1.2 Select a Generative AI Solution for a Business Need
Key Takeaways
- Run a decision path before a product name: outcome, data sensitivity, latency, cost, grounding need, and human review.
- Copilot-class productivity AI is the first fit when the work already lives in Microsoft 365 and the user is allowed to see the files.
- Microsoft Foundry and Foundry Tools are the path for a model catalog, custom generative apps, agents, or prebuilt vision, speech, language, and search APIs.
- Skip generative AI when the need is a stable numeric forecast, a closed-set class, or a hard rule with no language output.
- Build versus buy is only a teaser here: adopt Copilot when the packaged experience is enough; Foundry and Copilot extensibility are the later build-or-extend path.
1.2 Select a Generative AI Solution for a Business Need
Quick Answer: Start with the outcome, then test data sensitivity, latency, cost, grounding, and human review. Use Copilot-class productivity AI when the work already lives in Microsoft 365. Use Microsoft Foundry or Foundry Tools when you need a catalogued model, a prebuilt API, or a custom generative app. Skip generative AI when the real need is a label, a score, or a numeric forecast.
Microsoft's skill bullet is: select a generative AI solution to meet a business need. AB-731 candidates do this as leaders, not as developers. You will not configure endpoints. You will decide whether a packaged Copilot experience, a Foundry-hosted model or API, or a non-generative system is the first fit.
A decision path leaders can run in one meeting
Do not start with a product name. Start with six checks.
- Clarify the outcome. What artifact or decision must exist on Friday? A summary, a draft email, answers over policy PDFs, an image concept, a code suggestion, a class label, or a forecast number?
- Data sensitivity. Is the source web-only, Microsoft 365 content the user already can open, or a restricted system of record such as payroll, health, or unpublished financials? Sensitivity drives grounding, logging, and whether a consumer chatbot is even allowed.
- Latency. Does a person wait in a meeting for a draft, or can a batch job run overnight? Chat in Word is interactive. A custom evaluation pipeline is not.
- Cost. Generative usage is often token-driven and license-driven. Do not invent a US dollar fee. Compare shapes: per-user Copilot licensing versus consumption for Foundry Models and Foundry Tools. Later chapters cover Copilot license types and Foundry subscription models.
- Grounding need. Must answers cite approved files, or is a creative first draft enough? Grounding is how you keep the model on your facts. Retrieval-augmented generation (RAG) is the usual pattern; a later chapter treats RAG in depth.
- Human review. Who owns the send button? HR offer letters, legal notices, and customer credits need a named reviewer. Meeting notes for a private team may need a lighter check.
If step 1 produces a label, score, or forecast, stop. That is not a generative AI solution. Return to discriminative or predictive ML from the previous section.
Copilot-class productivity AI versus Foundry versus not generative
| First-fit family | Use when | Examples | Leader watch-outs |
|---|---|---|---|
| Copilot-class productivity AI | Work already happens in Microsoft 365 apps and the user is allowed to see the files | Draft in Word or Outlook, summarize a Teams meeting, Q&A over Graph-permitted work data with the appropriate Copilot license | License and grounding differ across Copilot Chat and Microsoft 365 Copilot experiences; web-grounded chat is not the same as work-grounded Copilot |
| Microsoft Foundry | You need a model catalog, evaluation, fine-tuning, agents, or a custom generative app with enterprise controls | Departmental assistant over a private corpus, an agent that calls tools, compare models before scale | Foundry is the platform; product mechanics belong in the Foundry chapter |
| Foundry Tools | You need a prebuilt API for vision, speech, language, document intelligence, search, or content safety | Read invoices, transcribe audio, search a knowledge index | Many of these APIs extract or search rather than write a paragraph; pair them with generative models when you need both |
| Not generative AI | The outcome is a closed-set class, a numeric forecast, or a hard rule | Demand forecast, fraud score, invoice code | Using an LLM here adds fabrications without improving the metric |
Microsoft Learn's AI platform guidance matches this split. Microsoft Foundry is the starting platform for generative applications and agents. Foundry Tools are pretrained APIs for vision, speech, language, search, and safety. Custom classical ML still belongs on Azure Machine Learning. AB-731 does not ask you to design that architecture. It asks you to pick the family that matches the business need.
Microsoft Copilot experiences are powered by LLMs, grounding in web and/or organizational data (Microsoft Graph and Work IQ), and access scoped by user permissions. That is why Copilot is the default for help my people write and summarize inside Microsoft 365. It is not the default for score every warehouse SKU for next quarter.
Concrete mappings
Map the need before you map the brand.
- Summarization. Long thread, meeting, or policy → Copilot in Outlook, Teams, or Word when the content is already in Microsoft 365. A Foundry app when the corpus lives in a line-of-business store you will retrieve from with Azure AI Search or similar.
- Drafting. Customer email, board narrative, job description → Copilot-class drafting with a human owner. Foundry when you need a shared departmental assistant with your own instructions and tools.
- Q&A over documents. Employees asking what our parental leave policy is → Copilot over permitted work files, or a grounded Foundry assistant if the corpus is curated and must cite a controlled set. Do not point a public chatbot at unpublished HR files.
- Image generation. Marketing concepts for human review → a generative image model, often reached through Foundry's model catalog. That is not a vision classifier. Azure Vision in Foundry Tools analyzes images; it does not replace a campaign art director.
- Code assist. Developers and analysts use coding assistants to suggest code or formulas. Mention it as a real use case. You still do not need to code to pass AB-731. A leader's job is to decide whether that assist is allowed on which repos, with which data, and with what review.
Build versus buy (teaser only)
If Copilot already drafts the email, adopt the packaged experience and govern it. That is the buy path in everyday language. If you need a custom agent, a specific model from the catalog, or APIs wired into a line-of-business system, you are on a build or extend path inside Microsoft Foundry and the Microsoft 365 Copilot extensibility framework. A later chapter covers when to build, buy, or extend. Do not over-commit here: most items at this bullet are which family fits, not write a make-versus-buy memo.
Worked scenarios for CFO, operations, HR, legal, and security
Legal. Counsel wants faster first-pass review of vendor paper. Outcome: a summary and issue list, not a binding opinion. Sensitivity: confidential contracts. Grounding: the uploaded agreement, not the open web. Human review: a lawyer. First fit: Copilot in Word on the file the lawyer opened, or a Foundry assistant restricted to a matter workspace. Not fit: a consumer chatbot, or a classifier that only labels NDA versus MSA.
Security. The chief information security officer (CISO) wants fewer phishing clicks. If the need is to detect a malicious message, that is discriminative detection, possibly plus Content Safety in the Foundry ecosystem. If the need is to draft a teachable simulation email and a staff advisory, that is generative, with communications review. Mixing those two needs into one AI project is how budgets stall.
HR. Policy Q&A for managers: Copilot or a grounded assistant, with HR review for high-stakes cases such as termination or accommodations. Résumé routing into stages: classification, not generation.
Operations vice president. Warehouse photos that must count damaged cartons: Azure Vision in Foundry Tools (analyze), not an image generator. Night-shift handoff narrative: generative draft from the shift log, human-confirmed.
CFO. Board narrative from an approved forecast: generative. The forecast itself: predictive ML, not an LLM.
How to choose on exam stems
Read the outcome first. If the company needs new language or media, stay in generative families and then choose Copilot versus Foundry or Foundry Tools by where the work and data already live. If the company needs a label or a number, the correct selection is often not generative AI. That not answer is in scope for this bullet.
Cost and latency still matter after the family is chosen. A Copilot draft in a live customer call needs interactive speed and a human mute button. A weekly research brief can tolerate a slower Foundry agent that retrieves more sources. Do not skip human review because the demo looked fluent. Fluency is not permission to send.
Official Microsoft pages to reopen
- Study guide for Exam AB-731
- What is Microsoft Copilot? — LLMs, Graph and Work IQ grounding, permission-scoped access
- What is Microsoft Foundry?
- What are Foundry Tools?
- AI and Machine Learning Products — Foundry for generative apps and agents; Foundry Tools for prebuilt APIs
A knowledge worker needs to summarize Word files they already can open and draft a reply in Outlook. What is the first-fit generative solution?
Which factor most strongly argues against using generative AI at all?
HR needs question-and-answer over policy documents, with answers limited to those files and a people partner reviewing high-stakes replies. Which requirement is most important to capture in the decision path?