Microsoft is positioning Copilot Cowork, its agentic AI system, as the layer that moves financial institutions from AI experimentation to governed execution.

Launched on June 16 after nearly three months in beta testing, Copilot Cowork is built to execute longer, multistep workflows across a user’s files and applications unlike Copilot Chat, which handles scoped question-and-answer tasks, Bill Borden, corporate vice president of worldwide financial services at Microsoft, told FinAi News.

“Copilot Cowork is Microsoft’s agentic AI system that runs complex, long-running, multitool tasks from start to finish, returning completed work rather than just drafts or suggestions,” Borden said. “This means you define the outcome, and it plans, reasons and acts across your files and apps to deliver a finished product.”

The tool operates within the Microsoft 365 trust boundary, inheriting existing policies and controls, and integrates with an organization’s IT control plane to make it easier to audit and govern centrally — which differentiates it from competing solutions that run on the desktop, he said.

Pilot to production

The new tool comes as major banks move beyond limited use cases. Several Microsoft partner banks are already deploying agents at large scale and seeing efficiency benefits:

• BNY has deployed nearly 140 agentic solutions that handle repetitive, rules-based tasks within governed, auditable processes;
• First Abu Dhabi Bank has deployed 18 agents in production across trade operations, customer service and finance, which has doubled processing capacity and reduced turnaround time by 50%;
• Commerzbank built an avatar assistant, AVA, that powers 30,000 monthly conversations and resolves 75% of calls autonomously; and
• Banco Bradesco‘s BIA handles all initial customer interactions and resolves 83% of inquiries across a base of 74 million customers.

One of the fastest-moving use cases is software development, Borden said.

“Developers in financial services were the first to leverage the power of gen AI back in 2022 to create code,” he said. “With agentic AI, the development use case has evolved beyond code creation to accelerate the entire development value chain with agents automating tasks such as unit testing, compliance and operational support.”

Commonwealth Bank of Australia has seen a 42% improvement in developer productivity with Microsoft’s AI tools, he said.

Taming token costs

Asked how financial institution clients and Microsoft are managing ballooning token costs, Borden said the answer starts with anchoring budgets to the most important business scenarios.

He offered the example of preparing for a high-stakes customer quarterly business review.

“You might use Copilot Chat at the beginning of this process to handle fast, scoped tasks, such as summarizing recent customer emails, extracting themes from prior QBRs or drafting an outline for an executive brief,” he said. Cowork, by contrast, is reserved for coordinated, multistep execution across systems.

The second lever is matching the tool to the task.

“One of the other ways we keep costs down is to automatically select the optimal model for the task,” Borden said. “Not every task requires the use of an advanced model. Model choice is a design principle though, so users can manually pick a model if they have a preference.”

Where the tool stops

Despite the “insight to action” framing, Microsoft draws a hard boundary around agentic autonomy. Cowork’s plugins — reusable capabilities that bundle a partner’s data and domain expertise — are not used for the most consequential financial actions such as payments or trade execution, he said.

“For these scenarios, customers can build AI solutions more effectively on platforms like Microsoft Foundry or Copilot Studio,” Borden said.

Register here for the FinAi Lending Summit, set for Oct. 7-8 in Las Vegas. This inaugural event will include speakers from Fifth Third and Capital One as well as a fireside chat with Piermont Bank founder and Chief Executive Wendy Cai-Lee.