All research

At Ignite, Microsoft Made AI Cost Control the Product

Microsoft built its Ignite keynote in San Francisco on November 17 around human ambition, intelligence and trust, and cost came up in nearly every segment along the way. Executives showed what their demos cost in credits, Agent 365 gained budgets and chargebacks, and Mustafa Suleyman pitched Microsoft's own models on price-performance. Microsoft is responding to customers who are struggling to keep AI spending on plan: in Creative Strategies' October survey of 200 CIOs at companies with at least 1,000 employees, 49% said token and API spending is running above their approved 2026 plan. Microsoft's answer gives customers better tools to monitor and cap that spend, while the pricing behind it stays metered and largely in Microsoft's hands.

The Subscription Covers Less of What Microsoft Showed

Copilot pricing changed shape this year: Copilot Cowork became generally available in June with pay-as-you-go pricingof $0.01 per Copilot Credit, and in September Microsoft split Copilot into a per-user license for everyday AI and credit-based billing for advanced work, which now covers Cowork, Code, Autopilot, frontier models and Work IQ APIs. The $30 per user per month license price did not move, while almost every compute-heavy new feature went onto credits.

On stage, Jacob Andreou, executive vice president of Copilot, said Copilot is the only way to draft a PowerPoint "without spinning a meter," which holds for the everyday AI the license covers. Most of what the keynote showcased runs on credits, including CoWork building a storyboard and a customer tracker, Autopilot keeping a spreadsheet current and sitting in on product meetings, and Code turning a markdown file into an interactive 3D demo.

Microsoft's Demos Cost More Than Its Default Spending Limit

Microsoft made a point of showing costs during its demos, starting with Judson Althoff, CEO of Microsoft's commercial business, who said his Sunday of preparation for an NRF keynote used about 10,000 credits, roughly $100, for two hours of work. Chevron showed a CoWork session that produced an engineering basis-of-design document and workbook for 4,500 credits, about $45, and said engineers' time on the task went from months to minutes.

Those numbers look reasonable against the labor they replace, and they also sit well above the default spending limit of 4,000 credits per user per month, about $40, that Microsoft is setting for new Microsoft 365 Copilot Business licenses and that admins can adjust. Althoff's Sunday alone came to two and a half times that default, and Chevron's single session exceeded it. Companies that want their people working the way Althoff described will be raising their limits, and the bill will grow with every person they raise them for.

Managing the Bill Becomes an IT Job

The rest of the cost story ran through Agent 365, starting with a member of Chevron's enterprise data and AI team explaining that IT has to set a spend limit before anyone can use CoWork, and that users who hit it must request an increase. Agent 365 now lets admins set alerts and limits by agent, department or individual, flag cost spikes when an agent switches models, impose hard caps on Microsoft agents, and charge spend back to cost centers through Azure subscriptions. Satya Nadella described the new Agent 365 Business Analytics as an ERP system for AI usage.

CFOs have been asking for these controls, and they are useful, even as they move the work of managing AI cost onto customers' IT and finance teams. Business Analytics pairs spend with value, but leaders calculate that value with their own task times and costs, and Microsoft's presenter cautioned that the figure is a conservative estimate of the work's value and does not represent cash savings. Credits, meanwhile, track models, context, tools and runtime, so a task that takes more steps to reach a mediocre result costs more than one that gets it right the first time.

Spending limits also decide who inside a company gets to work with agents. An executive spending $100 on a Sunday and an employee capped at the default are using the same product at very different levels, and every request for an increase is a point where someone may simply stop asking. Publicis Chairman and CEO Arthur Sadoun told the audience they could all get fired if they failed to bring the right level of AI adoption into their organizations, and budget policies will shape who gets the chance to adopt.

A Copilot customer and early adopter I met with sees the same mechanism differently. The request for more credits, they pointed out, starts a conversation a flat license never required: the people doing the work have to explain to management what the agents are producing, and management has to decide whether that work is worth funding. That brings the value discussion down to the people closest to the work, and whether those requests turn into conversations or quietly stall will come down to each company's culture.

Microsoft Is Using Its Own Models to Bring Costs Down

Suleyman, CEO of Microsoft AI, was direct about the problem, saying token spend is going through the roof and that at frontier prices the math is unsustainable, and he positioned MAI, Microsoft's model family, as the answer. By Microsoft's numbers, averaged across four widely used benchmarks for everyday enterprise coding, MAI Code 2 delivers 98% of GPT-6.1 Sol's performance at 40% of the cost, and MAI Cyber 2 Flash outperforms GPT-6 Astra on the CyberGym end-to-end benchmark at less than 2% of the cost per task. He argued customers should not pay on every token for scientific capability that a routine bug fix never needs. Microsoft also said Foundry's model router is up to 65% cheaper than sending every request to a frontier model, and Nadella said Decision 1.1 is 200 times less expensive than using an LLM.

In the Alaska Airlines demo, Microsoft's optimizer improved the accuracy of a maintenance-coding agent by adding instructions, which pushed token costs up, so the team fine-tuned MAI Code 2 Flash to carry that knowledge in the model itself and reached 97% accuracy on exact code matches.

The savings depend on routing, which Microsoft controls: in Copilot, auto-routing picks the model and effort level for each request, and Microsoft now has its own cheaper models to route to. Agent 365 shows customers which models their spend went to after the fact, and Chevron checked its cost with a /cost command once the session was done. Perception, Microsoft's new security system, already shows an estimated cost and time before a playbook runs, and extending that kind of upfront estimate to every CoWork and Autopilot task would give customers a say before the work begins.

A credit has a list price of one cent, but how many credits a task consumes depends on Microsoft's models, its router and its agent harness. Ignite gave customers far better tools to monitor and cap their spending, while the rate at which their work consumes credits remains largely in Microsoft's hands.

Trust Through Observability Applies to Microsoft Too

Microsoft tied trust to observability throughout the keynote: Nadella said trust starts with observability at every layer of the stack, and Judson Althoff closed by saying the only way to gain trust is through observability. Agent 365's cost controls bring that principle to customers' AI spend, and the Microsoft 365 outage five days before Ignite was a reminder that customers also need visibility into the service their agents run on. From 11:20 a.m. ET on November 12 until 11:10 a.m. ET the next day, customers lost access to Outlook, Teams, SharePoint and Copilot, following a September outage that blocked access to Microsoft 365 for days. Microsoft apologized in a statement on its X account, which noted the problem was unrelated to cyber or AI agents.

Satya Nadella addressed the outage from the Ignite stage, saying, "We know the trust you place in our services and take that responsibility extremely seriously," and adding that Microsoft has "work to do to provide that resilience that you can count on" and is committed to learning and continuously improving. That was the right message, and customers will now look for the detail behind it. Microsoft's preliminary post-incident report, shared in the Microsoft 365 admin center, points to a previously undetected bug introduced during an earlier network capacity upgrade. The final report is due by November 20.

The final report gives Microsoft a chance to apply its own standard to itself, and the keynote supplied the context: Nadella explained that agents are reshaping Microsoft's infrastructure, that GitHub is seeing exponential growth in pushes, commits and pull requests driven by agents, that Microsoft is rebuilding GitHub's core to keep up, and that the same work has to extend across Azure. If the pace of change needed to serve that demand played a part in the outage, saying so openly, and sharing how resilience now factors into the way Microsoft builds and ships, would give customers the kind of visibility into Microsoft that Agent 365 gives them into their own agents. As companies start budgeting for agents that work through the night, that visibility into the service underneath becomes part of the cost calculation too.

The Adult in the Room Has to Show the Way

Althoff closed the keynote by returning to human ambition, describing posters from a Microsoft event for employees' children, filled with their hopes for the future, and noting that none of them mentioned prompt engineering and that children will not want to live in a chat window. Technology does not have ambitions, he said, people do, and he called Microsoft "the adult in the room" for the transformation ahead.

For individuals, that case is easy to accept, since AI can put real capability behind a person's ambition and the Copilot demos showed one person getting through work that used to take a team. Organizations need more help, particularly in sectors that move carefully by design.

Our CIO survey shows how far most organizations still have to go: only 15% have scaled AI with monitoring, permissions and an operating KPI in place. Just 6% report a finance-reviewed return above full cost, and 57% have not measured full-cost return at all. Nearly half, 48%, cannot point to a single workflow where AI delivers measurable value. Employee adoption sits near the bottom of the list of barriers, cited by 9%, while security and permissions top it at 24%. In financial services, healthcare, energy and the public sector, security approval is the next gate to expansion for a third of CIOs, compared with one in five at technology companies.

Our companion survey of 200 security leaders at US enterprises with at least 1,000 employees helps explain why. Among organizations with business agents in production, fewer than 5% have a unique agent identity and an accountable owner on at least three-quarters of those workflows, and fewer than 3% have tested stop-and-rollback processes at that level. Only a quarter can verify most of what their agents touch downstream, from tools to data movement to calls between agents. For about a third of the organizations that reviewed an AI project for production in the last 90 days, the most recent one was delayed or rejected primarily for security.

Microsoft's tools address much of this: Agent 365 and Entra give agents identities and owners, Perception and Purview cover security, and the new cost controls give finance a view of spend. Building the tools that make adoption easier is one part of the job, and showing organizations how to get from a pilot to a governed, measured deployment, with the processes, roles and decisions that go with it, is another that the keynote spent far less time on. BNY CIO Leigh-Ann Russell described what that work looks like: defining the role of the humans and the role of the agent, the permissions that govern it, and where accountability lies, then testing before going live and keeping a human over the loop with the authority to stop it. If Microsoft wants to be the adult in the room, sharing playbooks like that one openly, including what went wrong along the way, would do as much for adoption as the next product launch. The children on those posters will one day work inside these organizations, and how well those organizations learn to adopt AI will shape how much of that ambition they get to realize.