AI工具Score B (67)

What AI at work really costs | Microsoft Copilot Blog

1 天前2 viewsSource: microsoft.com
As you hand off more work to AI, it’s worth comparing the total costs of the leading options, just like you would any other investment that helps you get work done. Today we’re sharing new research on how those costs compare across the same work, and a closer look at what drives the difference. Microsoft commissioned and funded research with GigaOm on Demystifying the TCO of Enterprise AI , which measured the total cost of ownership of Microsoft Copilot, ChatGPT Enterprise, and Claude Enterprise through hands-on testing in October 2026. Microsoft ran the testing and cost analysis. GigaOm contributed to the research design, validated the methodology, results, and cost analysis, and co-developed the findings. The headline is that, on the study’s modeled workloads, Copilot had the lowest estimated monthly cost for both everyday and advanced AI users, while quality was broadly comparable across all three products on the tasks tested. What the research tested The research modeled two kinds of AI users. Everyday users rely on AI across their workday for tasks like finding information, preparing for meetings, analyzing data, and creating content. Advanced users do all of that and also hand off complex, multi-step work to agents that run on their own. The evaluation covered 37 unique prompts across Copilot Chat and Cowork, plus Copilot in Word, Excel, and PowerPoint. Over a modeled month, that represented 100 tasks for an everyday AI user and 140 for an advanced AI user, including 20 agentic workflows. Each product was tested against the same fictional company and underlying data, with prompts run repeatedly to account for variation in AI responses. You can find the full testing methodology in the report. What the research found For the everyday AI user, Copilot costs $30 a month. Claude costs $205 and ChatGPT costs $81, once usage charges were added to their subscriptions. That’s more than 6x and 2.5x Copilot’s cost, respectively. For the advanced AI user, Copilot cost $96 a month. Claude cost $299 and ChatGPT cost $149. That’s more than 3x and 1.5x Copilot’s cost, respectively. All cost estimates are based on October 2026 list prices, or on an assumed price where list pricing isn’t published, and exclude negotiated discounts. Actual pricing may vary. Same AI work. Very different bills. Copilot had the lowest monthly cost for everyday and advanced AI users, with broadly comparable quality. Cost matters only if the work is done well, so the research measured quality too. Before testing, human evaluators wrote a checklist for every prompt: the right sources, the right people, the right facts and figures, a complete deliverable, and no actions the user didn’t ask for. A run passed only if it met every item. Partially correct work failed. Across the workload, the three products delivered broadly similar quality. Results varied by task, so this holds for the workload as a whole, not for every individual task. The Copilot cost advantage Why does Copilot cost less? Three factors explain much of the difference. First, everyday AI usage is included in the Copilot subscription, so those tasks don’t generate additional usage charges. Second, Work IQ , our workplace intelligence layer, adds another important dimension to Copilot: access to relevant business context. And third, Auto dynamically selects models and effort levels based on the work, helping manage the cost of each request. Why everyday AI belongs in the user subscription The biggest cost difference came from everyday work. Chat and Microsoft 365 app usage is included in the Copilot user subscription, so the study’s estimated cost for these everyday tasks remained within the user subscription. The assumed competitor subscription costs may be $10 less than Copilot, but once AI usage charges were added, that advantage disappeared within the first week of modeled use. When every task has a meter, each prompt adds to the bill. A successful rollout, where people use AI more and more, turns into a budgeting problem. Companies may respond by rationing access, so only some people get to use AI. AI innovation can come from anywhere within an organization, and the economics in this report show why a fixed price for everyday work can make that possible. Those economics carry through to advanced work. Agentic workflows on Copilot run on usage-based billing with Copilot Credits. Because Copilot’s subscription already covers everyday work, advanced AI users still had the lowest total monthly cost after their agentic charges were added. Work IQ gets the right context in fewer trips The second factor is context. Copilot runs on models from multiple labs. Two companies can use the exact same AI model and get completely different results. Context can be an important part of that difference. The scale of that context is enormous. Microsoft 365 holds 19.5 exabytes of data, with 10 billion new files added to SharePoint and OneDrive every week. The challenge isn’t simply giving a model access to all of it. It’s finding the information that matters for the work in front of you. A model’s training alone doesn’t tell it who edited a proposal, whether finance approved it, or what your manager said about the discount in Teams. That context goes beyond your email, meetings, chats, files, and business applications. It includes the people you work with, your relationships, your projects, and the activity around your work. The model needs access to that broader context to do the job well. In this research, Claude and ChatGPT accessed Microsoft 365 data through connectors and add-ins. Gathering the right context can involve multiple steps: finding a deck, checking the latest version, reading comments, following a discussion about a discount, or locating a relevant email or meeting. Depending on how the retrieval process works, those steps can add time and increase the amount of information the model needs to process, potentially increasing costs. Work IQ brings that context into the conversation by helping Copilot quickly find the right information across an enterprise, updated instantly. Several design choices make this possible. The path to context matters. Connectors can take a trip to find each piece of context. Work IQ gives Copilot a map of your work to start from. It starts with the semantic index, which organizes Microsoft 365 content into a map of what’s there. That map helps Copilot find relevant material across file formats without relying solely on keyword matches. When a request comes in, Work IQ pinpoints the key information that matters instead of re-reading and parsing entire files. Work IQ can also use your working set, the work you’re actively doing, to help identify documents and information relevant to the task. Memory can help carry useful context across tasks, such as preferences or information about an ongoing project. Getting context right is a balancing act. Too little, and the agent misses something and has to search again. Too much, and the model spends time and tokens on material that doesn’t help. By surfacing what a task actually needs, Work IQ helps reduce the amount of context the model has to process. Auto routing does the work behind the scenes The third factor is model selection. No single AI model is best at everything, and the most powerful model isn’t always the right one. A presentation, a data analysis, and a simple everyday task may each benefit from a different model and level of effort. The challenge is knowing which model to use when. Use a frontier model for everything, and you’re paying for power you may not need. Always use a smaller model, and you risk compromising quality on the hardest tasks. Auto routing takes that complexity away. Auto is designed to dynamically match each task and each part of the task to the right model and effort level. That means you don’t pay premium rates by default. Auto spends more compute where it improves the outcome and less where it doesn’t. Say you’re preparing for a customer renewal. You need to gather notes, research the customer, and prepare a deck. The expensive way to do that is to run every step through your most powerful model at its highest effort. Auto takes a different approach. It might use Opus to lead the work, GPT-5.5 at medium effort to gather your notes, Opus at high effort for the research that requires deeper reasoning, and GPT-Image 2.5 for the visuals. Auto routing brings it all together, designed to lower the cost of the overall task without compromising the quality of the result. Not every task requires the most powerful model. Finding relevant files doesn’t take the same effort as researching a customer. Auto dynamically selects models and effort levels based on the work, helping manage the cost of each request. Run the comparison yourself These results reflect the tasks selected, how often they’re done, the product settings, and current list or assumed prices. Your workloads and your negotiated pricing will differ, so read the findings as directional. The better test is your own. GigaOm recommends applying the same approach to your environment: choose tasks that reflect your organization’s work, run each repeatedly, assess quality against defined criteria, and apply your own pricing. Your Microsoft account team can share the evaluation methodology and full prompt list used in this research, so you can start from the same baseline. As AI becomes a core part of how people work, the useful question is what the work costs and how that cost changes as more people use AI more often. Read the full report , and talk to your Microsoft account team about running the evaluation in your own environment. LinkedIn Nicole Herskowitz Corporate Vice President, Microsoft 365 and Copilot Nicole is a seasoned leader with over 25 years of experience in the technology sector. She currently leads Microsoft 365 and Copilot product marketing, accountable for driving business growth, product direction, and go-to-market strategy. Nicole has also held key leadership roles across Microsoft Teams, Microsoft Azure, and Visual Studio. Her extensive tenure at Microsoft spans responsibilities in product management, product marketing, and sales, with experience in both global and European markets. See more articles from this author Work IQ Enterprise Tips and guides Microsoft Copilot

Read the full original article:

microsoft.com