Model builders are moving beyond simple AI chatbots to creating comprehensive assistants that, in the words of AI dignitary Ethan Mollick, “do real work” in enterprise workflows.

Anthropic is continuing its push in this area with a new feature, Agent Skills, which allows Claude to improve its execution of specific tasks. When relevant, the model can automatically access folders containing specific instructions, scripts, and other resources, then act on them with human approval.

“Think of Skills as custom onboarding materials that let you package expertise, making Claude a specialist on what matters most to you,” Anthropic said in its announcement.

Use Anthropic’s, or build your own Skills

Anthropic has provided four pre-built agent Skills:

• Microsoft Word: To create documents, edit content, or format text;
• PDF: To build formatted PDF documents and reports;
• Microsoft PowerPoint: To craft presentations, analyze content, and edit slides;
• Microsoft Excel: To generate spreadsheets, analyze data, and generate reports with charts.

Users can also build their own skills for Claude and use them across Claude apps, Code, and API. Claude automatically invokes the relevant skills when performing a task, without manual human intervention; for auditability, it provides its chain of thought (CoT) reasoning.

When building new Skills, human users get guidance from an AI “skill-creator” . Claude will ask about the workflow, build out a folder structure, format files, and bundle resources. Then, when performing a task, it will scan available skills, identify relevant matches, and load the minimum information it needs to quickly and efficiently do the job.

Users can also customize skills for specific use cases, and create, view, and upgrade different versions through the Claude Console. Skills are portable and have the same format across workflows, meaning users only have to create them once.

To build custom Skills, users must provide a “human-friendly” name (for instance, “brand guidelines”); a clear description of what the Skill should do and when to use it (such as, “apply brand guidelines to presentations and documents, including official colors, fonts, and logo usage”), and specify when Claude should reference it (for example, whenever creating external materials such as Word documents, marketing materials, client reports, or presentations that represent the company).

Developers can point the agent to resource folders containing all relevant materials such as document templates, and indicate any required software packages (like Python or Pandas), if necessary.

Skills are available to Pro, Max, Team and Enterprise subscribers; users can enable them in their settings, although Team and Enterprise users require admin approval and enabling as well.

Integration with Microsoft 365

Beyond Skills, Claude also now integrates with Microsoft 365.

It connects to Microsoft 365 via the open-standard model context protocol (MCP) and integrates with:

• Microsoft SharePoint and OneDrive: To search and analyze documents across sites and libraries.
• Microsoft Outlook: To analyze emails and other communications and provide insights on the status of projects, feedback from clients, or team alignment.
• Microsoft Teams: To search through conversations, discussions, and meeting summaries, and track project milestones.

Further, Claude now incorporates enterprise search across systems. For instance, when queried about a company’s policy on remote work, the model can pull information from HR documents in SharePoint and emails in Outlook.

“Enterprise search is particularly valuable for onboarding new team members, answering strategic questions like analyzing patterns in customer feedback, and quickly identifying the right internal experts to consult on any topic,” Anthropic said.

The Microsoft 365 integration and enterprise search are available to Claude Team and Enterprise plan customers.

A new arms race

“Claude’s new Skills move is a next step in what’s basically a new arms race,” said Wyatt Mayham of Northwest AI Consulting. It competes with a number of similar platforms from Google, Microsoft, ChatGPT, Perplexity and others.

He noted that model builders are all “chasing the autonomous worker,” but are approaching it from different angles.

For instance, Anthropic’s edge is trust. The company’s “Constitutional AI” design is built for predictability and safety, which is critical in finance or healthcare.

“Enterprises don’t want an AI that just can trigger a workflow,” said Mayham. “They want to trust it not to break one.”

Microsoft and Google, for their parts, are “ecosystem players,” embedding AI deeper into productivity suites by having Copilot and Gemini use internal data such as Microsoft Graph or Google Workspace to make AI feel native inside existing tools. Mayham pointed out that this makes adoption easy, but also locks enterprises into their platforms.

OpenAI, meanwhile, has a split focus: Custom GPTs are a “hit” for no-code builders, while the Assistants API serves developers. However, they’re not unified, and Anthropic seems to be sidestepping this hurdle by designing Claude for direct enterprise integration.

Different priorities

OpenAI is “closest in architectural ambition” to Anthropic, with its AgentKit providing a structured way to build modular, reusable agent workflows, noted Thomas Randall, a research director at Info-Tech Research Group.

However, adapting a tool like AgentKit to Skills requires more automation and lightweight authoring for broader enterprise workflows, he said. So, while Microsoft and Google may not provide comparable solutions to Skills in the short term, “this isn’t necessarily a put-down; these vendors just have different aims.”

“Microsoft and Google are choosing to prioritize platform lock-in and data integration as their competitive edge,” Randall explained. “Ultimately, Skills is a meaningful differentiator if Anthropic can make it work within the Microsoft/Google ecosystems.”

However, while all these capabilities are great, their strategic value lies less in the underlying technology and more in the ability to “institutionalize controlled, repeatable workflows,” he noted. “Enterprises will not gain value unless their internal workflows are already well-structured.”

For enterprises, the upside for all these enhancements is speed and safety, Mayham said. Agentic AI can automate cross-tool workflows such as onboarding processes, data pulls, or documentation prep. But, he said, the real differentiator will be governance: determining who has the authority to grant permission to connect systems, define what data that models can access, and decide how human review is built in.

“Companies should keep tight control over data privacy, access permissions, and audit trails, and assume a human-in-the-loop step for any AI that can change systems of record,” he advised.