misunderstanding is one of the most common ways AI budget gets wasted. Fine-tuning teaches a model a style or a narrow classification behaviour; it does not reliably teach it facts, and it cannot keep up with content that changes. If you want answers grounded in your documents, that is RAG. Fine-tuning earns its place for a specific tone at scale, or high-volume classification where latency and cost matter — and even then we would want the RAG baseline measured first.

• 04Is our data used to train the models?

  No. Content sent to Azure OpenAI and Microsoft Foundry is not used to train Microsoft’s or OpenAI’s models, and everything we build runs inside your own Azure tenant and subscription with the region you nominate. We can put the specific contractual and technical position in writing as part of the engagement, because with AI this is usually the first question a board asks and it deserves a specific answer rather than a reassuring one.

• 05How long until something is in production?

  A first production use case typically runs 10 to 14 weeks: two weeks scoping and scoring, two building the evaluation harness and guardrails, four tuning the pipeline against it, three piloting with real users, and the balance on production hardening and handover. Subsequent use cases are considerably faster because the platform, guardrails and harness already exist. A demo can be produced far sooner than that — but a demo is not the deliverable, and the gap between the two is where most projects are lost.

• 06How much do AI agent and RAG projects cost?

  Every engagement is scoped individually, so we quote after the scoping call rather than guessing beforehand. After the free scoping call we work in two commercial models, a fixed-price project or a managed AI retainer, with time and materials only where a scope is genuinely open-ended. The variables that move the number are how many use cases are in scope, document volume, how many source systems are involved, content readiness, permission complexity, whether the agent takes actions as well as answering, evaluation depth, expected query volume, and any compliance or residency requirements. Azure and model consumption bills from Microsoft directly to your own subscription and is never resold through us.

Reviewed 6 September 2026 by the Congnivex Systems Microsoft practice · Next review March 2027