Pattern 5: Serverless AI — Azure AI Patterns

Overview

The Serverless AI pattern is the lightest-weight deployment option. It uses Azure Functions (Consumption plan) for event-driven processing, Azure OpenAI with GPT-4o-mini for cost-efficient inference, and AI Search Basic for vector search.

This pattern gets you from zero to a working AI app in minutes. Ideal for hackathons, internal tools, MVPs, and PoC validation before investing in production infrastructure.

Note: Services use public endpoints with service-level auth. For production or regulated workloads, consider Pattern 6 (Secure AI) instead.

Key Characteristics

No VNets — Pure PaaS, public endpoints with service auth
Consumption Billing — Pay per execution, near-zero idle cost
GPT-4o-mini — Cost-efficient model (~10x cheaper than GPT-4o)
Optional APIM — API facade for rate limiting and key management
30-day Logs — Minimal retention for PoC scenarios

OpenAI Models

GPT-4o-mini	20K TPM
text-embedding-3-small	60K TPM

Functions Config

Runtime	Python 3.11
Plan	Consumption (Y1)
HTTPS Only	Yes
FTPS	Disabled

Storage

documents embeddings

Architecture Diagram

graph TB
    CLIENT["Client Apps<br/>Web / Mobile / Teams"]
    subgraph Optional["Optional API Layer"]
        APIM["Azure API Management<br/>Consumption Tier"]
    end
    subgraph Compute["Serverless Compute"]
        FUNC["Azure Functions<br/>Consumption Plan Y1<br/>Python 3.11"]
    end
    subgraph AI["AI Services"]
        AOAI["Azure OpenAI<br/>GPT-4o-mini"]
        AIS["Azure AI Search<br/>Basic Tier"]
    end
    subgraph Storage["Storage"]
        BLOB["Azure Blob Storage<br/>Documents"]
        FUNCSA["Functions Storage<br/>Runtime"]
    end
    subgraph Monitor["Monitoring"]
        MON["Log Analytics<br/>30-day retention"]
        APPINS["App Insights"]
    end
    CLIENT --> APIM
    APIM --> FUNC
    CLIENT -.->|"Direct"| FUNC
    FUNC --> AOAI
    FUNC --> AIS
    FUNC --> BLOB
    AIS --> BLOB
    FUNC --> APPINS
    APPINS --> MON

Bill of Materials

#	Service	Resource Name	SKU / Tier	Purpose	Monthly Cost
1	Azure OpenAI	`{base}-{env}-openai`	S0	GPT-4o-mini (20K TPM) + embeddings (60K TPM)	~$450
2	Azure AI Search	`{base}-{env}-search`	Basic	Vector + keyword search (2 GB)	$73.73
3	Azure Functions	`{base}-{env}-func`	Consumption (Y1)	Event-driven compute (pay per execution)	~$20
4	Functions Storage	`{base}{env}funcsa`	Standard_LRS	Functions runtime storage	~$2
5	Blob Storage	`{base}{env}sa`	Standard_LRS (no HNS)	Document and embedding storage	~$15
6	Log Analytics	`{base}-{env}-law`	PerGB2018, 30-day	Centralized logging	~$10
7	Application Insights	`{base}-{env}-ai`	Workspace-based	APM and request tracing	Incl.
8	API Management (opt.)	`{base}-{env}-apim`	Consumption	API facade, rate limiting, key mgmt	~$3.50/M calls
Estimated Total (Moderate PoC/Internal)					~$890/mo

Consumption pricing means near-zero cost when idle. Token costs based on ~200M tokens/month with GPT-4o-mini.

Service Breakdown

Azure OpenAI

S0~$450/mo

GPT-4o-mini for cost-efficient chat completions (20K TPM) and text-embedding-3-small for lightweight vector embeddings (60K TPM). ~10x cheaper than GPT-4o per token.

Azure Functions

Consumption Y1~$20/mo

Event-driven serverless compute. Python 3.11 runtime. Pay only for execution time. 1M free executions/month. HTTPS-only, FTPS disabled.

AI Search (Basic)

Basic$73.73/mo

Lightweight vector and keyword search. 2 GB storage, 1 replica, 1 partition. Semantic search on free tier. Upgrade to Standard for production.

Blob Storage

Standard LRS~$15/mo

Simple blob storage for source documents and pre-computed embeddings. No HNS (flat namespace). Cost-efficient for PoC-scale data.

APIM (Optional)

Consumption~$3.50/M

Optional API facade providing rate limiting, API key management, and usage analytics. Consumption tier means pay per call with no fixed cost.

Monitoring

PerGB2018~$10/mo

Log Analytics with 30-day retention. Application Insights for request tracing, dependency tracking, and exception logging.

Security & Networking

Public Endpoints

No VNet / Private Endpoints — services use public endpoints
Service-level authentication (API keys, managed identity)
Storage enforces TLS 1.2 and HTTPS-only
FTPS disabled on Functions
Optional APIM for API key management

Identity

Functions use System-Assigned Managed Identity
OpenAI endpoint injected via App Settings
AI Search name injected via App Settings
Optional APIM provides rate limiting
HTTPS-only on all endpoints

Use Cases

Internal Copilots

Slack/Teams bots, internal knowledge assistants, document Q&A for teams.

MVPs / Rapid Prototyping

Validate AI concepts before investing in full infrastructure. From idea to API in under 30 minutes.

Hackathons / Innovation

Deploy a full AI stack in minutes for innovation sprints and competitive hackathons.

Low-Ops Applications

No clusters, no VNets, no private endpoints. Minimal operational burden and near-zero idle cost.

Constraints & Considerations

Constraint	Mitigation
No VNet isolation — public-facing services	Acceptable for PoC; upgrade to Pattern 1/6 for production
Cold start latency on Consumption plan	Use Premium plan for latency-sensitive workloads
Basic AI Search has limited capacity (2 GB)	Upgrade to Standard for production scale
Not ideal for strict enterprise environments	Add private endpoints if moving to production
Limited scaling control	Acceptable for PoC; use dedicated plans for production

Serverless AI