9.3 Match an AI Model to a Business Need

9.3 Match an AI Model to a Business Need

Quick Answer: Start with capability (text, vision, reasoning, embeddings), then score cost, latency, data residency, grounding, and evaluation. Do not train a foundation model from scratch for AB-731. Microsoft Foundry is where you compare catalog models, run evaluations, and deploy. Grounding private content is a Search-plus-model design, not a bigger pretrained checkpoint.

Microsoft's July 22, 2026 skills list asks you to match an AI model to a business need. The audience is still a business decision-maker. You sponsor a bake-off. You do not write a training loop. Microsoft Foundry Models is documented as the destination for discovering, evaluating, and deploying models—whether you are building a custom copilot, an agent, or enhancing an existing application.

Capability class before brand name

Name the job in plain language, then pick a class. Brand names in the catalog change monthly; Microsoft currently says the catalog includes over 10,000 models, with about 50 new models published each month.

Worked examples:

• HR needs semantic search over job descriptions. Start with an embeddings model and Azure AI Search, not a 70-billion-parameter chat model that rereads every description on each keystroke.
• Operations needs defect tags on still photos. Start with Azure Vision in Foundry Tools or a vision-capable catalog model, not a speech model.
• Finance needs a board narrative from an approved forecast. A text model in Copilot or Foundry can draft; the forecast numbers still come from the spreadsheet, not from the model's prior.
• Legal needs cited answers from a matter index. Pair a text model with grounding through Azure AI Search. A reasoning model without retrieval will still fabricate a case citation.

Cost, latency, residency, grounding, evaluation

After the class is right, score the operating constraints. Microsoft's Foundry Models overview and manage-costs article give leaders a vocabulary without turning this exam into a pricing spreadsheet.

Cost. Language and vision models bill on tokens (and related meters that vary by model and deployment type). Microsoft documents pay-as-you-go (serverless API) usage and commitment tiers as billing approaches for Foundry resources. Fine-tuned Azure OpenAI models are charged three ways: training, hosting (hourly even if unused), and inference. Hosting while idle is the trap the CFO should hear in the first steering meeting. Partner and community models are often Marketplace-billed; Azure Prepayment (monetary commitment) can apply to models sold by Azure, not to those Marketplace meters. Full Foundry Tools subscription mechanics belong in chapter 13. Here, know that cheaper-per-token is not cheaper-per-outcome if you must call a reasoning model five times.

Latency. Interactive agents feel slow above a couple of seconds. Batch extraction can wait. Global deployments may be faster and cheaper; they may also process data outside a single geography. Microsoft documents serverless deployment types that include global, data zone, and regional processing so you can trade latency against residency.

Data residency. If the CISO or a regulator requires prompts and completions to stay in a named geography, filter the catalog by region and choose a regional or data-zone deployment rather than a global one. Microsoft also documents that pay-per-token offers depend on the billing account country and on project resources in a region where the model is available. Confirm availability before you promise a country launch.

Grounding. A chat model without retrieval answers from training data and the prompt. A grounded solution retrieves from your corpus first. That is Azure AI Search (classic index or knowledge base) plus a generator, not a request to “train on all of SharePoint.” Microsoft documents RAG with models deployed as serverless deployments and vector indexes in Foundry. If the need is “answers must cite the policy PDF,” grounding is a must-have, not a nice-to-have.

Evaluation. Foundry's catalog supports side-by-side comparison, leaderboards, and benchmark tabs on model cards. Foundry also evaluates agents on dimensions such as groundedness, task adherence, and jailbreak exposure. A leader's bake-off uses a fixed set of real tasks, a named pass bar, and Content Safety on both prompt and output. Microsoft is explicit: customers are responsible for selecting an appropriate model, reviewing model cards, and implementing measures including Azure AI Content Safety.

Foundry as the compare-and-deploy surface

Microsoft organizes the catalog in two buckets you should be able to explain to finance.

Deployment options the catalog exposes: managed compute (model weights on dedicated virtual machines, billed on core hours) versus serverless (API access to a Microsoft-hosted model, billed on inputs and outputs, typically tokens). Instant access (preview) lets some models be called by name without a pre-created deployment. Managed compute for many open models still uses hub-based projects in Foundry (classic). New investment is on Foundry projects. For AB-731, remember the business implication: dedicated GPUs are a capacity decision; serverless is a consumption decision.

You do not train from scratch on this exam. Fine-tuning exists for certain models when you have labeled examples and a measured gap. Fine-tuning is not the default, and hosting charges continue until you delete the deployment. Pretrained catalog models plus prompting, grounding, and evaluation cover most leader scenarios Microsoft is testing.

Worked matching by role

CFO. “Summarize variance commentary” is a text model with a small context window and Copilot or Foundry, grounded on the approved workbook. “Cluster 200,000 vendor comments” is embeddings plus Search. “Train our own large model on ten years of ledgers” is out of scope for AB-731 and usually out of budget.

Operations vice president. Photo defects: Vision API or a vision catalog model. Radio transcripts: Speech to text, then a text model if you need a summary. Do not buy a reasoning model to count boxes.

HR. Policy Q&A: embeddings plus Search plus a modest chat model, with PII detection on uploads. Image-generation models for recruiting posters are a brand exercise, not the policy engine.

Legal and security. Residency and Content Safety outrank a one-point leaderboard win. If a model cannot deploy in the required region, it is not a candidate. If embeddings skip content filtering (Microsoft notes filtering is not available for certain embedding and time-series serverless models), your grounded chat still needs Safety on the generator.

Official Microsoft pages to reopen

• Microsoft Foundry Models overview — catalog buckets, filters, deployment options, customer responsibilities
• What is Microsoft Foundry? — agents, models, tools
• Plan and manage costs for Microsoft Foundry — tokens, pay-as-you-go, commitment tiers, fine-tune hosting
• Introduction to Azure AI Search — grounding and RAG