Enterprise AI Agents: Workday and Claude for Australia

Enterprise AI agents moved from slide decks into HR and finance systems this week. At Workday Rising on 13 October 2026, Workday and Anthropic announced a partnership that lets people find Workday information, take action and approve work through Sana agents inside Claude — while Workday’s permissions and business rules apply to every request, and the user confirms every change. For Australian businesses, the useful story is not another chatbot. It is what “governed” actually means when agents touch pay, roles, budgets and invoices.

Here is what was announced, what the vendors claim about controls, and a buyer checklist you can use whether you run Workday or not.

What Workday and Anthropic announced

Workday positions Sana as its agentic experience and orchestration layer for HR and finance. Anthropic’s Claude supplies frontier reasoning for individual steps. The two exchange information and tasks through the Model Context Protocol (MCP), an open standard, so work can start in either place and carry forward rather than restarting, the joint press release says.

• In Sana: Sana orchestrates Workday agents through long-running processes — for example a compensation cycle or the financial close — keeps process context, builds the workspace each task needs, and keeps data inside Workday. When a step needs reasoning (such as drafting a manager note about a raise), Sana calls Claude and shares only the context that step needs; exceptions go to a person for review.
• In Claude: Employees can access Sana AI agents, analyse Workday HR and finance data, and take action inside Claude. Vendor examples include a hiring manager comparing open roles to headcount and budget, or a finance leader checking invoice exceptions against purchase orders and approving them in Claude or routing them to Sana. Every action follows the customer’s existing permissions and audit controls in Workday; Claude surfaces only what that person is permitted to see.
• Agent-Ready Tools: Joint customers can build their own agents in Claude using Workday Agent-Ready Tools, with the same governance, security and audit trail. Admins connect once — no new permission model and no per-user setup, Workday says.

Availability, as stated by Workday: the Sana-in-Claude plugin is in early access now and expected to be generally available in late 2026; Workday Agent-Ready Tools are generally available through Workday Extend AI and will arrive via the plugin later this year. Treat those dates as vendor roadmap, not a purchase commitment.

What “governed” enterprise AI agents actually mean

Vendor language about “lawful AI” and “trusted agents” only helps if you can translate it into controls your auditors and privacy officer recognise. From the Workday–Anthropic materials, four ideas matter for Australian buyers:

1. Permissions stay in the system of record. Actions follow existing Workday permissions and audit controls; the assistant does not invent a parallel access model.
2. Data stays put for the process. Sana is described as keeping HR and finance data inside Workday while calling Claude only with the context a step needs. A same-day companion release adds that third-party AI tools “never get raw access” to Workday data — they reach it through Sana, which applies policies before acting.
3. Humans confirm changes. The headline claim is that the user confirms every change. That is the difference between an agent that drafts and one that silently mutates payroll or approvals.
4. Approvals and audit trails remain yours. Agent-Ready Tools are sold as carrying governance and an audit trail from the start. If you cannot export who approved what, when, and under which policy, you do not have enterprise governance — you have a demo.

If you are still deciding how much autonomy you need, read AI agent vs chatbot and our note on AI agent safety in Australia.

Sana in the tools people already use

On the same day, Workday said Sana will work inside Claude, Google Cloud’s Gemini Enterprise, Microsoft Copilot, Microsoft Teams and Slack. One connection is meant to carry each company’s permissions, policies and approvals so HR and finance agents can act where work already happens. Self-service with Sana is described as generally available today in Gemini Enterprise, Copilot, Teams and Slack, with fuller Sana availability on those platforms expected in late 2026; Sana in Claude Cowork is in early access.

The pattern is familiar: agents show up inside chat surfaces, but the rules stay anchored in the HR/finance platform. That is also the pattern Australian mid-market teams should copy when they wire agents into their own stacks via AI integrations and a governed LLM integration layer — not a free-for-all of personal accounts.

Same week: HCLTech’s Anthropic Global Premier Partnership

On 12 October 2026, HCLTech announced a Global Premier Partnership with Anthropic focused on “governed Agentic AI”, safety, cybersecurity and governance. HCLTech says it will set up a dedicated Anthropic Business Unit and Center of Excellence for Claude deployments, and that HCLSoftware will embed Claude and Claude Code into selected products (including HCL BigFix and HCL AppScan) as an Anthropic ISV partner. It also describes integrating Claude into its AI Force engineering platform with policy guardrails, exception handling and human oversight for high-impact changes.

That is a systems-integrator story, not a Workday product story — but it underlines the same buyer theme this week: large vendors are competing on governed agents, not unconstrained automation.

Why Australian HR and finance teams should care

Pay, promotions, terminations, invoice approvals and budget changes are high-stakes. In Australia they also sit near automated decision-making and privacy expectations. If an agent proposes or executes a decision that affects someone’s employment or financial position, you need a clear human step, a record you can explain, and limits on what the model can see. Our guide to automated decision-making and the Privacy Act is the right companion reading for any HR/finance agent pilot.

Workday’s examples (compensation notes, invoice exceptions, internal mobility against budget) are exactly the class of work where a clever draft is useful and an unlogged auto-approve is not. The partnership’s value proposition — permissions, audit, confirmation, data staying in Workday — maps cleanly onto what Australian risk committees already ask for. Whether you buy this stack or build beside another ERP, those four controls should be non-negotiable.

Buyer checklist: enterprise AI agents you can defend

1. Name the system of record. HRIS, ERP, finance ledger — agents read and write only through APIs you control, with the same roles humans have.
2. Map permissions before prompts. Least privilege per agent; no shared “god” service account for payroll or banking.
3. Require human confirmation for irreversible steps. Anything that pays, posts, deletes, offers employment terms or changes access needs an explicit approve.
4. Keep an audit trail you can export. Who asked, what the agent saw, what it proposed, who confirmed, what changed.
5. Minimise context shared with the model. Step-scoped payloads beat dumping whole employee files into a chat window.
6. Separate draft from act. Start read-only and draft-only; promote write actions only after a scored pilot.
7. Stay model-agnostic where you can. Put Claude, Gemini or others behind your own integration so you can switch when vendors change terms or regions. Ground answers in your policies with RAG knowledge assistants where needed.
8. Train the people who will supervise agents. A free starting point is our AI agents course.

Custom work — connectors, approval UIs, logging — is where most Australian mid-market teams actually spend. That is the day-to-day of AI agent developers in Australia and custom AI development, not a press-release plugin alone.

Where regulation is heading

Federal policy is catching up to agent risk. Assistant Minister Andrew Charlton signalled national AI standards by the end of the year and legislation in 2027, with frontier developers expected to show their safety systems work, the ABC reported on 8 October. On 13 October, coverage of the Joint Select Committee on Artificial Intelligence’s early advice highlighted independent safety checks, incident reporting and stronger oversight of agents and models, the ABC reported. Those debates target labs and national settings first, but buyers will still be asked the practical question: is there a serious system to detect risks and prevent incidents in your HR and finance workflows? A documented, permissioned pilot answers more of that question than a screenshot of a chat.

This is general information, not legal or privacy advice. Check your obligations, including Privacy Act and automated decision-making expectations, with a qualified adviser.

Featured image: Photo by fauxels on Pexels.