Your AI agents are burning through tokens and budget because they're relearning your business context on every single task. Frontier Tuning changes that equation. Instead of injecting thousands of context tokens at runtime, you teach the model your workflows once, and it learns from your organization's data. Here's how this new approach to post-training, announced at Build 2026, can make your agentic systems both smarter and cheaper to run.
Why the guidance is shifting: The economics have changed
Traditional guidance has previously steered AI practitioners to reserve fine-tuning for very specific domain workflows. Before jumping straight to fine-tuning a model, the best practice was to first exhausting prompt engineering, RAG approaches, and context engineering. Historically, fine-tuning's training and hosting costs were more expensive than they were worth, and fine-tuning was often overkill.
However, for agentic workflows, generic models don't understand a customer's terminology or task sequences, so RAG injects the context without teaching behavior every time at runtime. For complex workflows that require thousands of tokens per task, injecting context at runtime adds up quickly.. Now, the economics have inverted, so by tuning models to be task, organization, and line-of-business specific, performance and efficiency improve. Fine-tuning now can be a cost-reduction lever depending on token consumption. When your models are Frontier Tuned with your data, their performance naturally enables less token consumption through more optimal task solving.
In summary:
- Before: Fine-tuning training and inference costs were usually not justified when RAG and context engineering could provide strong results
- Now: For high-volume agentic workflows, tuning costs are often cheaper than repeated context injection
- Result: Frontier tuning becomes both a cost-reduction lever and a performance add
Reinforcement fine-tuning vs. classic fine-tuning
There is now massive value in learning process and workflows in our agentic systems. Reinforcement learning is based on a rewards system, and when we apply Reinforcement Fine-Tuning (RFT), we reward model behavior for landing the correct sequence in an agentic workflow (think tool calls within traces). Historically, Reinforcement Learning excels at sequence and process-based scenarios. For example, think game-playing with complex games like GO or Chess. The system understands the current state and determines the next most optimal move based on that state. Now extend this concept to agentic workflows, where the agent has the current context (or state) and determines the next most optimal move (tool call, action, etc.) based on it's context. This is a shift from our former method to apply Supervised Fine-Tuning (SFT) with supervised data (i.e. ground truth labeled data of inputs and good outputs). With classic fine-tuning, we are only enabling the model to understand strong token outputs given a user prompt based on that labeled data, without considering an optimal sequence of actions that RFT enables our models to learn.
| Supervised Fine-Tuning (SFT) | Reinforcement Fine-Tuning (RFT) | |
|---|---|---|
| Learns from | Labeled input/output pairs | Traces of actions and outcomes |
| Optimizes for | Correct outputs | Optimal action sequences |
| Best for | Single-turn Q&A, classification | Multi-step agentic workflows |
| Analogy | Determining correct answers | Learning to play chess |
Example: Customer Service Agent Optimization
Consider a customer service agent that handles refund requests. Without tuning, each interaction requires injecting your refund policy (500 tokens), product catalog context (1,200 tokens), and customer history patterns (800 tokens) - roughly 2,500 tokens per request. At 10,000 daily requests, that's 25M context tokens monthly.
With Frontier Tuning, the model learns your refund workflows from historical traces. It knows when to check inventory, when to escalate, and how to apply your specific policies. Context injection drops to ~400 tokens per request for customer-specific details only, which is an 84% reduction in token consumption.
When to Frontier Tune?
Consider Frontier Tuning when:
- You have workflow data: Traces of tool calls, decisions, and evaluation signals from your agentic systems
- You can define success: Clear evaluation rubrics for what optimal task completion looks like
- The economics make sense: Your token consumption is high enough that tuning costs less than repeated inference
Get Started with Frontier Tuning
Ready to explore Frontier Tuning for your agentic workflows?
- Explore the fundamentals: Microsoft Frontier Tuning - understand the technical capabilities and how it works
- Nominate your use case: Request early access - tell us about your agentic workflow and nominate it for the preview
- Check out the MAI Blog: Microsoft AI Blog - get the latest news, stories, and perspectives from Microsoft AI
- Experiment with the MAI models: Microsoft AI Model Playground - explore the models without a subscription
- Not sure if Frontier Tuning is right for you? Start by instrumenting your agentic workflows to capture traces. You'll need this data regardless, and it will help you assess whether the economics make sense for your scenario.
Authored by the Americas Markets and Industries Office of the CTO