Module 3 of 11
Design overall AI strategy for business solutions
Use the Cloud Adoption Framework to build an enterprise AI strategy, and decide when to extend Copilot versus build custom.
Exam questions Q21–Q30 19 units on Learn
Open module 3 on Microsoft Learn ↗ Watch the episodes (93 min)
Why this module matters
This is the largest module in the path — nineteen units — and it carries proportionally more of the skills outline than any other. It spans the Cloud Adoption Framework, multi-agent design, knowledge source selection, prompt libraries, small language models, the AI Center of Excellence, and regional data regulation.
The connective tissue is standardisation. Almost every correct answer favours a governed, reusable, centrally-supported approach over per-team improvisation.
The one big ideaStrategy questions reward the answer that scales: shared governance, reusable patterns, a portfolio prioritised on value and risk, and a Center of Excellence that makes the good path the easy path.
What Microsoft Learn covers
Units in this module, as published on Microsoft Learn:
- Introduction
- Implement AI adoption process with Azure
- Design AI agents for business solutions
- Design a multi-agent solution
- Develop use cases for prebuilt Microsoft 365 Copilot agents
- Define solution rules and constraints for AI components
- Determine generative AI knowledge sources for agents built in Copilot Studio
- Determine when to build custom agents or extend Microsoft 365 Copilot
- Determine when custom AI models should be created
- Provide guidelines for creating a prompt library
- Develop use cases for customized small language models
- Provide prompt engineering guidelines and techniques
- Identify key business user roles for AI workloads
- Evaluate regional and local AI data regulation compliance requirements
- Include elements in a Microsoft AI Center of Excellence
- Design AI solutions using multiple Dynamics 365 apps
- Design user prompt training for AI solution adoption
- Module assessment
- Summary
Watch this module
Microsoft Learn publishes a free video course for AB-100 on YouTube. This module is covered across two episodes, 49 min and 44 min. An instructor walking through the module is often faster than reading it twice.
Key concepts and definitions
| Cloud Adoption Framework (CAF) | Azure's adoption methodology. The stages you must recognise: Strategy (motivations, outcomes, justification), Plan, Ready (landing zones), Adopt, Govern, Secure, Manage. |
| Landing zone | A pre-configured, governed environment baseline — identity, network, policy, monitoring — that workloads deploy into so every team inherits the same controls. |
| AI agent lifecycle | Ideate → design → build → test → deploy → operate → improve or retire. Strategy must fund and govern all of it, not just the build. |
| Multi-agent solution | An orchestrator agent interprets the goal and delegates to specialised child agents, then composes the result. Split when scope, knowledge, permissions, or ownership genuinely diverge. |
| AI Center of Excellence | The cross-functional body owning standards, reusable patterns, review gates, enablement, and the shared backlog. Establish it before scaling. |
| Prompt library | A curated, versioned, owned collection of approved prompts and instructions, shared so teams do not each reinvent them. |
| Small language model (SLM) | A smaller, cheaper, faster model suited to narrow, well-defined, high-volume tasks — and to latency- or cost-sensitive scenarios. |
| Model router | Routes each request to the most suitable model for that task, balancing quality against cost and latency. |
| Shadow agent | An ungoverned agent built outside the standard path. Causes inconsistent data access, duplicated cost, and compliance exposure. |
| Solution rules and constraints | The explicit boundaries — what the AI component may and may not do, which data it may touch, which actions need approval. |
Design guidance
Extend Microsoft 365 Copilot, or build custom?
- Extend when users already work in Microsoft 365, the need is knowledge or a modest action, and you want to inherit Copilot's orchestration, identity, and grounding.
- Build custom when you need bespoke orchestration logic, a specific or fine-tuned model, a non-Microsoft surface, or behaviour the host cannot express.
- Custom models are a last resort. Prefer grounding, retrieval, and prompt engineering first — they are cheaper, faster, and easier to govern than training.
Portfolio and funding
- Prioritise initiatives by scoring value, feasibility, risk, and strategic alignment — not by who asked first.
- Fund in stages with go/no-go gates after pilot and after scaled rollout, rather than one large upfront request.
- Consider opportunity cost explicitly: capacity spent on a low-value agent is capacity denied to a high-value one.
People and compliance
- Define roles, upskilling paths, and a maker-enablement plan supported by pro-dev capability. Skills do not emerge on their own.
- Invest in user prompt training — adoption failures are frequently prompt-literacy failures, not product failures.
- Evaluate regional and local AI data regulation early: residency, cross-border movement, sector rules, and how they constrain architecture.
Good practice vs. anti-pattern
Do
- Standardise on a common platform and governance baseline.
- Establish the Center of Excellence before proliferation.
- Version and own prompts like code.
- Right-size the model per task; consider SLMs and a model router.
Avoid
- Let each business unit pick its own stack and controls.
- Jump to fine-tuning before exhausting grounding and prompting.
- Treat strategy as a one-off document rather than a governed portfolio.
- Ignore residency rules until deployment.
Exam traps to watch for
TrapStrategy is the CAF stage about motivations, outcomes, and business justification. Do not confuse it with Plan, Govern, or Manage.
Trap"Always build" and "always buy" are both wrong. The expected answer is *build for differentiation, buy for commodity, decided case by case on total cost and time to value*.
TrapA landing zone is an environment baseline, not a physical place, a Dataverse table, or a project phase.
TrapWhen a question describes multiple teams each creating overlapping agents, the risk being tested is ungoverned proliferation — inconsistent access, duplicated cost, compliance exposure.
Readiness checklist
You are ready to move on when you can do all of these without notes.
- Name the CAF stage that owns business justification, and what the others own.
- Give three concrete triggers for splitting one agent into a multi-agent design.
- State when to extend Microsoft 365 Copilot and when to build custom.
- Explain when a small language model beats a frontier model.
- List four elements an AI Center of Excellence must provide.