- Towards AGI
- Posts
- Workday's AI Agent Now Finishes Your Work. Who Audits It?
Workday's AI Agent Now Finishes Your Work. Who Audits It?
three moves that change who controls your enterprise data
Shen Pandi
October 16, 2026
Today, we're diving into:
Gen AI: Workday's AI agents now finish the work instead of answering questions
Hot Tea: Anthropic buys its way into confidential AI
OpenAI: Business AI adoption flattens as OpenAI closes the gap
Workday's AI Agent Now Finishes Your Work. Who Audits It?
For two years, enterprise AI has mostly meant a chat box that answers questions. This week, one of the largest HR and finance platforms in the world decided that was no longer enough.
At Workday Rising, the company made its Sana agent the new home screen for every employee. Workday says more than 11,500 organizations use its platform, including over 65% of the Fortune 500. That is a lot of payroll, hiring and finance work about to run through an agent first.
Is Your Data Ready For Agents That Act, Not Just Answer?
Find out whether your metadata, lineage and access rules can hold up before an agent starts approving work on top of them.
From Copilots To Autopilots
The bigger shift sits underneath the new interface. Workday also introduced four autopilot agents for talent acquisition, payroll, accounting and deployment.
Unlike a copilot, which waits for you to log in, an autopilot keeps running whether you are online or not. The Talent Acquisition Agent already monitors open roles, recommends candidates, drafts outreach for recruiter approval and schedules interviews.
"From co-pilots to autopilots, that's a big transition that you will see across our agents today." Joel Hellermark, Chief AI Officer, Workday
One Front Door For Every Agent
Workday is also opening Sana to the tools your people already use, including Claude, Gemini Enterprise and Microsoft Copilot. Those tools do not get raw Workday data. Requests pass through Sana, which applies your permissions, approvals and policies and records the work.
A new Agent Harness lets you build custom agents that run in isolated environments with built-in memory, including third-party agents that connect through the Model Context Protocol.
The Real Question Is Not What Agents Can Do
Most leaders will read this launch as a productivity story. That is the wrong reaction. This fundamentally changes who, or what, is making decisions inside your HR and finance systems.
Analyst Melody Brue of Moor Insights & Strategy put it plainly: "An agent can be authorized and still make the wrong call."
She also warned that customers need to separate permission to recommend from permission to execute. Most enterprises have never written that distinction down, because until now only humans executed.
What You Should Do Before You Switch It On
Split recommend from execute. List which agent actions need a named human sign-off, starting with hiring and payments.
Name an owner for every cross-platform workflow. When an agent spans Workday and another system, someone must own approvals, exceptions and the audit trail.
Pilot before you trust. Most autopilots reach early adopters in early 2027, so test the controls on real workflows first.
Check the data under the agent. An agent that acts on a wrong cost center or a stale job code does the wrong thing faster.
The better alternative is to treat agent rollout as a governance project with a productivity payoff, not the other way around.
The organizations moving first will not just automate faster. They will be the ones who can prove, line by line, what their agents did and why.
Anthropic Just Bought Its Way Into Confidential AI
Every enterprise AI conversation eventually hits the same wall. Your most valuable data is also your most sensitive, and nobody wants it sitting in plain view while a model works on it.
This week Anthropic licensed technology from Opaque Systems and hired part of its team. According to The New Stack, that includes CEO Aaron Fulkerson and CTO Rishabh Poddar. Terms were not disclosed.
What Confidential AI Actually Means
Standard encryption protects data on disk and while it travels across networks. The gap has always been the moment the data is actually being used.
Confidential AI means running workloads on sensitive data while the data stays protected by the hardware itself, even while it is being processed.
Opaque grew out of UC Berkeley research about a decade ago and raised a $24 million Series B earlier this year. Fulkerson describes the goal as "verifiability for the AI stack."
Why This Matters For Your Roadmap
Constellation Research reads the deal as part of Anthropic's push toward enterprises that want on-premises deployments and sovereign AI. If that is right, the gap between "we cannot send that data to a model" and "we can" just got narrower.
For regulated industries like banking, insurance and healthcare, that changes which use cases are even on the table.
The Catch Nobody Is Talking About
Confidential computing exists so you do not have to trust the operator. Now one of the largest model operators owns a big piece of the expertise.
"The whole point of confidential computing is that you don't have to trust the operator." Heinrich, 0G Labs, speaking to The New Stack
Neither company has committed to maintaining Opaque's open projects, such as Confidential MCP. Experts quoted by The New Stack advise developers to pin versions and mirror repositories now.
Hardware Will Not Fix A Data Problem
The real bottleneck is not encryption. It is knowing which data is sensitive in the first place, who owns it, and which agents should ever touch it.
A secure enclave will happily protect data that was misclassified, duplicated or shared with the wrong agent. Protection without classification is just a locked box with the wrong things inside.
Building confidential AI starts with trusted enterprise context. DataManagement.AI helps organizations create that foundation by unifying metadata, lineage, governance, business glossaries and knowledge assets into a single intelligence layer, so you know what is sensitive, who owns it and which agents are allowed to reach it.
The enterprises that win the confidential AI race will not be the ones with the strongest vault. They will be the ones who know exactly what belongs inside it.
Business AI Adoption Just Hit A Plateau. OpenAI Is Closing In.
For two years, the AI story has been simple: more companies, more spend, every month. The latest data from corporate card provider Ramp suggests that story is changing.
Ramp's October AI Index, published October 14, puts paid business AI adoption at 55.9%, up just 0.62 points. Its lead economist, Ara Kharazian, titled it "Cracks in the AI Thesis Part 3."
The Race At The Top Is Tightening
Anthropic still leads, used by 43.8% of businesses on Ramp's platform. But OpenAI grew more than four times faster this month, adding 1.12 points to reach 40.6%.
"It feels like we're at the top of the S-curve." Ara Kharazian, Lead Economist, Ramp
His conclusion is that future growth "will have to come from more intense adoption rather than more customers." In other words, the next phase is not about who signs up. It is about who gets real work out of it.
Why Spend Is Stalling
Ramp recorded two straight months of virtually no growth in token spend, even as token volumes kept rising. The main driver, according to Kharazian, is falling prices at the frontier.
You Have More Leverage Than You Think
Ramp's read is blunt: "the customer is in the favorable negotiating position." Contracts are not especially restrictive, and vendors are cutting prices to compete.
Renegotiate now. If your AI contracts were signed a year ago, you are likely paying for yesterday's prices.
Stop single-sourcing by habit. With the two leaders this close, your workloads should be able to move between them.
Measure depth, not seats. Licenses tell you who has access. Completed tasks tell you who is getting value.
The Wrong Reaction Is To Cut Back
A flattening adoption curve tempts some boards to pause. The better alternative is to use cheaper tokens to push AI deeper into fewer, higher-value workflows.
Know Your Inference (KYI) is becoming a core enterprise capability. By understanding how every inference affects cost, latency, accuracy and governance, you can see what each model really costs per completed task, choose the right one for each workload, and negotiate with evidence instead of instinct.
The next AI advantage will not belong to the companies that adopted first. It will belong to the ones that went deepest, at the lowest cost per outcome.
Journey Towards AGI
Research and advisory firm guiding on the journey to Artificial General Intelligence
Know Your Inference Maximising GenAI impact on performance and Efficiency. | Model Context Protocol Connect with us, and get end-to-end guidance on AI implementation. |
Your opinion matters!
Hope you loved reading our piece of newsletter as much as we had fun writing it.
Share your experience and feedback with us below 'cause we take your critique very critically.
How's your experience? |
Thank you for reading
-Shen & Towards AGI team