Workday brings its Sana HR and finance agents into Claude

What did Workday and Anthropic announce?

Workday and Anthropic used the opening day of Workday Rising 2026 to announce a partnership that “connects their platforms and agentic experiences.” The deal brings Claude’s reasoning together with Sana, Workday’s agent and orchestration layer, and with the HR and finance data that Workday holds for its customers.

The idea is that people can start a task in either product and carry their context and controls with them. In Workday’s words, “Claude and Sana exchange information and tasks through the Model Context Protocol (MCP), an open standard, so work carries forward rather than starting over.”

Workday says more than 11,500 organisations, including more than 65 percent of the Fortune 500, use its software. That makes HR and finance one of the largest pools of structured business data an AI assistant can be connected to, and one of the most sensitive.

Ankur Sinha, Anthropic’s head of engineering for enterprise and verticals, said in the release that “HR and finance decisions shape every employee’s pay, role, and career,” and that Claude “brings the reasoning to help people make those decisions faster and with more confidence.” Gerrit Kazmaier, Workday’s president of product and technology, said the combination would “unlock lawful AI across the enterprise.”

What can you do with Sana inside Claude?

The more visible half of the deal is Sana in Claude, a plugin listed in Claude’s plugin directory. Inside Claude, employees can reach Sana agents, analyse Workday HR and finance data and take action without switching apps.

Workday gives two examples. A hiring manager can compare open roles against existing headcount and budget, then build an internal mobility plan that fills those roles with people already on the team. A finance leader can check invoice exceptions against purchase orders, then approve them in Claude or route them to Sana.

Customers can also build their own agents directly in Claude using Workday Agent-Ready Tools, Workday’s MCP tools for reading Workday data and taking action. Workday says admins connect once, with “no new permission model to build and no per-user setup.”

This follows a pattern Anthropic has pursued with other workplace software. Last week Claude gained the ability to edit Google Docs, Sheets and Slides from a sidebar, and the Workday plugin pushes Claude further into the systems of record that companies run on.

How does Sana use Claude for its own workflows?

The other direction runs from Sana to Claude. Workday describes Sana as orchestrating its agents through complex HR and finance processes, such as a compensation cycle or the financial close, and turning work that used to take weeks into processes that run continuously in the background.

Sana holds the context of a process from start to finish and builds the workspace each task needs, such as a live compensation grid or an exceptions queue, without moving data out of Workday. When a step calls for language or reasoning, Sana calls on Claude.

Workday’s example is drafting the note a manager shares to explain an employee’s raise. For that step, Sana “shares only the context that step needs,” and exceptions go to a person for review. That is a narrower data flow than sending a whole compensation file to a model, though Workday has not published the exact fields passed in each case.

How are permissions and audit handled?

The partnership is pitched on governance. Workday says every action follows the customer’s existing Workday permissions and audit controls, and that “Claude surfaces only what that person is permitted to see.” The release’s subheading adds that “the user confirms every change.”

Workday also explained why it thinks this matters in a companion announcement about its Trusted Agent Platform. It argues that pointing an agent at an ungoverned copy of HR data in a data lake lets it read compensation it should never see, and that connecting an agent through a system-level account lets it act without approval. “Headless can’t mean lawless,” said Gabe Monroy, Workday’s chief technology officer.

With Agent-Ready Tools and a feature called Live Data Query, Workday says every request carries its rules: who is asking, what they can see and do, which approvals apply and how the action is recorded. Live Data Query lets teams and agents ask questions of live Workday data in SQL through Workday Data Cloud.

Workday also cites its own research showing that large language models can misapply HR policies even when the correct policy is in front of them, which is why it runs every agent on a layer it calls Agent Harness. Agent Harness supplies each agent with the person’s role, reporting line, permissions and applicable policies, governs each action and routes tasks to models based on cost and performance.

What else did Workday launch at Rising?

The Claude deal sits alongside a broader opening of Workday’s platform. The company says its Trusted Agent Platform is now “headless,” meaning agents and apps can work with Workday data without anyone logging in to Workday itself.

Developer Agent, now generally available, turns plain-language requests into custom agents and apps, and Workday says customers can use it from Claude Code, OpenAI’s Codex, Cursor or the Workday Developer Site. Custom Agents, also generally available, let teams pick the model behind each agent, including Claude, Google’s Gemini, Meta’s Llama, Mistral models, OpenAI’s GPT and NVIDIA’s Nemotron open models. LangChain support is generally available, and Google’s Agent Development Kit is open to design partners.

Workday says it will add Agent Context Service in early access in late 2026 and Grounded Reasoning in early 2027. Grounded Reasoning is meant to check requests against company policy before an agent acts, so an agent asked to move someone into a role the manager cannot fill would stop and explain why. Rivals are building similar layers; Google Cloud recently named a Gemini agent as its universal workplace agent.

The relationship also runs deeper than one plugin. Workday says Anthropic runs its own finance and people operations on Workday Financials and Workday HCM, while Workday has rolled out Claude Enterprise, plans to use Cowork, and uses Claude Code in its product and development teams.

When is it available?

The Sana in Claude plugin is available to early access customers now and is expected to be generally available in late 2026. Workday Agent-Ready Tools are generally available to all customers through Workday Extend AI and will be available through the plugin later this year.

Neither release mentions pricing. The plugin is aimed at joint customers, meaning organisations that use both Workday and Claude.

What are the open questions?

The governance claims are Workday’s descriptions of its own system and have not been independently audited. How well permission checks hold up when an agent chains several tools together, or when a request is ambiguous, will only be clear once more customers use it.

HR and finance are high-stakes areas. A wrong headcount figure or an approved invoice that should have been held has real costs. Workday says users confirm every change and exceptions go to people, but the quality of what Claude presents for confirmation still matters. Teams connecting agents to business systems may want to review how MCP tools differ from resources before granting write access.

Finally, the timeline is staggered. The most interesting safeguard, Grounded Reasoning, is not due until 2027, and the plugin itself is still in early access. For now, this is a partnership announcement with a working preview rather than a product most Workday customers can switch on today.