Microsoft MAI Models in Copilot: What Changes for Your Business

Updated on October 6, 2026

Microsoft is quietly removing OpenAI and Anthropic from Copilot and replacing them with its own in-house models. If your company uses Excel or Outlook with AI, the engine under the hood has already changed, and no one told you.

Is Copilot still worth relying on after the model change?

It depends on what you use it for. For structured tasks like Excel formulas and email summaries, MAI models hold up well. For high-reasoning work, such as complex analysis, persuasive copy and strategic briefs, the answer is still open.

What really happened: did Microsoft change Copilot's engine?

Yes. Microsoft has started a quiet transition: AI models from OpenAI and Anthropic are being gradually replaced with proprietary MAI (Microsoft AI) models inside products like Excel and Outlook.

According to The Decoder, tens of thousands of weekly queries already run through the new in-house models. The move wasn't announced to end users with a dedicated communication.

Mustafa Suleyman, CEO of Microsoft AI, told Bloomberg explicitly: "Anthropic is extremely expensive and I think many people are urgently looking for alternatives. We pay a lot of money to Anthropic, so our goal is to reduce and ultimately eliminate that cost." This is reported by SiliconANGLE.

The direction is clear: reduce reliance on third-party providers and lower the cost of every single AI interaction. For companies paying for a Copilot subscription, this means the model processing their requests can change without notice.

Are MAI models really on par with GPT-4o and Claude?

On structured tasks, yes. On complex reasoning, it's still too early to say for sure. MAI models are designed for speed and cost, not depth of processing.

At Build 2026, Microsoft unveiled 7 MAI models. The most advanced is MAI-Thinking-1: 35 billion active parameters, a 256,000-token context window (meaning it can process very long documents in one go), and trained entirely on commercially licensed data, without using outputs from OpenAI or other third-party models. Windows Central confirms it.

OpenAI models, particularly GPT-4o, and Anthropic models like Claude 3 were chosen in the past for their complex reasoning and natural language handling. MAI models seem to prioritize speed and cost per query over depth of processing.

This isn't a neutral detail. If you use Copilot for cognitively demanding work, such as analyzing qualitative data, generating persuasive copy or summarizing complex documents, you may notice differences in the outputs.

How many companies already use Copilot, and how widespread is it in Italy?

Copilot passed 30 million paid licenses in July 2026. Growth in Italy has been fast, but actual adoption among license holders remains low.

According to Industria Italiana, Microsoft 365 Copilot passed 30 million paid licenses in July 2026, with new licenses more than doubling quarter over quarter. The number of companies with over 50,000 licenses grew more than 7x year over year.

In Italy, Industria Italiana again reports a 70% increase in customer companies and daily users doubling quarter over quarter. Globally, 70% of Fortune 500 companies have already integrated the technology into their processes.

But the figure that should worry you most is another one: only 35.8% of employees with access to a license actively use Copilot, according to Avantiico. Fewer than 4 in 10 workers with a license really use it.

What changes in practice for people who use Copilot every day?

If you've built Copilot into your workflows, you may notice changes in output quality, and Microsoft hasn't guaranteed equal performance.

The problem isn't just technical. Companies pay for a Copilot subscription without knowing exactly which model processes their requests. Silently swapping models raises an AI governance issue, meaning control and accountability over how AI tools are used in the company, that IT leaders and marketing managers can't ignore.

If you use Copilot for advanced Copilot features such as agents and coding, a change in the underlying model can alter behaviors that seemed stable.

Three real-world scenarios: how does the model change affect an Italian company?

The impact varies widely depending on how you use Copilot. Here are three typical situations and what they mean in practice.

Illustrative examples: the numbers and situations are meant to show the reasoning, they are not real customer data.

Why did Microsoft make this move, and what's behind it?

It's a financial and strategic choice at once. Microsoft wants to reduce its dependence on third-party providers and lower the marginal cost of every AI interaction, but the reputational risk is real.

Last year, the race to invest in AI put pressure on the margins of all the big players. Microsoft, which has invested billions in OpenAI, now has to balance operating costs against shareholder expectations. Bringing models in-house is a rational move financially: VentureBeat reports claimed savings of up to 89% on API costs thanks to hardware-level optimizations on the Azure infrastructure.

The reputational risk, though, is concrete. Copilot was sold as a tool powered by the best models on the market. If perceived quality drops, companies may reconsider adoption or reduce the seats they buy. Competitors, from Google with Gemini to Salesforce with Einstein AI, could use this moment to position themselves as more transparent alternatives.

It's also worth following how the other players' models evolve: Anthropic is building an independent go-to-market strategy that could make it less dependent on Microsoft as a distribution channel.

Why does the Italian AI market make this news even more urgent?

The Italian AI market is growing fast, but most SMEs are still on the sidelines. If you already have Copilot, you need to understand what you're really using.

According to the Artificial Intelligence Observatory of Politecnico di Milano, the Italian AI market reached €1.8 billion in 2025, up 50% from 2024. 84% of large Italian companies have purchased Generative AI licenses, with Microsoft Copilot among the most widely used solutions.

The picture for SMEs is the opposite. According to the 2025-2026 Research of the Digital Innovation in SMEs Observatory at Politecnico di Milano, 76% of Italian SMEs have not invested in AI and don't plan to, and only 7% have launched structured training programs. If you already have Copilot, often without a clear strategy, you risk paying for a tool you don't use or use poorly.

How do you get started: what to do in the next 30 days?

Before waiting for official word from Microsoft, there are five actions you can take right now with your in-house resources.

What are the most common mistakes companies are making right now?

The biggest one is not noticing the change. The second is having no comparison data to measure it.

• Doing nothing while waiting for official word. Microsoft hasn't announced the change to end users. If you wait for a notification, you may never get one.
• Renewing licenses without an audit. Auto-renewal is convenient, but right now it's a risk: you're paying for a service that may no longer match the one you originally chose.
• Having no comparison baseline. If you haven't saved examples of past outputs, you can't measure whether quality has changed. Start doing it now.
• Treating Copilot as a monolithic tool. Copilot isn't a single product: it's an interface that swaps the underlying model over time. You need to evaluate it the way you'd evaluate any service provider.
• Ignoring internal training. With only 35.8% actual adoption among license holders, the problem often isn't the model: it's that employees don't know how to use it well. Investing in AI training for companies is the first step.

How can you tell if Copilot is still working well for your business?

You need concrete indicators, not impressions. Here is a reference table to track output quality over time.

So, is it worth keeping Copilot, or is it time to consider alternatives?

It's worth it if you use Copilot for structured tasks (formulas, summaries, internal searches), have an enterprise plan with dedicated support, and are willing to audit outputs periodically. In that case, the cost savings Microsoft passes on (at least in part) justify staying the course.

Somewhere in between: you use Copilot for mixed tasks, some structured, some requiring heavy reasoning. Run the benchmark now, before your next renewal. If the results are acceptable, keep going. If not, explore plans with selectable models or hybrid solutions.

Not yet if you've built critical automated workflows on Copilot and haven't audited the outputs yet. In that case the operational risk is high: a silent model change can break logic that seemed stable. Act before your next renewal.

How we see it at SHM

For us, the news about Copilot's model change isn't a crisis: it's a reminder. AI should be integrated to free companies from routine, not to create new opaque dependencies. When the engine behind a subscription changes without notice, the problem isn't technical — it's governance.

We've worked with companies that use complex AI tools — from technical catalog management to international SEO — and the starting point is always the same: understand what you're really using, not just what you bought. That's true for Copilot as for any other tool.

If you're looking at how to integrate AI into your processes in a controlled way, our strategic AI integration service starts with an audit of what you already use. If you want to understand how AI models affect your online visibility, we also work on SEO for AI. And for teams that need to learn to use these tools properly, we offer business process automation training programs.

You can dig deeper into all these topics in our section on AI and process automation, where we collect practical analysis for people who have to make decisions, not just stay up to date.

It's also worth keeping an eye on how the model landscape evolves: OpenAI is developing models for browsers and computer use that could shake things up again, and security incidents involving AI models show that governance isn't optional.

Frequently Asked Questions

The real problem isn't which model Microsoft uses: it's that your company is paying for a tool whose specifications it doesn't control.