Cloud Infra, Data & AI Platform
AI Apps
Applications and agents with retrieval, evaluation, and a human handoff designed in.
Build applications and agents as one chain rather than one link of it: the use case that justifies the work, the knowledge and data the answer is grounded on, retrieval and index design on Azure AI Search, then the agent, its tools, and the channel it is used through. Evaluation, content safety, and a controlled human handoff are designed into the workflow rather than added after the pilot demonstrates well, so a change to a model or a prompt is measured against a fixed question set instead of judged by impression. The approach is Microsoft Foundry for the platform and Microsoft Agent Framework where the work spans several systems and a single retrieval step is not enough.
Microsoft FoundryAICopilot StudioThe Microsoft products behind it
What each product protects or produces, how it works, and the problem it answers.
- Microsoft Foundry
Where an AI workload is built, released, and watched
The platform an AI application or agent is built on, with the model catalogue, evaluation, and tracing in one place rather than assembled per project.
- Model Catalogue
- Agent Service
- Prompt Engineering
- +3 more
Azure OpenAI ServiceEnterprise language models inside your own tenant
The language, reasoning, and embedding models a workload calls, deployed inside the customer's own Azure subscription and network.
- Language Models
- Reasoning
- Embeddings
- +3 more
Azure AI SearchThe index an answer is grounded on
The retrieval layer under a knowledge assistant, combining keyword and vector search with the access rules that decide what a given user may be answered from.
- Retrieval
- Vector Search
- Hybrid Search
- +3 more
Azure AI Document IntelligenceTurns documents into data a workload can use
Extraction that reads scans, tables, and forms into structured text, which is what makes a document estate retrievable in the first place.
- Document Extraction
- OCR
- Table Recognition
- +3 more
Copilot StudioAgents a business team can build and keep
The low code path to an agent, where the people who own the process maintain it instead of queueing behind a development team.
- Low Code Agents
- Topics and Triggers
- Knowledge Sources
- +3 more
Microsoft Agent FrameworkOrchestration for work one prompt cannot finish
The framework behind a multi agent workload: planning, registered tools, and state that survives a long running task.
- Agent Orchestration
- Tool Calling
- Planning
- +3 more
Six service stages, from use case to a workload in production
An assistant that demonstrates well and an assistant that survives production are different pieces of work. These six stages are the order the work runs in, from choosing the use case to operating the workload against an evaluation set.
Use Case and Value Framing
- Choosing the process where an assistant or agent changes the outcome, not the demo
- Naming the decision or task the workload takes over, and who stays accountable for it
- Defining what a good answer is before anything is built, so success is measurable
Knowledge and Data Grounding
- Locating the documents, records, and systems the answer has to come from
- Extraction and chunking with Azure AI Document Intelligence, including tables and scans
- Carrying the source entitlement through, so a document a user may not read cannot reach them
Retrieval and Index Design
- Index and field design on Azure AI Search, keyword, vector, and hybrid
- Ranking, filters, and the security trimming that enforces access at query time
- Citations returned with the answer, so a reader can check where it came from
Agent and Orchestration Build
- Model selection and prompt design on Microsoft Foundry
- Tools, planning, and specialist agents with Microsoft Agent Framework
- The channel the workload is used through: web, Microsoft Teams, messaging, or voice
Evaluation and Guardrails
- An evaluation set built from real questions, scored before release and after each change
- Groundedness, relevance, and safety measured rather than asserted
- Content filtering and a controlled human handoff for what the workload should not answer
Operate and Improve
- Monitoring latency, cost per request, and failure modes in production
- Reviewing unanswered and low confidence questions as the backlog for the next index change
- Model and prompt changes rolled out against the evaluation set rather than by impression
Two architecture paths
The default path is a grounded assistant: one retrieval index, one model, citations returned with every answer. The multi agent path is for work that spans several systems and needs planning and tools rather than a single retrieval step.
Grounded assistant
Default path- Source documents
- Extract and chunk (Document Intelligence)
- Index (Azure AI Search)
- Retrieve and rank
- Ground the model (Microsoft Foundry)
- Answer with citations
Medallion layers
Multi agent orchestration
Alternative- Request
- Planner
- Specialist agents
- Registered tools
- Durable orchestration
- Shared memory (Cosmos DB)
- Answer with citations
This path is for work that spans several systems, where one retrieval step is not enough and the workload has to plan, call tools, and hold state across a long running task.
Product stack by layer
The components used on both paths, grouped by the layer they serve. Components that need a licence or capacity of their own are marked, so the scope is clear from the start.
Cloud: Intelligence
Cloud- Azure OpenAI in Foundry Models
- Model, the language and reasoning models the workload calls
- Microsoft Foundry
- Platform, where agents and AI apps are built, deployed, and observed
- Azure AI Search
- Retrieval, the index the answer is grounded on
- Microsoft Agent Framework
- Orchestration, planning and tool calling across agents
- Azure AI Document Intelligence
- Extraction, turning documents into structured, chunkable text
- Azure AI Speech
- Speech, transcription and synthesis for a voice channel
- Azure AI Vision
- Vision, image and video understanding
- Azure AI Language
- Language, classification, extraction, and translation
- Azure AI Content Safety
- Safety, prompt and response filtering with configurable thresholds
- Azure Machine Learning
- Machine learning, the lifecycle for predictive and custom models
- Microsoft Copilot Studio
- Low code, agents built and maintained by a business teamNeeds Fabric capacity
Cloud: Experience
Cloud- Azure App Service
- Application, the hosted web or API surface
- Azure Static Web Apps
- Application, static front end with a managed API
- Azure Container Apps
- Application, containerised services with scale to zero
- Azure AI Bot Service
- Channels, one agent reaching Teams, web, and messaging
- Azure Communication Services
- Channels, voice and telephony for a realtime assistant
- Azure SignalR Service
- Transport, streaming tokens and events to the browser
Cloud: Integration
Cloud- Azure Functions
- Compute, event driven functions for tools and ingestion
- Azure Durable Functions
- Compute, long running orchestration with checkpointing
- Azure Logic Apps
- Workflow, connectors into line of business systems
- Azure Event Grid
- Events, routing between ingestion, indexing, and evaluation
- Azure Stream Analytics
- Streaming, continuous signals into the workload
- Azure Container Registry
- Supply chain, the images the services run from
Cloud: Data
Cloud- Azure Cosmos DB
- Storage, conversation state, memory, and vector search
- Azure AI Search
- Storage, the retrieval index the assistant queries
- Azure Blob Storage
- Storage, the source documents and generated artefacts
- Azure SQL Database
- Storage, managed relational data the agent reads and writes
- Azure Database for PostgreSQL
- Storage, open source relational data with vector support
- Microsoft Fabric
- Analytics, the governed data the AI workload reads fromNeeds Fabric capacity
Cloud: Foundation
Cloud- Microsoft Entra ID
- Identity, who the user is and what they may see
- Azure Managed Identity
- Identity, service to service access without stored secrets
- Azure Key Vault
- Secrets, keys and connection strings held outside the code
- Azure Monitor
- Operations, latency, cost, and failure signals
- Azure Log Analytics
- Operations, the query surface over traces and evaluations
- Azure Virtual Network
- Network, private access between the services
Delivery models
Five ways the work is run. The choice follows the shape of the requirement rather than the other way round, and models can be combined within one engagement.
Projects Based
Scope, timeline, and deliverables are defined up front.
Suited to
A first workload taken from use case to production, with a fixed scope and a defined evaluation set.
Man Month
Resourcing provided as a monthly allocation.
Suited to
Continuous AI development with a dedicated team, where the backlog is set by what production shows.
Quota Mandays
A quota of effort drawn down as it is needed.
Suited to
Incremental work on an assistant already running: new sources, new tools, index and prompt changes.
Training
Enablement on Microsoft Foundry, retrieval design, agents, and evaluation.
Suited to
Building the customer's own internal capability so they can run and extend the workload themselves.
Support and Maintenance
Post-implementation support, both preventive (monitoring, evaluation runs, cost and quality review) and corrective (fixing issues after they occur).
Suited to
Keeping a live assistant grounded and answering as the underlying documents and models change.
How the work is scoped
Engagements delivered by this capability, each with defined modules, deliverables and a timeline.
- Cloud Infra, Data & AI Platform
AI Readiness Assessment
An assessment that turns scattered AI experiments into a plan for a governed production platform.
- Work modules
- 3
- How scope is agreed
- Scope depends on the number of use cases and systems under review, agreed with the practice team before work begins
- Cloud Infra, Data & AI Platform
Agent Pilot Programme
A pilot engagement that puts one agentic workflow into a real process on Azure OpenAI Service.
- Work modules
- 3
- How scope is agreed
- Scope depends on the complexity of the chosen process, agreed with the practice team before work begins