Study guide

1.3 Pretrained Versus Fine-Tuned Models

Microsoft Certified: AI Transformation Leader Exam AB-731

Track your progress as you read

Finished sections get a check mark.

1 Introduction & Exam Overview

0.1 Current AB-731 Exam Facts 0.2 Audience Profile and the July 22, 2026 Blueprint 0.3 Independent Study Path for AB-731

2 Generative AI Foundations

1.1 Generative AI Versus Other Types of AI 1.2 Select a Generative AI Solution for a Business Need 1.3 Pretrained Versus Fine-Tuned Models

3 Cost, Risks, and Business Value

2.1 Tokens and ROI Cost Drivers 2.2 Fabrications, Reliability, and Bias 2.3 Scalability, Automation, and When Generative AI Adds Value

4 Prompt Engineering

3.1 Impact of Prompt Engineering 3.2 Prompt Engineering Techniques

5 Grounding, RAG, and Data

4.1 Business Requirements for Grounding Solutions 4.2 Retrieval-Augmented Generation (RAG) 4.3 Data Type, Quality, and Representative Datasets

6 Secure AI and Machine Learning

5.1 Importance of Secure AI 5.2 Application, Data, and Authentication Security for AI 5.3 When Machine Learning Adds Value 5.4 Lifecycle of a Machine Learning Solution

7 Microsoft 365 Copilot and Microsoft Copilot

6.1 Map Business Processes and Use Cases to Copilot 6.2 Differences in Capabilities Between Versions of Copilot 6.3 Microsoft 365 Copilot Chat on Web and Mobile 6.4 Copilot in Microsoft 365 Apps

8 Copilot Studio, Microsoft Graph, and Integrated AI

7.1 Microsoft Copilot Studio Capabilities 7.2 Microsoft Graph Capabilities 7.3 Integrated Microsoft AI, Risk Mitigation, and Safety

9 Agents, Extensibility, and Stack Mapping

8.1 Map Processes to Microsoft AI Apps and Services 8.2 When to Use Researcher or Analyst in Copilot 8.3 Build, Buy, or Extend Copilot

10 Microsoft Foundry and Foundry Tools

9.1 Map Business Processes to Foundry Tools 9.2 Azure Vision, Azure AI Search, and Microsoft Foundry 9.3 Match an AI Model to a Business Need 9.4 Scalability and Security Benefits of Foundry

11 Responsible AI and Governance

10.1 Importance of Responsible AI 10.2 Governance Principles for AI Use 10.3 Establish an AI Council 10.4 Microsoft Responsible AI Standards

12 Adoption Teams, Barriers, and Champions

11.1 Establish an Adoption Team 11.2 Common Barriers to AI Adoption 11.3 Establish an AI Champions Program

13 Impacts, Copilot Licensing, and Foundry Cost Models

12.1 Impacts to Data, Security, Privacy, and Cost 12.2 Copilot License Types 12.3 Foundry Tools Subscription Models

1.3 Pretrained Versus Fine-Tuned Models

Key Takeaways Explain This Topic AI

A pretrained foundation model already learned general patterns from large-scale data; teams consume it with prompts before any custom training.

Fine-tuning continues training on labeled examples so style, format, or a narrow task becomes more consistent without training from scratch.

Prompt engineering plus retrieval-augmented generation is usually enough when the gap is missing or changing facts.

Fine-tuning adds training and hosting cost, dataset governance, and a refresh cycle; do not invent a Microsoft SKU price.

Organizations catalog, evaluate, and deploy these models in Microsoft Foundry; deeper product mechanics belong in the Foundry chapter.

Last updated: September 2026

1.3 Pretrained Versus Fine-Tuned Models

Quick Answer: A pretrained foundation model already learned general patterns from large-scale data; you use it with prompts. Fine-tuning continues training on your labeled examples so style, format, or a narrow task becomes more consistent. Prompting plus retrieval-augmented generation (RAG) is usually enough when the gap is missing facts. Fine-tune when the gap is missing skill, voice, or schema after prompting has grown painful.

Microsoft's skill bullet is: describe the differences between AI models, including fine-tuned and pretrained models. AB-731 candidates describe this as buyers and governors. You will not run a training job. You will decide whether use the catalog model as-is, ground it on our files, or sponsor a fine-tune is the proportionate next step.

Pretrained foundation models

A pretrained model (often a foundation model ) has already been trained on a large, diverse dataset before your company ever shows up. You consume it through prompts. Microsoft Foundry is where organizations discover, compare, and deploy these models. Microsoft documents a catalog of thousands of models from Microsoft, Azure OpenAI, Anthropic, Meta, Mistral, Cohere, Hugging Face, and others. Capabilities range from chat completions to embeddings, vision, and speech. You do not memorize the catalog. You remember that pretrained means general capability, no extra training on our examples yet.

Leaders should still read the model card: task type, language, deployment option, lifecycle (preview versus generally available), and the provider's terms. Microsoft notes that customers remain responsible for lawful use, for choosing an appropriate model, and for safety measures such as Content Safety. Models sold by Azure carry Microsoft support and product terms. Partner and community models are typically supported by their providers. That distinction matters for procurement, not for coding.

In Microsoft 365, Copilot experiences also call pretrained LLMs. Users do not pick weights. They pick a Copilot experience whose grounding and license match the work. In Foundry, a team can choose a named model for a departmental assistant. Same idea: start from a pretrained base.

Foundry Tools APIs are a related but different pretrained story. A vision or speech API is a pretrained capability you call for a specific task, often without a generative paragraph at all. Do not confuse consuming a pretrained LLM in Copilot or Foundry with calling a Foundry Tools classifier. Both are pretrained. Only one is typically generating new language.

What fine-tuning actually changes

Fine-tuning takes a pretrained model and continues training on a smaller, task-specific dataset so the weights shift toward your examples. Microsoft Foundry documents this as adapting a pretrained language model to a task or dataset without training from scratch. Teams often work with hundreds to thousands of labeled prompt-response pairs rather than millions of records.

Microsoft's fine-tuning guidance lists leader-relevant reasons to consider it:

  • Shorter prompts, because examples move out of the request and into the model
  • More consistent style and tone, such as brand voice or support-desk voice
  • More consistent formats or schemas, such as the same report sections every time
  • Better tool-calling patterns when listing many tools in a prompt becomes noisy
  • Sometimes, a smaller fine-tuned model that is cheaper to run than a large general model on a narrow task

Foundry also describes technique families you should recognize by name, not implement: supervised fine-tuning (SFT) on labeled pairs; reinforcement fine-tuning (RFT) when many solutions exist and a grader rewards better reasoning; direct preference optimization (DPO) when you supply preferred versus non-preferred replies. Details stay with the technical team. Your job is to know that fine-tuning is extra training with extra governance, not a magic accuracy switch.

Fine-tuning does not reliably teach the model new facts that change every week. If the employee handbook changes on Monday, a model fine-tuned on last quarter's handbook can still speak with perfect confidence and last quarter's policy. Changing facts belong in retrieval.

Prompt plus RAG versus fine-tune

Approach What you change Best when Leader cost and governance

Prompt only Instructions and examples in the request Tone is good enough, facts are in the prompt, volume is modest Fastest to try; long prompts raise tokens and latency

Prompt plus RAG Prompt plus retrieved passages from approved sources The gap is knowledge — policies, tickets, contracts — that must stay current You govern the corpus, permissions, and citations; you do not host a custom model

Fine-tune Model weights, using labeled examples The gap is behavior: voice, schema, or a narrow skill after prompting is painful Training and hosting cost, dataset quality, bias checks, and a refresh when the base model or policy changes

Train from scratch Entire model Almost never the AB-731 answer for a business leader Out of proportion for this exam's audience

RAG pulls relevant snippets at run time so the pretrained model can condition on your files. A later chapter covers RAG. Here, remember the split: RAG supplies facts ; fine-tuning supplies habits .

Microsoft also notes that fine-tuning and retrieval can work together: the model learns how to use retrieved context and ignore junk. That is an advanced pattern, not the first experiment.

  • When prompt plus RAG is enough
  • HR Q&A over the current policy set
  • Legal Q&A over a matter's documents
  • Copilot answering from Word, email, or Graph-permitted work data
  • Any corpus that changes faster than you would retrain
  • When to sponsor a fine-tune
  • Support replies must match a strict template after few-shot prompts became huge
  • Brand voice is inconsistent across thousands of generations per day
  • You have a clean, representative set of example pairs and owners who will relabel when the process changes
  • A technical team has already shown that prompting and RAG plateau on the quality bar
  • When to refuse both generative customization paths

If you needed a closed-set invoice code, you never needed a generative model. Fine-tuning an LLM to mimic a classifier is usually the expensive way to rebuild discriminative AI.

Cost, data, and governance for a business leader

Do not memorize a Microsoft SKU price. Prices vary by model, deployment option, region, and commitment, and Microsoft publishes them on pricing pages that change. What you must be able to explain:

Pretrained, hosted: you pay for usage (often tokens or transactions) and, for Copilot, for the relevant license. You govern prompts, data access, and review.

RAG: you also pay to index and retrieve. You govern which sources are in the index and who may query them.

Fine-tune: you pay to train and to host a custom model. Microsoft's fine-tuning considerations warn that you need high-quality, representative data; that poor data produces overfit or biased behavior; that you may need to fine-tune again when data or the base model changes; and that experimentation is normal. Import and storage rules can collide with a locked-down enterprise network — that is a security architecture issue, not a reason to skip review.

Data governance questions a leader should ask before signing a fine-tune:

Who labeled the examples, and do they include rare but high-risk cases?

Are personal data, secrets, or privileged legal text in the training file?

How will we evaluate fairness, privacy, and safety before production?

Who retires the custom model when a new base model ships?

Microsoft Foundry is where organizations catalog, evaluate, deploy, and — when supported — fine-tune models, with observability and content safety in the platform. Deep product mechanics such as deployment SKUs, classic versus new portal, and agent APIs belong in the Foundry chapter. For this bullet, remember the destination: do not leave custom models as unmanaged experiments on a laptop.

Worked scenario

A customer-support director wants every chat reply in a five-section template. Prompting with five examples works in a pilot and then blows up token use in production. Facts such as order status and return window still come from systems of record, so RAG stays. The director sponsors an SFT job on approved reply pairs so the template is habitual, keeps RAG for live order data, and requires a supervisor sample-review. That is pretrained plus RAG plus fine-tune, each doing one job. The CFO still sees usage cost as tokens plus training and hosting — not as a single invented list price.

Legal and security should sit on the same review. Fine-tuning on unredacted case files can freeze privileged language into weights. RAG with permission trimming is often the safer way to keep those facts available without copying them into a custom model.

  • Official Microsoft pages to reopen
  • Microsoft Foundry Models overview — catalog, providers, evaluation, and deployment options

Microsoft Foundry fine-tuning considerations — when to fine-tune, data quality, and extra training and hosting cost

What is Microsoft Foundry?

Study guide for Exam AB-731

Loading diagram... Choose prompt, RAG, or fine-tune from the quality gap

Illustrative relative effort for a business leader (not Microsoft pricing)