Strategies for Managing AI Cost with Cowork
Preview
Microsoft introduced Microsoft Cowork as an add-on feature for M365 Copilot users, but usage and processing costs skyrocketed, forcing Microsoft to build a new – and problematic – Cowork pricing model. Companies need to adopt new approaches to allow for valuable pay-as-you-go scenarios to avoid sprawling costs.
Microsoft Cowork’s new cost model is problematic
The new pay-as-you-go model tied to Cowork use will force companies to choose between eliminating Microsoft Cowork as an option for users or accepting unknown costs and unpredictable spending by allowing users to continue using it.
Most companies will likely struggle to quickly rationalize a budget line item they had not prepared for and either turn it off or face a potential wild west of new AI costs that are untethered to measurable business outcomes.
That’s not to say Cowork is not valuable. On the contrary, the pricing model reflects how useful Cowork is—so useful that Microsoft could not sustain Cowork usage costs under the existing licensing revenue from Copilot.
This new wave of AI power will require companies to evaluate the structures, processes, and oversight needed to balance business outcomes with business costs, while accelerating AI governance and support for more thorough, thoughtful business solution engagement.
How Cowork pricing has changed and why that’s a problem
During the Microsoft Cowork preview, unprecedented use and adoption showed that AI usage, complexity, and processing costs had increased significantly. In response, Microsoft announced a new pricing model that extends the pricing structure of the Microsoft 365 Copilot license. Instead of including agent and model processing in the general month-to-month Copilot license, as is the case with the basic Copilot Chat experience throughout M365, companies now need to pay for the processing power required to handle Cowork tasks. Through a pay-as-you-go model, companies will now pay between $1 and more than $10 per Copilot task, depending on task complexity.
This new model creates several problems:
Costs are only understood after the model has run, meaning employees do not know how much a task will cost until after they have incurred the cost.
AI users have been conditioned by “unlimited” models, such as M365 Chat and many other $30-per-month options on the market, to be less deliberate about AI use.
Most organizational financial controls and department budgets are not well equipped to support a large, flexible fund for AI use.
Cowork costs can hide a significant amount of unnecessary processing power caused by limited architecture planning, insufficient data or content preparation, and poorly designed processes.
Usage scenarios and costs are not well connected to administrator oversight. Cost insight is limited to a token rollup tracked by user.
Suffice it to say, organizations will likely need more than one strategy to get a handle on this new capability and its potential demand.
What Microsoft offers to manage costs
The currency for Cowork use is tokens, and tokens cost $0.01 per use (USD).
Basic cost control is configured in the tenant administration settings at admin.microsoft.com. To set it up, an administrator must navigate to Copilot → Cost Management → Configuration and create one or more spending policies. Policies define:
The user group the policy applies to, currently limited to security groups or “everyone”
The maximum tokens allowed per month and per user, including thresholds for alerts
The scope of the pay-as-you-go features (currently Cowork and WorkIQ are the only two options)
The billing method, using an Azure subscription with a valid associated billing profile
Spending policies enable organizations to assign credit budgets, alerts, and access controls for Cowork usage.
Once Cowork is in use, administrators have access to high-level reporting that shows who is consuming Cowork credits and provides a simple dashboard for viewing consumption trends by user.
Monitor Cowork usage and enforce spending limits through built-in Microsoft 365 cost management controls.
Individual users can track how many tokens a task or request used after it is complete. Unfortunately, there is no way to predict how many tokens a request might use before the transaction. Once the task is complete, the user can type “/cost” into the chat window to display the number of tokens used.
The /cost command gives users visibility into how many credits a Cowork request consumed after processing is complete.
What is missing is any understanding of the request types, impact areas, or value scenarios driving Cowork use.
Building organizational capabilities that can help triage and intervene in AI cost
Cowork use represents a bottom-up approach to organizational process improvement. In our short experience, we have found that Copilot use and output are verifiably saving time and effort. If we assume Cowork use is valuable because it reduces effort, saves time, or improves output quality, the work of AI cost oversight should:
Ensure users evaluate value and cost with each use
Provide education on AI options, with an emphasis on cost considerations
Review AI usage scenarios and evaluate lower cost alternatives
Provide low-cost solutions for repeated meta-scenarios across departments, groups, and teams
Improve the factors that lead to AI token waste
Ensuring users evaluate the value and cost of each use
Early adoption of Cowork should focus on helping the organization build a stronger understanding of cost and benefit. Organizations should consider inviting early adopters into a working group that helps the company better understand the productivity scenarios being addressed and how those needs map to costs.
In our organization, Cowork users have been instructed to evaluate their costs using the “/cost” option in the chat window. Short weekly meetings allow for quick sharing and review of how Cowork is being used and provide a more nuanced view of usage trends. Finally, follow-up one-on-ones are being used to review recurring use cases, such as our sales team using Cowork on each RFP we respond to, to identify lower-cost alternatives and opportunities to create agents for users or teams.
Providing education on AI options
Some Cowork use is likely the result of users seeing Cowork as the easy path to AI-enabled support. We have found that some Cowork use is because users are unaware of features, capabilities, or techniques in Copilot that could produce similar results.
Cowork users share that Cowork is better at research, producing reports or PowerPoint presentations, and working across larger document sets to summarize or organize insights. To counter this, organizations can increase education on:
How to use first-party Copilot agents, such as the PowerPoint agent or Copilot within PowerPoint, which are much better at building presentations than the default Copilot chat window
How Copilot RAG works with contentand how to properly add or reference content so tasks such as document summarization work best, especially across multiple documents
How Researcher and Analyst agents are better suited for research, investigative reporting, and analysis tasks than the default Copilot Chat window
How spending time organizing knowledge, content, or data, and explicitly referencing it, can help Copilot provide better summaries, analysis, and citations
In addition to the above, organizations may want to arm Cowork users with tips and techniques that reduce agent token use. For example, directly referencing the content Cowork should use can reduce the need for Cowork to crawl and evaluate content before using it. Skill files can also help reduce token overhead by enabling just-in-time use, reducing the initial context-window cost that drives token use up. Education on initial prompting techniques can reduce rework and iteration, which often result from poor initial instruction. As Cowork use increases, sharing stories of effective, streamlined use will be a valuable learning activity.
Evaluating lower cost alternatives
Once foundational Microsoft agents and Copilot experiences have been explored as alternatives, organizations should work to understand additional alternatives to Cowork that may be more cost-effective.
For example, we have found that many organizations we work with have not yet invested in agent-building expertise or custom agents. One of the most impactful alternatives to combat Cowork costs is to create custom-built agents using Copilot Studio.
Copilot Studio agents, which can be created by any Copilot user, fall under a “fair use” clause with a premium M365 Copilot user license. In other words, if you create an agent using Copilot Studio and that agent is used by a Copilot-licensed user, the organization does not have any additional per-use costs. There are some caveats: the agent cannot consume premium features, such as premium connectors, and the agent cannot run autonomously; it must be initiated and used by a human. The term “fair use” has been introduced into the licensing clause for M365 Copilot to indicate that Microsoft may also monitor scenarios that appear to work around intended licensing requirements.
What is especially compelling about this option is that Copilot Studio agents can run under specific OpenAI and Anthropic models (as enabled by the organization) and can duplicate or mimic the intelligent agent experience in Cowork. A few examples of agents we created include:
One of our “vibe coders” found that Copilot was poor at building PowerShell scripts. They used Cowork to generate PowerShell scripts for demo and test SharePoint configuration. Using Copilot Studio, we built a script agent that uses Anthropic’s Opus model.
After client interviews, our team would use Cowork to take our rough notes and the interview transcript generated from Teams and fill in the gaps so our notes included more of the missed detail. Using Copilot Studio, we built an agent designed to tackle that task with more precision than Copilot could provide alone.
Our sales team would use Cowork to create project kick-off deck drafts that summarized and pulled in content from our master proposal and statement of work documents. A custom Copilot Studio agent for this purpose returned similar results to Cowork.
Reducing token waste
One of the most compelling architectural challenges organizations will need to understand is how to better control the factors that increase AI costs. These costs often result from poor or ineffective prompting, disorganized data or content, and little to no AI architecture.
As AI model effectiveness continues to increase, so will the ability of those models to navigate around unclear instructions and disorganized content. The key question is: is that a good use of AI?
In one customer scenario, we attempted to use AI to evaluate a dataset that had not been cleaned up or organized for reporting. We found that while the model could determine the content structure and ultimately return an accurate report, each transaction required the model to repeatedly work through disorganized data, creating high token use per transaction. Token use caused by poor data or content structure is simply waste.
To combat this, organizations should strengthen their architectural understanding of AI, including the prompting expertise, data science, data engineering, and AI architecture decisions that drive AI efficiency and accuracy. We have also helped organizations consider the broader content management and oversight needed to ensure AI returns accurate, high-quality results from existing SharePoint and M365 repositories.
Conscious and controlled learning
Cowork represents an exciting evolution in useful AI productivity tooling for end users. But with new cost-control considerations, this capability also requires organizations to increase their capacity and capabilities for AI oversight so they can grow and evolve AI use while effectively managing AI costs.
Using out-of-the-box reporting, paired with thoughtful user engagement and education, AI architecture oversight, and stronger management of the AI tool portfolio, will help ensure waste is contained and AI use continues to provide value to both employees and the organization.
