Siloed data across systems
ERP, MES, PLM and SCADA each stand alone. Equipment, process, quality and order data are hard to correlate, diagnosis means a human stitching several systems together, and the AI never sees the full context.
Microsoft Foundry · Enterprise AI agent platform
Buying the toolis only step one
GitHub Copilot lets your developers build prototypes faster. Microsoft Foundry turns those prototypes into production systems with identity, observability and governance. NovaTech delivers the part in between — from a demo that runs, to a system your company is willing to put into production.
Reality Check
The gap between prototype and production
What stops enterprises is never “the model isn't smart enough”. It is that once the prototype works, nobody is willing to let it touch real data, real systems or a real production line.
Knowledge walks out of the door
Fault handling depends on veteran engineers. SOPs, equipment manuals and historical work orders were never turned into searchable, traceable knowledge assets — when people move on, the capability moves with them.
No governance, no observability
Agents lack unified identity, logging and call-chain governance. Their behaviour cannot be explained or attributed. Fail the production security review and you stay in POC forever.
Root cause
The break is not in how smart any single tool is. It is the absence of an enterprise foundation that connects data, intelligence, orchestration and governance.
From Copilot to Foundry
From Copilot to Foundry: the full path
If you are already running GitHub Copilot, this section is about your next step.
Stage 1 · Prototype
GitHub Copilot
A developer uses agent mode and has something working in two or three days. This step is fast now; it is not the bottleneck.
→
The gap · why it stalls here
IT says no
The prototype runs on someone's laptop. Whose identity does it use to reach data? How many model calls, at what cost? How do you trace a failure? Nobody can answer.
→
Stage 2 · Production
Microsoft Foundry
Hosted Agents accept agents written with MAF, the GitHub Copilot SDK, LangGraph or the Claude Agent SDK without a rewrite.
What deployment gives you
Every agent gets its own Entra Agent ID, so no secrets sit in the image · OpenTelemetry traces flow straight into Application Insights · scale-to-zero means you pay nothing while idle, and the next request brings it back with session state and files intact.
Capability
Microsoft Foundry: what matters
Microsoft publishes the full feature list. Here are only the four things that decide whether it reaches production.
Foundry Models
11,000+ foundational, open, reasoning, multimodal and industry models behind one API. Switch models when costs bite — your business code stays put.
Foundry Agent Service
Hosted Agents package your code as a container on managed infrastructure, with built-in identity, automatic scaling, persisted session state and versioning.
Foundry IQ
One SLA-backed retrieval endpoint across Work IQ, Fabric IQ, Azure SQL, File Search and MCP sources — so every team stops building its own RAG.
Foundry Tools / Toolboxes
A single managed MCP endpoint for every tool type; skills are versioned in a project-scoped catalogue and discoverable by any agent in the project.
Security & Governance
Where Microsoft genuinely pulls ahead
Model capability is converging fast. Security and governance are not. This is the layer that is hardest to match with a self-built stack — because it is not one feature, it is identity, data, threat and compliance closing at the same time.
01
Model layer · Foundry Guardrails
Content filters, Prompt Shields against prompt injection, and abuse detection protect managed inference endpoints directly.
More importantly, Guardrail Policies let you mandate minimum guardrail controls at subscription or resource-group level, evaluated automatically by Azure Policy, with one-click remediation for non-compliant deployments — security is enforced by the platform, not left to developer discipline.
Content FiltersPrompt ShieldsAbuse DetectionGuardrail Policies
02
Threat layer · Microsoft Defender for Cloud
Security posture recommendations surface misconfigurations and risk. With threat protection enabled for Foundry Tools, it detects jailbreak and user input attacks and raises them as Defender alerts.
Those alerts can be correlated with Purview audit records of AI interactions to support investigation and attribution.
Security postureJailbreak detectionInput attack detectionCorrelated investigation
03
Data layer · Microsoft Purview
Once enabled on a subscription, AI interaction data from every application and agent in that subscription flows into Purview, opening up enterprise data compliance in one step.
AuditSensitive info types (SIT)DSPM for AIInsider Risk ManagementCommunication ComplianceData LifecycleeDiscovery
04
Identity layer · Entra Agent ID
Each agent is a first-class identity, inventoried in the Entra Agent Registry (including third-party agents), governed by RBAC, conditional access and least privilege, with secrets centrally held in Key Vault.
Entra Agent IDAgent RegistryRBAC / conditional accessKey Vault
What NovaTech does
Microsoft provides all four layers, but none of them configure themselves. Which guardrail policies to mandate, how to wire Defender and Purview, where to draw the identity and network boundaries — that is what NovaTech actually delivers on every project.
Reference Architecture
A multi-agent platform for a global manufacturer
Microsoft Foundry + Microsoft Agent Framework
Background
The client operates production sites across several regions and faces exactly the three breaks described above: data scattered across systems, fault-handling experience that cannot be reused, and agents without unified governance that cannot pass a security review.
Architecture: one orchestration layer, six business agents
With Microsoft Foundry as the platform foundation and Microsoft Agent Framework as the unified orchestration layer, every agent registers natively with the orchestrator, which handles routing, prioritisation and escalation — so no new silos are created.
MonitoringDiagnosticsTroubleshootingPredictive maintenanceProduction optimisationAutonomous action
End-to-end loop: from alarm to work order
- 1Alarm raised
- 2Monitoring agent confirms anomaly
- 3Diagnostics agent correlates sources
- 4Knowledge graph lookup
- 5RAG retrieval
- 6Troubleshooting agent drafts a fix
- 7Recommended action with confidence
- 8Execution agent · human approval
- 9Work order written back to business systems
What RAG means here
Azure AI Search with embedding, hybrid search and rerank turns scattered manuals, SOPs and historical work orders into searchable knowledge with citations back to source. On top sits an equipment–process–fault–SOP knowledge graph, so a diagnosis can always explain what it is based on.
What MCP means here
Azure API Management acts as the single gateway, carrying REST, gRPC, MCP Server and webhook traffic together. The agent landing zone reserves native MCP registration, so new capabilities extend smoothly without touching the orchestration layer.
Delivery philosophy
Close one loop first, then replicate at scale. One scenario matures, one scenario goes live. Each is signed off against a single quantifiable business metric rather than a big-bang rollout.
NovaHub Gateway
Where NovaTech adds its own: a cross-ecosystem model gateway
Foundry unifies everything inside the Microsoft ecosystem. But your reality is that you will not use only one model, and not only one ecosystem.
The problem
Complex reasoning goes to GPT; cost-sensitive batch work goes to DeepSeek or Kimi. The result is API keys scattered across teams, and nobody can say what was spent this month, by whom, or whether it was worth it.
NovaHub Gateway
Built by NovaTech, part of the NovaHub enterprise AI platform.
- Central key distribution — developers no longer hold vendor keys individually
- Tiered quotas — caps per team and per project, so one experiment cannot burn the monthly budget
- Token-level metering — usage and cost attributed to departments, ready for FinOps
- One endpoint, many models — GPT, DeepSeek, Kimi and others; switching does not change business code
How it relates to Foundry
Complementary, not a replacement. Foundry unifies models, guardrails and governance inside the Microsoft ecosystem; NovaHub Gateway unifies calls and cost across ecosystems — including models that are not in the Foundry catalogue. In real projects the two usually run side by side.
Delivery
How NovaTech delivers
Selling a subscription is not the work. Choosing the scenario, designing the architecture, configuring security, controlling cost — that is the work.
Scenario selection and feasibility
- Inventory existing systems and data sources, confirm data is reachable
- Pick one high-frequency, fast-payback scenario to run first
- Agree a single quantifiable acceptance metric
Feasibility assessment · Scenario priority list
01
02
Architecture and model selection
- Foundry foundation and orchestration layer design
- Model selection and cost model, including NovaHub Gateway
- Security and compliance architecture: identity, network, data boundaries
Architecture design · Security & compliance plan
Agent / RAG / MCP build and integration
- Agent development and multi-agent orchestration
- Knowledge assetisation and RAG with citation traceability
- Business system integration through APIM (REST / MCP)
Working system · Integration contracts
03
04
Productionisation and ongoing operation
- Guardrail policies, Defender and Purview onboarding
- End-to-end observability and alerting
- Cost review and quota tuning
Live operation · Monthly operating report
Why NovaTech
Why NovaTech
NovaTech is a Microsoft Solutions Partner. We have worked on the Microsoft stack since 2011 and delivered end-to-end services — strategy through to ongoing operations — for more than 1,000 enterprises.
Full-stack Microsoft credentials
All six Microsoft Cloud Solution Partner designations:
- Modern Work · Data & AI (Azure) · Digital & App Innovation
- Infrastructure (Azure) · Business Applications · Security
Advanced Specialisations and dual-channel credentials:
- AI on Azure · Infra & Data Migration
- Adoption & Change Management · Custom Solutions for Microsoft Teams
- 21V CSP / OSPA / NCEI · HK 1T/2T CSP · SG 2T CSP · ECIF Certified Partner
Microsoft hackathon champion two years running
2025
“Eva” took first place at the Microsoft China Copilot AI innovation competition
2026
“Sales Deal Agent”, a multi-agent solution, won the Grand Prize (first place) at the Microsoft Frontier Agentic Hackathon, alongside the FY26 Microsoft Market Expansion Pioneer Award
Winning two years running shows more than technical depth — it shows the ability to turn frontier technology into something deliverable, operable and scalable.
FAQ
Frequently asked questions
Q1. How is this different from building it ourselves?
A. The expensive part of building it yourself is never the model calls. It is three things: giving every agent an enterprise identity that can be audited, turning the whole call chain into traceable logs, and making safety guardrails a platform obligation rather than developer discipline. Foundry turns all three into platform capabilities — Entra Agent ID, OpenTelemetry tracing, Guardrail Policies. Building that layer yourself usually takes longer than the business logic, and then you maintain it forever.
Q2. Can we change models later, or are we locked in?
A. Foundry Models puts 11,000+ models behind one API, so switching does not change your business code. If you also need models outside the Foundry catalogue, NovaHub Gateway adds a cross-ecosystem entry point above it and lowers the switching cost further. We do not recommend locking yourself to any single vendor.
Q3. Where does our data sit, and is it used for training?
A. Customer data is not used to train models. Foundry supports dedicated deployment and private networking, with secrets centrally held in Key Vault. With Microsoft Purview enabled, auditing, sensitive information classification and data security posture management for AI interactions are brought under one roof. The exact data boundary and network design is something NovaTech confirms with you item by item during the architecture phase.
Q4. How do we connect existing systems, and what is MCP?
A. MCP (Model Context Protocol) is the standard for letting agents call external tools and data. In practice we use Azure API Management as a single gateway carrying REST, gRPC, MCP Server and webhooks together — existing systems connect over their current protocols while new capabilities register via MCP. You do not have to re-engineer your business systems before adopting AI.
Q5. How quickly will we see something?
A. We close one high-frequency, fast-payback scenario first, sign it off against a single quantifiable business metric, then replicate by line or department — one scenario matures, one goes live. The timeline depends on data reachability and integration complexity, and we give a firm schedule at the feasibility stage rather than promising a number up front.
Q6. How is it charged?
A. Two parts. Azure bills by tokens, compute and individual service usage, which varies considerably with the architecture; NovaTech charges for implementation and ongoing operation. Because there are too many variables, we do not publish a falsely precise estimate here — the feasibility stage produces a cost model and range based on your actual scenarios.