Back to Blog

    Agentic AI

    RAG on Azure: which cost blocks to plan for

    Which cost blocks a RAG system on Azure creates and how to estimate Azure AI Search, models, availability and operations in a traceable way.

    August 9, 20263 min read
    Marcel Haas

    Marcel Haas

    Solution Architect, CEO

    marcel.haas@cnext.ch
    In this article

    A custom RAG stack has fixed and variable costs: search service, replicas and partitions, embeddings, ranking, model calls, data preparation and operations. The information was checked on 8 September 2026 against the official documentation on Azure AI Search pricing, SKUs and capacity, availability and Azure OpenAI pricing.

    What the price depends on

    Azure prices depend on region, currency, agreement, SKU, partitions, replicas, model, deployment type and token volume, among other things. The official pricing pages provide configurable values and point out that actual prices may differ. A reliable figure therefore only comes from your own configuration.

    The cost blocks

    BlockCost drivers
    Azure AI SearchSKU, number of partitions and replicas, runtime
    Ranking and agentic retrievalEnabled feature and usage according to the current pricing page
    EmbeddingsChosen model and processed tokens
    Answer modelModel, deployment type, input and output tokens
    PreparationOCR, parsing, chunking and reprocessing
    OperationsPermissions, monitoring, evaluation, updates and support

    A search unit is replicas × partitions. For the query SLA, Microsoft requires at least two replicas; for the SLA covering queries and indexing updates, at least three. Whether you need this SLA follows from your business requirements.

    Two reliable estimation patterns

    Small internal agent: Record index size, rate of change, requests, tokens per request and the availability you need. Cost production and any test environment separately. A free tier is a development offer, not a production baseline.

    Larger document collection: Measure the actual index size after chunking and vectorisation. The number of documents alone does not determine a particular number of partitions. Explicitly plan for re-indexing, OCR, permission filters and growth.

    Use the Azure Pricing Calculator for each option and record region, retrieval date and assumptions. Do not transfer prices or model availability from another region to Switzerland North or West Europe without checking.

    Microsoft 365 content without your own index

    Copilot Studio can connect SharePoint as a knowledge source and generate answers in the context of the authenticated person. This can avoid a custom indexing pipeline; licence, credit, connector and governance costs still need to be considered.

    Microsoft 365 Copilot uses Microsoft 365 data according to the permissions of the signed-in person. The cost lies in the required Microsoft 365 and Copilot licences; scope and rights are governed by your agreement.

    In a custom RAG solution with Azure AI Search, SharePoint permissions are a design task in their own right: check the supported data source, the status of any preview features and the security filter architecture. An application must never rely solely on correct ACLs existing in the source system.

    When your own index can make sense

    • Sources outside Microsoft 365 need controlled retrieval.
    • Ranking, chunking, evaluation or the interface need to be controlled individually.
    • Several sources are to be combined with their own permission and freshness logic.
    • Contractual or technical requirements call for an explicitly defined Azure architecture.

    Conversely, native grounding can reduce operating effort for clearly scoped SharePoint content. The decision should be based on quality, permissions, residency, availability and total cost of ownership.

    Check data residency too

    With Azure OpenAI, prompts and responses in regional Standard deployments are processed within the selected Azure geography; Global and Data Zone deployments follow different rules (Microsoft data privacy documentation, checked on 8 September 2026). Model and feature availability depends on the region. An Azure resource alone therefore guarantees neither processing exclusively in Switzerland nor the availability of every model in Switzerland North.

    Conclusion

    The search service can be a significant fixed cost block; how large it is depends on SKU, replicas, partitions and runtime. A verifiable RAG estimate states SKU, region, replicas, partitions, models, tokens, environments and operating effort, and links the pricing sources used on the reference date.

    Further reading: Company Brain platform comparison

    Agentic AIAzure AICopilot StudioSharePointSchweiz
    Teilen:

    AI tools may have supported the creation of this article. CNEXT is responsible for the published content. Product features and prices can change; check the linked primary sources for current details.

    Marcel Haas

    Marcel Haas

    Solution Architect, CEO

    6x Microsoft Applied Skills

    Build a RAG cost estimate with CNEXT

    We estimate search, models, environments and operations for your use case with you.