Copilot Tuning paused: what applies now and why waiting is the plan

Copilot Tuning through Agent Builder templates is paused. Microsoft is rebuilding the entire tuning architecture and moving it to Copilot Studio. If you are planning a tuning project right now, take a close look before you promise dates to customers or management.

This is one of those moments where „Hold on, let me check the Message Center again” is the most honest sentence a consultant can say. So: what does MC1454393 really say, and what is just hope?

What does „tuning” actually mean?

Tuning literally means fine-adjusting, and that is exactly the point. A normal Copilot agent is like a good barista with a recipe card: for every question, it looks things up in your knowledge sources and builds an answer from them. Copilot Tuning goes one step further. The barista learns your house blend so well that it does not need to look it up anymore.

Microsoft puts it like this: Copilot Tuning helps agents understand and follow an organization’s own terminology, workflows and business processes. Microsoft staggers this into levels and explicitly recommends starting from the bottom — not every agent needs all of them:

• Level 0 – Start without tuning: First work with a tunable template without tuning. If the result is good enough, you are done.
• Level 1 – Tune context: The agent learns which information matters for its task.
• Level 2 – Tune tools: The agent learns to use its tools better.
• Level 3 – Tune model: The language model itself is retrained with your data, classic fine-tuning. This level comes at the very end.

And one more myth that keeps going around: tuning is not „data in, done”. You define goals, data and evaluation criteria, and Copilot Tuning gives you evaluation results based on them. It is an iterative process and often needs several rounds. And even after that: the results stay drafts that a human checks for accuracy.

What exactly was paused

Until now, Copilot Tuning ran through tunable templates in Agent Builder. Since 20 August 2026, the Tune option is no longer available for five templates:

• Document Writing
• Document Summary
• Style Editor
• Document Validation
• Expert Question Answer

The Optimization template is being retired completely. And a detail that hurts if it affects you: Microsoft discarded all tuning runs that were not finished by 20 August.

Microsoft deliberately talks about a pause, not an end. Tuning comes back, just in a different place and in a different form.

What keeps running

The most important thing first, because I know some admins’ pulse is rising right now: existing agents created from these templates keep working. Agents with fine-tuned models keep running with exactly these models. Nobody pulls productive workloads out from under you.

What	Status since 20 August 2026
Existing agents from the templates	Keep running
Agents with fine-tuned models	Keep running with these models
New agents from the templates	Possible, but without tuning
New tuning via the Tune option	Paused
Unfinished tuning runs	Discarded
Optimization template	Being retired

You can still create new agents from the templates. They just do not get tuning and run as completely normal declarative agents until the new tuning experience is available. To keep existing agents running, you do not have to do anything.

The hurdle hardly anyone talks about

Before we talk about the new architecture, an honest look at the target audience.

For most mid-sized companies in the DACH region, this means: Copilot Tuning has simply been out of reach so far. Microsoft has not yet communicated whether the new preview in Copilot Studio comes with the same threshold. For many, that is the really exciting question.

The new architecture: skill-based in Copilot Studio

The new tuning no longer runs through templates in Agent Builder, but in Copilot Studio. Microsoft describes it as a skill-based architecture with enhanced tuning capabilities for business scenarios. There are no more official details so far.

My expectation, and I deliberately label it as an expectation: skills will probably be more granular than templates. Instead of shaping an agent to one context as a whole, you could adjust individual capabilities more precisely. For everybody who has ever tried to teach a tuned agent one single new habit afterwards, that would be real progress. Whether it turns out that way, only the preview will show.

The timeline according to Microsoft: public preview from September 2026, general availability from December 2026. So between now and then there is a preview that is exactly what the name says. Publicly testable, but not production-ready in the sense a mid-sized company means when it hears „production-ready”.

What this means for your project

If you are planning a tuning project right now, one simple rule applies: do not build on preview sand if you want to stand on GA foundation. In practice this means:

• Define requirements now: which tasks should the agent do better than a normal declarative agent?
• Prepare content: which SharePoint content is clean, current and suitable for tuning?
• Plan go-live realistically: at the earliest after general availability from December, rather later.
• Clarify the license question early: as soon as Microsoft names the requirements for the new preview.

For existing tuned agents: watch, do not touch. In the Agent 365 portal you see all tuned agents of your tenant in one place. There is no automatic migration. If you want to move to the new tuning later, you rebuild the agent. Whether you can reuse already trained models is something Microsoft still wants to explain.

If you have a tuning project on hold yourself, or already gathered first experiences with the preview, tell me on LinkedIn. I am curious how other companies communicate this transition phase to their customers.