A custom RAG stack has fixed and variable cost blocks: search capacity, replicas and partitions, embeddings, ranking, model calls, data preparation and operations. This article was reviewed on 8 September 2026 against the official documentation for Azure AI Search pricing, SKUs and capacity, reliability and Azure OpenAI pricing.
Why a universal monthly figure is misleading
Azure pricing varies by region, currency, agreement, SKU, partitions, replicas, model, deployment type and token volume. Microsoft's configurable pricing pages warn that actual prices may differ. Non-reproducible monthly totals and unsupported CHF conversions have therefore been removed.
| Cost block | Main drivers |
|---|---|
| Azure AI Search | SKU, partitions, replicas and runtime |
| Ranking and agentic retrieval | enabled feature and usage under current pricing |
| Embeddings | model and processed tokens |
| Answer model | model, deployment type, input and output tokens |
| Preparation | OCR, parsing, chunking and reprocessing |
| Operations | permissions, monitoring, evaluation, updates and support |
A Search Unit is replicas × partitions. Microsoft requires at least two replicas for a query SLA and at least three for an SLA covering queries and indexing updates. Whether that SLA is required must follow the business requirement.
Defensible estimation
For a small internal agent, record actual index size, change rate, requests, tokens per request and availability target. Cost production and any test environment separately. The Free tier is a development offer, not a production baseline.
For a larger corpus, measure index size after chunking and vectorisation. Document count alone does not establish a partition requirement. Include 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 a price or model availability from another region to Switzerland North or West Europe without checking.
Microsoft 365 content
Copilot Studio can use SharePoint as a knowledge source and answer in the authenticated user's context. This can avoid a custom indexing pipeline, but it does not remove licence, credit, connector or governance costs.
Microsoft 365 Copilot uses Microsoft 365 data under the signed-in user's permissions. Calling indexing “free” in isolation is misleading: the required Microsoft 365 and Copilot licences are still paid products.
Azure AI Search does not universally and automatically inherit SharePoint permissions for every custom RAG design. Developers must verify the supported source, the status of any preview feature and the security-filter design.
Data residency
For 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, reviewed 8 September 2026). Models and features vary by region. “On Azure” alone does not guarantee processing exclusively in Switzerland.
Conclusion
Search can be a material fixed cost, but no universal monthly range is defensible. A reviewable RAG estimate states SKU, region, replicas, partitions, models, tokens, environments and operating effort and links the dated source prices.
Further reading: Company Brain platform comparison · Contact us