MAI: A Work in Progress and its Impact on Partners

When Satya Nadella trumpeted Copilot at Ignite 2024, it sounded like he was announcing Microsoft’s entry into the LLM race with OpenAI and Anthropic. He was even pulling into the lead by introducing us to “The Agentic Age” before any of them.

Two years later, it’s no mystery that Copilot is a harness, connecting LLMs to the tools and resources needed to offer users a convenient interface of a routed selection of models. If you doubted this, it didn’t take long for them to end their exclusive agreement to use ChatGPT as the “back end” and add Claude. Later, they would expand the model routing capabilities to include other models, and I suspect that will continue.

Finally, Microsoft built seven AI models of its own, making them available along with its rivals' models.

Copilot being a harness is not a terrible thing. It’s a harness that is ideally tuned to work most effectively with the entire Microsoft suite, providing solutions focused on almost every platform they offer. Microsoft believes customers will pick the best model for each job, and that the best one will not always be a Microsoft model. Your business should be built on that same belief.

Mustafa Suleyman stopped running consumer Copilot in March to focus entirely on models. The result, shown at Build in June, is seven MAI models, most of them specialized.

The main general-purpose model is MAI-Thinking-1, a medium-sized model that competes in its weight class. Microsoft's headline comparisons were against Claude Sonnet 4.6 and Opus 4.6, the previous generation. Copilot already offers Anthropic's Claude Opus 5 and OpenAI's GPT-5.5.

Unlike OpenAI, Anthropic, Gemini, and SpaceXSI (previously SpaceXAI), Microsoft does not claim MAI is “best-in-class” or “the biggest fastest model of all.” Microsoft is being uncharacteristically humble and self-effacing about where its models fare. Every claim arrives wrapped in qualifiers: weight class, mid-weight price, matches last year's model.

Instead, Microsoft has decided to lead with “Price-performance.” It says a custom MAI model for Excel matches GPT-5.4 at up to ten times the efficiency. Those are Microsoft's numbers.

For your customers, the translation is simple. MAI is a cost play. It makes routine AI work cheaper, and Microsoft keeps the recognized “frontier” models next door for the hard problems.

Satya Nadella calls Copilot a new OS for work that spans every model. Read that as a statement of neutrality. Microsoft wants to own the place where models are chosen, not win every model contest. My interpretation: they want to be the harness of choice. Not necessarily a bad position to be in, but not easy to achieve in a crowded marketplace.

Their adoption evidence will help your sales conversations with customers. Microsoft reports more than 30 million paid Copilot seats. It's your customers' peers who are buying that.

Regular readers of The Evolving MSP have repeatedly heard me remind them that it’s all about applications. Skip the constantly cited benchmarks. The MSP’s most meaningful metric is the array of jobs customers pay to be done using AI. I've listed some of Copilot's best selling points below:

• After-hours phone and meeting follow-ups: Copilot now has a meeting recaps app for Teams and a Teams Phone agent. The customer gets no missed calls and no lost action items. You get call-flow design, setup, and a monthly review.
• Monthly reporting in Excel: Excel gained skills, personalization, and Power BI context, plus data preparation. The customer's owner gets reports in minutes instead of days. You build the reusable skills for each customer's recurring reports and charge for maintaining them.
• Finding answers in company files: Copilot Search is now the default in SharePoint and OneDrive. Staff find the contract, the price list and the policy. Before you switch it on, audit permissions, because search surfaces everything people have permission to open. That audit is a billable project.
• Agents that do the work: Copilot Cowork turns answers into actions. Customers who adopt it need someone to watch cost and behavior. That is the managed service Microsoft's own MSP playbook calls AgentOps.
• Your own help desk: Run MAI-Transcribe-1, priced at $0.36 per audio hour, on your recorded support calls. If the summaries hold up, you have a proof point and a lower cost of delivery.

Copilot is licensed per user per month. Copilot Cowork adds usage charges on top of the license. Usage is measured in Copilot Credits, one cent each at pay-as-you-go rates.

On December 1, usage billing turns on by default for new Copilot Business licenses bought through CSP. Microsoft moved that date from November 2. The default limit is 4,000 credits per user per month, which works out to $40. That is a ceiling, not a forecast, but it exceeds the license price. Microsoft already moved the date once, so it's reasonable to believe there will likely be an adjustment coming soon.

Cowork is not active by default; customers need an Azure subscription to carry the charges.

Incentives: the Cowork CSP Activation Incentive has run since September 1. Through December 31, eligible SMB customers get a 25-user, 30-day Copilot Business trial.

• Invest now in Microsoft agent operations. The Frontier Partner specialization has been open since September 30. Earn it. Sell AgentOps as a monthly service.
• Invest now in model selection as a service. Microsoft offers MAI, GPT, and Claude in one place, and your customers lack the means to judge among them. You do the judging. Test each customer's top three tasks on every available model. Report accuracy, speed, and cost. Repeat quarterly, because the models change quarterly. This moves you from license reseller to the adviser who decides what runs where.
• Wait on MAI as a premium offer. Do not build a practice around MAI-Thinking-1 until Microsoft publishes its price and independent testers confirm its results. Transcription and image models are the exception. They are priced and available now.
• Do not bet on any single lab. Microsoft does not, and neither should you. In fact, recent examples of the value of model-routing suggest you need to have several models of varying size and capability in your arsenal, routing each task to the least token-hungry one for savings.

With the AI market evolving faster every day, you need to stay constantly in motion, keeping up with the latest relevant AI news for MSPs in AgenticMSP, not grabbing on too tightly to any one new introduction, and staying truly flexible to roll with the pace as the new ones replace it. Here’s my recommended action plan:

1. Right now: Put Copilot on your own help desk before you sell it. You will answer the first customer question with experience instead of slides.
2. This week: Pick three Copilot customers and their top three tasks each. That is your test set.
3. By November 2: Use the Partner Center sandbox to practice ordering Copilot Business with usage billing.
4. Before December 1: At every Copilot customer, name one owner for AI spending, set a spending policy below the 4,000-credit default, and confirm an Azure subscription is attached through CSP.
5. By December 31: Offer the 25-user trial to your best prospects. Propose Business Basic with Copilot at renewal while the promotion runs.
6. By January 2027: Run your test set across MAI, GPT, and Claude where each is available. Deliver the results to those customers as a paid report.

There’s a clear role for Microsoft partners who know what to do when their customers face multiple models, harness choices, and many confusing bills.  You want to be the expert adviser who explains and resolves all of it for them. The one they’ll always turn to for AI advice.

Posted by Howard M. Cohen on October 28, 2026