The HR Tech Deal Moves to Agent Governance
On this page 8 sections
An illustrative procurement review
The following opening is an illustrative composite, not a reported account of a specific employer. At 9:06 a.m. on a Tuesday, a software selection committee at a large employer sat through its third AI platform demo in eight days.
The CHRO was in the room. So was the CIO. So were the general counsel, the shared-services leader, the head of employee support, and a procurement lead who had already decided that the old evaluation template was useless.
Each vendor had prepared the familiar story.
Agents could answer employee questions. Agents could open cases. Agents could route approvals. Agents could summarize policy documents, draft responses, recommend next steps, and take action in other systems. The slides showed better productivity, faster resolution, cleaner handoffs, fewer tickets, and more self-service.
Then the procurement lead interrupted with a question that changed the meeting.
Where do the agents live once we approve them?
Not on the slideware level. In operational terms.
Who assigns them an identity? Who determines what systems they can touch? Who approves a new model endpoint? Who logs which agent took which action on which employee record? Who can see whether an action was executed under a policy rule, a human approval, or an LLM choice? If a regulator, auditor, works council, or internal investigator asks what happened six months later, which product reconstructs the answer?
That is a practical enterprise AI procurement question in HR tech.
For the last two years, most buying conversations were still framed around visible surfaces: copilots, chat interfaces, drafting tools, search, summaries, and workflow shortcuts. Those surfaces matter. They are how the user feels the product. But as AI has moved closer to hiring, employee service, payroll-adjacent requests, internal mobility, compliance workflows, and cross-functional case work, the center of gravity has shifted.
The assistant is only one part of the product. The governance layer determines what it may see and do.
That is why Workday, ServiceNow, and Salesforce are suddenly colliding in a more direct way. They are not just competing to offer AI features inside their respective domains. They are trying to own the operating layer that sits above those features: the place where agents are registered, permissions are defined, handoffs are controlled, models are chosen, audit trails are retained, and business outcomes are measured.
This is why the category feels different from a year ago. The market is no longer asking only, “Which platform has agents?” It is asking something harder and more valuable.
Which platform can be trusted to govern them?
That is a much bigger question than feature breadth. It reaches into budget ownership, risk policy, architecture decisions, and workflow power. It also explains why the next strategic fight in HR tech will not be won by the vendor with the most charming demo.
The current product race therefore favors vendors that can demonstrate control across identity, permissions, actions, and evidence. Whether one vendor can own that layer across the enterprise remains unproven.
Two independent security baselines make the requirement more concrete. The NIST Generative AI Profile treats governance, measurement, documentation, and lifecycle risk management as connected work. Anthropic’s trustworthy-agents framework emphasizes human control, transparency, secure interactions, and privacy. OpenAI separately argues that prompt-injection defense must constrain the impact of manipulation even when an attack succeeds. None of these sources certifies the enterprise platforms discussed below; they supply criteria against which buyers can test vendor claims.
First AI Wave Created a Visibility Problem
The first enterprise AI wave was easy to sell because it looked incremental.
A recruiter drafted outreach faster. An HR rep summarized a case. A manager searched a policy page with natural language instead of clicking through a portal. None of that seemed to require a new software architecture. Vendors could bolt a generative layer onto an existing product, call it an assistant, and show immediate time savings.
That logic broke down as soon as companies tried to scale.
Once a business has multiple agents, multiple model providers, multiple approval paths, and multiple systems of record, the value of the interface starts to depend on something deeper: whether the organization can see, govern, and coordinate what is actually happening beneath the interface.
That is the point where procurement changes shape.
The old questions were surface questions:
| Old buying question | Why it was acceptable in the assistant era |
|---|---|
| Does the agent answer common questions well? | Early deployments were narrow and mostly informational |
| Does it save recruiter or HR time? | Labor savings were the easiest ROI story to tell |
| Does it work in Slack, Teams, or the portal? | User adoption was the immediate battle |
| Can it integrate with our HRIS or ITSM? | Integration was seen as a connector problem |
The new questions are infrastructure questions:
| New buying question | Why it matters in the agent era |
|---|---|
| Who assigns trusted identity to agents? | Execution rights, approvals, and traceability depend on it |
| Where are model, tool, and workflow choices governed? | Cost, risk, and consistency now vary by route |
| Can third-party and homegrown agents be managed together? | Most enterprises will not run a single-vendor agent stack |
| Can HR, IT, legal, and shared services see the same logs? | AI risk now crosses functional boundaries |
| Can the platform connect action to outcome? | Buyers are paying for governed execution, not chat volume |
This shift is why the market language is changing from assistant to agent, from feature to runtime, and from productivity to governance.
It is also why the platforms with the best starting positions are not identical.
Workday starts with people data, business process logic, and HR-finance system-of-record credibility. ServiceNow starts with workflow orchestration, service operations, and a cross-functional control posture. Salesforce starts with employee-facing conversational surfaces, service workflows, Slack distribution, and increasingly explicit agent runtime controls.
Each company is trying to move toward the same valuable center.
But they are approaching it from different directions.
That difference matters because buyers are not simply choosing an AI brand. They are choosing the place where digital labor will be governed.
Workday Wants to Govern Agents Like It Governs Workers
Workday’s thesis is the most explicit and, in some ways, the most elegant.
On February 11, 2025, Workday announced its Agent System of Record. Workday described it as a place to manage agents from Workday and third parties, with controls for roles, cost, impact, and compliance. The announcement also warns that unreleased features are subject to change, so the stated architecture should not be read as proof that every capability was then generally available.
That choice of language mattered.
Workday was saying that agents should be treated less like plug-ins and more like workforce objects. They need onboarding. They need role definitions. They need scoped permissions. They need cost tracking. They need compliance controls. They need performance measurement. And they need to be managed in the same disciplined environment where enterprises already manage sensitive people and finance processes.
This is not a cosmetic distinction. It reveals where Workday believes its advantage lives.
The company does not win this market by having the most novel model. It wins by arguing that enterprise AI becomes trustworthy only when it operates inside a system that already understands organizational structure, business rules, approvals, segregation of duties, and sensitive workforce data.
That is why Workday keeps pushing the phrase “people, money, and agents” as a single platform story.
The supporting moves through 2025 made the same point. In June 2025, Workday announced a unified AI developer toolset and Agent Gateway. The company said Agent Gateway would connect third-party agents to its Agent System of Record and described Illuminate as drawing context from more than 1 trillion transactions a year. That figure is Workday’s platform-scale claim, not a measure of agent activity, and the release said Agent Gateway was planned for early adopters by the end of 2025. By early 2026, Workday had begun reporting usage and commercial measures rather than only describing an architecture.
In its fiscal 2026 results, Workday said it had more than 11,500 customers and had delivered 1.7 billion AI actions during the fiscal year. Management also reported more than $400 million in annual recurring revenue from newer AI solutions and more than 400 customers using Workday-built agents. These are company-defined commercial and usage measures. They do not show how many actions were autonomous, how outcomes were audited, or whether customers earned a positive return.
That does not prove the control-plane thesis is won. It does show that the market is not treating governed AI inside Workday as science fiction.
More important, the product evidence started to move from architecture language to workflow language.
Workday’s May 2025 agent announcement described a Self-Service Agent designed to answer employee and manager inquiries and execute routine actions. The release did not provide the 25% case-volume and 20% productivity figures previously repeated here, so this article does not use them as evidence. In March 2026, Workday launched Sana globally, positioning it as an interface that can find answers, take action, and automate workflows inside and outside Workday. These are vendor-described capabilities, but the product direction is clear: the control layer is intended to surface through an employee-facing interface rather than remain only in administration settings.
Still, Workday’s approach has limits, and those limits are exactly what make the market interesting.
Its natural strength is strongest where people data and governed business process already matter most: HR, finance, compensation, payroll-adjacent flows, internal mobility, and structured employee transactions. That is also where AI governance has real stakes. If an agent recommends a candidate disposition, updates a worker profile, changes a benefit election, or routes a mobility action, the governing system cannot be an afterthought.
Carl Eschenbach, Gerrit Kazmaier, and the broader Workday leadership team are clearly leaning into that argument. Their core pitch is not that AI should float above the enterprise. Their pitch is that AI becomes useful only when it is anchored in governed enterprise context.
The question is whether that advantage travels far enough.
A true enterprise control plane eventually has to coordinate not only HR and finance actions, but also IT support, employee service, procurement, legal requests, customer operations, and an increasingly mixed estate of partner-built and internally built agents. Workday has been moving toward that with acquisitions like Flowise and Paradox, and with its partner network expansion. But its strongest proof still comes from the domains it already owns.
That may be enough for some buyers.
For others, it may feel too system-of-record-centric for a world where the most valuable agent workflows cross departmental lines.
ServiceNow Wants to Sit Above the App Layer
If Workday’s thesis is “govern agents like workers,” ServiceNow’s thesis is “govern agents like enterprise work.”
That is an important difference.
ServiceNow does not begin with the deepest native people dataset. It begins with a different claim to power: it already sits in the middle of ticketing, service delivery, workflow orchestration, approvals, automation, and cross-functional execution. That makes it a natural candidate to become the platform where enterprises try to standardize AI governance across more than one function.
The company’s messaging has become progressively more direct about this. In May 2025, ServiceNow launched AI Control Tower as a centralized command center for ServiceNow and third-party agents, models, and workflows. In its May 2026 expansion announcement, it described discovery, observation, governance, security, and cost measurement across other clouds and applications. The company said some enhancements entered its Innovation Lab in May with general availability expected in August 2026, an important distinction between announced and available capability.
That wording tells you exactly what market ServiceNow thinks it is in.
Not a copilot market. A command-center market.
Then the company pushed further. On January 20, 2026, ServiceNow said more than 1,000 existing partners would transition into its redesigned Build Program, which was intended to help outside builders test, certify, and distribute solutions on the ServiceNow AI Platform. The count describes partners entering a program, not deployed agents or successful customer outcomes. In February 2026, just two months after closing the Moveworks acquisition, ServiceNow launched Autonomous Workforce and EmployeeWorks. ServiceNow said the combination joined Moveworks’ conversational AI and enterprise search with its portal and autonomous workflows for nearly 200 million employees; that reach figure is company-reported and should not be read as active product adoption.
That number is strategically revealing.
ServiceNow is not merely trying to be the AI layer for ServiceNow admins. It is trying to become the governed request-and-execution surface for the everyday employee.
This is where its position becomes unusually strong. In its May 2026 AI Control Tower expansion, ServiceNow said more than 100 billion workflows run on its platform each year. That company-reported scale does not measure agent-governance effectiveness or customer outcomes. It does explain the installed workflow context behind ServiceNow’s effort to compete for approvals, exceptions, escalations, and audit logs across business functions.
It also has a more credible cross-functional story than a pure HCM suite.
An Employee Service Agent, a Level 1 Service Desk AI Specialist, and a Security Operations Analyst may belong to different budget owners, but ServiceNow can argue that they all share one governance substrate. Bill McDermott and Amit Zavery are effectively selling the idea that enterprise AI should be administered like a workflow portfolio rather than purchased tool by tool.
That is attractive for CIOs, shared-services leaders, and risk teams.
It is less obviously attractive for HR teams that want control to remain anchored in workforce context rather than service architecture.
This is ServiceNow’s central tension. Its cross-enterprise position is a strength precisely because it is less HR-native. It can govern more workflows because it is not confined to HR. But that same strength can make HR buyers worry that the employee layer becomes just one workflow family inside a broader control regime owned elsewhere.
In other words, ServiceNow may be the most natural neutral ground for multi-function governance. It may also be the clearest example of why HR can lose the seat at the table once agent governance becomes a platform decision instead of a talent-tech decision.
Salesforce Wants to Own the Front Door and the Runtime
Salesforce is taking a different path, one that becomes more coherent the longer you watch it.
At first glance, the company’s AI push looked like a broad agent expansion built on CRM, Slack, Data Cloud, and Service Cloud. That is true. But by 2026, the more important detail is not just that Salesforce has Agentforce. It is that Salesforce is trying to connect three layers that enterprises often buy separately:
- the employee-facing interface,
- the service and workflow runtime,
- and the governance fabric that coordinates agents across vendors and models.
You can see that strategy in the sequence of product moves.
In May 2025, Salesforce launched Agentforce for HR Service. The practical use case was straightforward: employees could ask HR questions, update profile details, manage PTO, and track cases conversationally in Slack or an Employee Portal. Salesforce said its own HR team was seeing a 96% self-service resolution rate from Agentforce HR Service and the Employee Portal. That is a company-reported “customer zero” result from Salesforce’s own deployment, not an external benchmark.
That is not merely a chatbot metric.
It is a claim that the employee front door can absorb a meaningful share of support activity if the agent is connected tightly enough to policy, case management, and execution paths.
Then Salesforce broadened the argument beyond HR. In a February 2026 press release, it said more than 180 organizations had selected Agentforce IT Service within four months of general availability. That vendor-reported selection count does not establish deployment depth or category dominance. It does show how Salesforce framed the product against legacy ITSM and presented CoolSys, Sunrun, and Cornerstone as customer or partner evidence for a Slack-first service layer.
That sentence captures the whole market shift.
The real product is not the answer. It is the governed outcome.
Salesforce’s March 26, 2026 announcement about the U.S. Department of Labor made the same point in a higher-stakes setting. Salesforce said the department had modernized its national contact center and was rolling out DOLA, an Agentforce-based digital worker. The release also listed 2.8 million citizen support cases and inquiries, more than 9.7 million multichannel interactions, 236,000 OSHA logs, and 41,000 Job Corps applications handled by the broader Salesforce-supported service environment. It did not say DOLA autonomously resolved all of that volume. The more relevant product detail was its description of deterministic guardrails grounded in existing applications and predefined workflows.
That is governance language, not chatbot language.
The control-plane story became explicit in April 2026, when Salesforce expanded Agent Fabric. The company described automated discovery for MCP servers and other agent platforms, guided determinism for multi-agent handoffs, centralized LLM governance, agent identity, and unified monitoring. The announcement distinguished beta and planned general-availability dates, so these capabilities should not all be treated as fully available on the announcement date. Salesforce also said customers were already managing hundreds of agents across diverse platforms, a vendor-supplied adoption signal rather than an audited deployment count.
This is the crucial point.
Salesforce is no longer just trying to sell agents into business functions. It is trying to make itself the place where heterogeneous agents are discovered, governed, routed, and observed.
That matters because Salesforce’s distribution advantage is different from Workday’s and ServiceNow’s.
Workday’s natural gravity comes from the system of record. ServiceNow’s natural gravity comes from workflow coordination. Salesforce’s natural gravity comes from the interface and collaboration surface: Slack, service consoles, portals, case flows, and increasingly a unifying runtime that claims to broker actions across a mixed AI estate.
That can be powerful in enterprise buying because the front door shapes adoption. If the employee asks for help in Slack, the system that captures the request can become the system that governs the response. If the same platform also controls routing rules, agent identity, model policies, and observation dashboards, it stops being “just the interface.”
It becomes the runtime.
Marc Benioff has been selling that broader idea aggressively, and the company reports commercial traction behind it. In its fiscal 2026 results, Salesforce said Agentforce annual recurring revenue reached $800 million and that it had closed more than 29,000 Agentforce deals since launch. These company-defined measures cover Agentforce broadly; they do not isolate governance products, production depth, customer returns, or control-plane share. They support the narrower conclusion that enterprises are paying for Agentforce products attached to business workflow.
Still, Salesforce’s strongest strategy also creates its biggest challenge.
It can own the front door, the collaboration layer, and a growing governance fabric. But in many HR scenarios it still depends on another system for authoritative employee records, payroll context, org structure, and certain governed people actions. That means its control-plane ambition gets stronger as the workflow expands beyond HR into employee service, IT, citizen support, or multi-function case work. It gets weaker when the buyer insists that the most sensitive agent decisions must be anchored inside the core HCM.
Buyers Compare Control Surfaces, Not Feature Lists
This is the part of the market that many product launches still obscure.
Enterprises are not buying agent governance because they love governance. They are buying it because the cost of fragmented agent operations is rising fast.
The first cost is operational confusion.
One team launches an HR assistant in the portal. Another launches an IT agent in Slack. A third connects a model to a service workflow. A fourth experiments with a partner-built recruiting agent. At small scale, that looks like healthy innovation. At larger scale, it becomes a mess. Nobody knows which agent has which permissions, which model is being used where, which prompts or tools changed last month, or whether similar employee requests are being handled under different policies in different systems.
The second cost is risk.
Hiring, employee service, payroll-adjacent changes, and workforce support are not neutral domains. They involve protected data, approvals, employee records, notices, case history, escalation rules, and increasingly explicit regulatory expectations. If an organization cannot show who did what, why it happened, what policy applied, and how a human could intervene, the AI project stops looking modern and starts looking reckless.
The third cost is economics.
Once companies run multiple models and multiple agents, routing decisions become cost decisions. A cheap model may be adequate for a knowledge lookup but not for a sensitive case classification. A deterministic handoff may be required for a policy action but unnecessary for a FAQ. A system that cannot route intelligently ends up either overspending on premium model usage or under-governing important work.
This is why the control surface is becoming the buying surface.
The comparison now looks something like this:
| Platform | Natural advantage | Weak spot | Most credible buyer |
|---|---|---|---|
| Workday | Deep HR and finance context, governed people processes, workforce object model | Harder to dominate cross-enterprise workflows outside its native core | CHRO, HRIS, CFO, enterprise architecture teams that want governance anchored in core records |
| ServiceNow | Cross-functional workflow orchestration, service operations, centralized control posture | Less native people context; HR can become one workflow among many | CIO, shared services, enterprise operations, risk and governance leaders |
| Salesforce | Employee-facing interface, Slack distribution, service runtime, growing multi-vendor control layer | Often depends on another system for authoritative employee and payroll context | Employee service, customer service, CIO, digital workplace and transformation buyers |
This is also why “who is paying” has become inseparable from “who is governing.”
If the budget sits with HR alone, Workday often looks strongest because it can connect AI authority to the existing system of record. If the budget sits with CIO or shared services, ServiceNow becomes more attractive because it can govern AI across multiple operational domains. If the budget is framed around employee interaction, contact-center logic, and a unified request surface, Salesforce has a stronger narrative than many people assumed a year ago.
That does not mean the market splits neatly by org chart.
The most consequential deals now involve multiple buyers at once. HR wants policy fidelity. IT wants control. Legal wants auditability. Procurement wants outcome metrics. Shared services wants cross-functional reuse. Finance wants proof that agent usage is tied to lower case volume, faster resolution, or lower cost to serve. The winning platform has to satisfy all of them.
That is why the next procurement template will look less like a feature checklist and more like a governance document.
Expect buyers to ask for five things up front:
- A trusted agent identity model
- Role- and policy-based permissions
- Model routing and cost controls
- Unified action and decision logs
- Outcome reporting tied to business workflows
Any vendor that cannot answer those five cleanly will increasingly look like a point tool, no matter how polished the demo appears.
Independent HR Tech After the Control Plane
The most important implication is not that one of these three platforms automatically wins.
It is that the middle of HR tech gets squeezed again.
A standalone vendor can still win if it owns hard operational complexity, external demand, or a uniquely valuable data asset. But if its AI strategy depends on living downstream of someone else’s identity model, someone else’s approval system, someone else’s workflow runtime, and someone else’s audit layer, then its bargaining power weakens.
That is what agent governance changes.
In the old software stack, a specialist could still defend itself by owning a workflow. In the new stack, that workflow may be useful but insufficient. The enterprise increasingly wants the workflow to sit inside a governed agent environment where actions can be routed, constrained, logged, and measured. That pushes value upward toward the platform that coordinates the work.
This is why the control-plane fight matters so much for recruiting and HR service specifically.
Those categories are emotionally easy to discuss in terms of experience. Better candidate communication. Better employee support. Faster answers. More self-service. But the harder truth is that the money moves when the buyer believes the AI layer can be governed like real operating infrastructure.
That is also why HR teams should pay attention to where this layer lands.
If HR does not help define the agent governance architecture, the architecture will still be defined. It will simply be defined by someone else: CIO, security, enterprise architecture, legal, procurement, or a shared-services function that sees HR as just one workflow family among many. Once that happens, HR does not just lose budget. It loses leverage over how workforce policy gets encoded into digital labor.
The strategic risk is not only that HR buys the wrong tool.
It is that HR mistakes an interface decision for a governance decision.
A team can love a portal or a Slack experience and still lose control of the rules that govern what the agents are allowed to do. It can be impressed by an agent’s fluency and still fail to notice that approvals, logs, and model choices live somewhere else. Over time, that is how workflow ownership drifts.
The next 12 months will likely produce a market pattern that looks familiar in enterprise software. Buyers will experiment broadly at the surface and consolidate slowly at the control layer. They will tolerate multiple assistants for a while. They will tolerate multiple agent builders too. But they will become far less willing to tolerate multiple uncoordinated governance models once real workflow execution, compliance exposure, and cost management are involved.
That is where the durable market power will accumulate.
Not in the prettiest assistant.
In the system that governs the work.
Administering Digital Labor Becomes the Real Competition
The selection committee from the opening meeting will eventually buy software. It may buy more than one system. Most large enterprises will.
But the more important decision will not be which product demos best.
It will be which platform the company trusts to administer digital labor as a real part of the enterprise.
Workday wants that right because it already administers the most sensitive people and finance processes inside many companies. ServiceNow wants it because it already coordinates work across departments and can present itself as neutral governance ground. Salesforce wants it because the interface, the service runtime, and the agent fabric are starting to converge into one employee-facing execution layer.
All three arguments are credible.
None is complete.
That is why this market is worth watching now. The category is moving past the theatrical phase of enterprise AI. It is leaving behind the period when every launch could be explained by a better chatbot, a cleaner summary, or a faster draft.
The next phase is stricter.
Buyers want a platform that can tell them which agent acted, under whose authority, through which workflow, using which model, at what cost, with what result, and under what policy constraint.
Once that becomes the standard question, agent governance stops being an admin feature.
It becomes the product.
This article provides a deep analysis of agent governance in HR tech. Published April 20, 2026.