The Software Debate Through Workday’s Eyes
February 27, 2026
Periods of confusion can be valuable for the curious. Confusion can cause those with deep knowledge in the area to start talking and sharing more. And the good kind of talking and sharing where they’re trying to clarify their thoughts in real time.
We are in one of those moments right now. AI continues to progress. OpenClaw and Claude Cowork marked another ChatGPT-type moment. Agentic advancements led to new levels of confusion for many, particularly in software. When behaviors suggest that nobody fully understands, you sometimes get more candor, deeper explanations, presentations, and interviews where management teams wrestle with the question out loud.
There have been several recent earnings calls where management has attempted to cut through the noise and clarify why they believe their respective software companies are well positioned for the future. But Workday’s was the cleanest we’ve seen in framing the current software debate and boiling it down to what matters most.
It started with Co-Founder Aneel Bhusri who is now back in the CEO seat. He said:
[See attached PDF for quote.]
This framed a couple of key questions within this debate: Just how hard is it to build systems of record? And how much does this type of domain expertise matter? In Workday’s case, the reality is these systems must process payroll accurately to the penny across many countries with different tax codes, labor laws, and regulatory requirements. They must enforce security models that determine who can see what data (e.g., manager can see their direct reports’ compensation but not their peers’). They must comply with statutory requirements that change constantly across jurisdictions. They must handle benefits administration, financial close processes, audit trails, and regulatory reporting with zero tolerance for error. The complexity is in the interaction of thousands of rules, exceptions, and compliance requirements across geographies and industries, accumulated and refined over decades.
The pushback tends to be: Even if you can’t vibe code this type of system because it is too complex for that today, is it just a current limitation? When Aneel left PeopleSoft to start Workday, incumbents made similar arguments about why cloud-based systems couldn’t handle the complexity and security requirements of enterprise HR and finance. They were wrong because the new architecture eventually caught up on the hard parts while delivering advantages the old architecture couldn’t match. The question is whether AI follows the same pattern on a compressed timeline or whether the deterministic requirements of these systems represent a harder barrier than the cloud transition did. At the very least, it’s probably a multi-year effort.
The next question in this debate is: well, then what does better software look like? We all know by now that AI’s progress means the software we use will allow us to be far more productive. If agents can’t fully take over, at least for a while, what does progress look like from here? Aneel said:
[See attached PDF for quote.]
Now he starts getting into the role of agents, which covers the next section of the current debate. It’s widely agreed that both the system of record / truth underneath and agents that help improve user experience and business process automation are important. But whose agents can do this and whose will work best? Particularly for agents that are more focused on user experience and will go across apps like Workday HCM and Financials, will it be first-party agents from Workday or third-party agents from a Claude Cowork-type model that win?
Annel said:
[See attached PDF for quote.]
And Gerrit Kazmaier, Workday’s President of Product & Technology, said:
[See attached PDF for quote.]
Clearly Workday believes first-party agents will play a major role and is all-in on winning the agents that can both automate work within its apps and go across multiple apps. And Workday believes that executing here will allow for accelerating growth.
Being able to articulate a plan for accelerating growth is central to the debate on whether the current pacing around the ~$200bn (and growing) in annual capex from Amazon, ~$185bn from Google, etc. still makes sense. We likely need to see widespread and effective consumer and enterprise adoption of agentic workflows (due to the higher levels of productivity it can unlock) within the next 12 months to avoid valid “plateau” concerns. This will require more companies like Workday (or disruptors) to demonstrate that enterprise software spend can accelerate due to winning AI agent product releases.
Next, Gerrit added:
[See attached PDF for quote.]
Workday is hoping its customers adopt its agents not just to find information and automate workflows within its own apps but also across third-party apps as well. Why do Workday agents have a right to win workflows that involve third-party apps? Why don’t Claude and other frontier labs instead have the right to win both the workflows that go across apps and within apps like Workday? This is important.
Gerrit said:
[See attached PDF for quote.]
Workday should have the right to win these types of agentic workflows if only its agents have access to the underlying data. But could Claude (and others) do the same or do it better and get there quicker IF they had access to the same underlying data? Workday is incentivized not to give third party agents this type of access, but are new entrants that start with a clean sheet of paper incentivized to provide the frontier labs with full access? If so, how long would it take new entrants to build up the required domain expertise and functionality at the system of record level?
Frontier models like Claude are already exceptionally good at reasoning, language understanding, summarization, and working through complex logic. If you gave it a well-structured prompt with the right context, it could draft a performance review, analyze a financial close checklist, or help plan a workforce restructuring. It could reason about HR and finance problems at a high level. The gap Workday refers to throughout this earnings call is about access, context, and authority.
Access: Workday’s agents operate inside a system that already contains the actual employee records, compensation data, benefits elections, org charts, financial transactions, journal entries, and payroll runs for a given company. Claude sitting outside that system would need to be given all of that data every time it is asked to do something, which raises security, compliance, and integration challenges. Workday’s agents don’t need to be given the data.
Context: Gerrit described 70,000 core business process archetypes instantiated across thousands of customer variations. That means Workday has both the data and the encoded logic of how companies run their HR and finance operations. When a Workday agent handles a financial close process, it understands the specific sequence of approvals, the compliance rules, the journal entry structures, and the exception handling procedures for that particular customer’s configuration. A general-purpose model would need all of that context reconstructed from scratch. It would be reasoning about HR and finance in the abstract rather than operating within the specific reality of how a given company works.
Authority: A Workday agent can execute. It can change an employee’s benefits election. It can post a journal entry. It can move a candidate through a hiring pipeline. It has permissioned write access to the system of record. Claude can tell you what should happen. Workday’s agent can make it happen, within the security and approval frameworks the company has already configured.
What is not commoditized is the data substrate, the process logic, and the permissioned execution layer. Those are the pieces that have taken 10+ years to build and that customers cannot easily rip out and replace.
The risk, of course, is Claude and other general-purpose AI platforms ultimately being “pulled in” and connecting to those data and execution layers from the outside. If Claude or a similar model could securely access Workday’s data, understand a company’s specific process configurations, and execute actions with proper permissions, the value of Workday’s proprietary agent layer would erode. Another wrinkle is the legacy platforms that Anthropic, OpenAI, etc. would be disrupting are their customers. The general-purpose AI platforms must be mindful of competing with key customers, particularly prior to having a clear line of sight into how they would obtain the access to the data substrate and process logic they’d need to offer winning agents within these domains. These questions also came up in Workday’s Q&A:
[See attached PDF for quote.]
Workday’s responses highlight the core obstacle for disruptors to win these types of domain specific agentic workflows today. LLMs like Claude are trained in general knowledge. They know broadly how payroll works, what tax withholding is, what FMLA leave entails. But it does not know that Company X’s policy is that vacation days roll over up to 40 hours in California but expire entirely for employees in Germany, that the rollover resets on the employee’s hire date anniversary rather than the calendar year, that employees who transferred from the London office retain their UK accrual schedule for 18 months under a legacy policy, and that all of this interacts with a union agreement that overrides the standard policy for hourly workers in three specific facilities. That is one company’s vacation policy. Multiply it by every HR and finance process across over 10,000 customers, each with their own configurations, exceptions, union agreements, regulatory jurisdictions, and historical policy changes. That is what Workday’s business process framework encodes.
There’s a structural barrier today that prevents general-purpose AI labs from replicating enterprise systems like Workday. It lies in the distinction between raw data and institutional logic. While an LLM can process data points, it lacks the “Business Process Framework” required to interpret them within a stateful, deterministic environment. In addition, replicating Workday’s integrated domain security model is a big undertaking; a third-party AI “dumping data” into a model cannot inherently understand the complex web of permissions that prevent a peer from viewing a manager’s salary, for example.
For an LLM to handle this well, it needs access to the company-specific rules and configurations. Today these live inside Workday’s system. An LLM would need structured access to all of this configuration data for a given customer. It would need the rules engine: which policies apply to which employee populations, what the approval chains look like, what the exception handling logic is, how different jurisdictions interact. This is technically solvable. If a company exported or exposed its full configuration logic through APIs, an LLM could ingest it. The question is whether that configuration logic is currently accessible outside Workday’s system in a structured, complete way. For most customers, it probably isn’t. It lives inside Workday’s proprietary framework. But it’s not impossible to extract or reconstruct. And perhaps in a world of agentic “screen-reading” or advanced ETL (Extract, Transform, Load) tools, AI can “watch” how a company handles its German vacation rollovers for six months without needing the formal API to reconstruct the rule-base. But then it would need to do the same for the next customer, and so on, all while these customers already have software that does the job.
Then there’s deterministic execution within probabilistic reasoning that was discussed. Payroll cannot be 99.7% accurate. But this is a framing problem as much as a technical one. The solution isn’t to have the LLM guess the payroll amounts. The solution is to have the LLM orchestrate deterministic computation. The model reasons about which rules apply, which tax tables to use, which deductions to calculate, and then calls deterministic functions that execute the math with precision. This is already how tool use appears to work in AI systems. The model decides what to do, and structured tools execute with exactness. The model doesn’t calculate the tax withholding itself. It calls a tax calculation function with the right parameters. This architecture, which is probabilistic reasoning orchestrating deterministic execution, seems to be advancing rapidly. It’s essentially what Workday itself is building with its agents. The question is whether this architecture requires Workday’s specific process framework or whether it can be built on top of a well-structured alternative.
Then there’s regulatory and compliance knowledge. Payroll in Germany follows different rules than payroll in Brazil. Tax withholding in California differs from Texas. Statutory reporting requirements change constantly. Workday maintains this knowledge through continuous updates from a large team of domain experts who track regulatory changes across dozens of countries and update the system accordingly. For an LLM to replicate this, it would need either continuously updated training data on regulatory changes across all relevant jurisdictions, or access to structured regulatory databases that are kept current. Neither exists in a comprehensive, machine-readable form today. But companies like Thomson Reuters, Wolters Kluwer, and government agencies maintain much of this information. If that information were structured into machine-readable regulatory APIs, which is a direction the industry could be moving, an LLM could consume it. This could be the layer where the timeline is longest because it requires not just AI capability but an ecosystem of regulatory data providers making their content available in formats AI can reliably consume.
Then there’s the testing and auditability requirement. Enterprise systems need to prove they got the right answer. Every payroll run needs an audit trail. Every financial close needs documentation of which rules were applied. Regulators need to examine the logic. This means the AI system can’t be a black box. It needs to show its work in a way that satisfies auditors and regulators. This is a real constraint for current LLMs, which don’t naturally produce auditable reasoning chains that map to specific regulatory requirements. But it’s a solvable engineering problem. If the system is designed so that the LLM’s decisions are logged as structured rule applications like “applied German tax table 2026-Q1, line 43, to employee classification B2”, auditability becomes possible.
So how would Workday’s intelligence layer stop mattering? You’d need all these things to converge. Full access to company-specific configuration logic, or a new system that captures it from scratch. Reliable orchestration of deterministic execution by probabilistic models. Comprehensive, machine-readable regulatory databases that stay current. And auditable AI reasoning that satisfies enterprise compliance requirements. Each of these is individually solvable. None of them is solved today at the scale and reliability required for enterprise deployment. And they all need to work together seamlessly, which is a systems integration challenge on top of the individual technical challenges.
It seems Gerrit is right that this is “utterly unviable” today. The full stack does not exist today. At the same time though, the threat probably doesn’t need to be complete to start mattering. It probably doesn’t need to handle every edge case across 40 countries on day one. It needs to be good enough for a 200-person company operating in two countries to choose it over Workday at a fraction of the cost. Then good enough for a 1,000-person company in five countries. The disruption would likely start at the bottom of the market and work its way up, which is the pattern Workday itself followed against PeopleSoft and Oracle.
The scenario Workday should be most worried about probably isn’t a frontal assault on their largest, most complex enterprise customers. It’s new entrants like Darwinbox with MCP integrations, or an AI-native startup we don’t know yet, winning SMBs with a simpler, cheaper, AI-first system that handles 80%+ of use cases, then gradually expands the scope of what it can do.
A lot of this comes back to some of the most important pillars of business sustainability like internal talent, agility, and culture of innovation. How fast and effectively can Workday (and its peers) deliver the types of agents that work within specific apps and that can go across its apps in a way that allows thousands of customers to get to quick yeses? Incumbents also face innovator’s dilemma issues that could lead to suboptimal product and pricing decisions. That said, Workday’s customers would be happy for it to win the agentic layer within its areas of domain expertise because this would be more convenient than having to identify and work with additional third-party software providers. But with all signs pointing to agents being ready to add huge productivity soon, its customers likely won’t have much patience for anything short of highly effective agents at affordable prices.
Legal Disclaimer
DENMARK Capital Management LLC (the “General Partner”) is not registered as an investment adviser with the Securities and Exchange Commission (“SEC”) or any state securities authorities. The limited partnership interests (the “Interests”) in DENMARK Capital Partners LP (the “Fund”), are offered under a separate private offering memorandum (the “Offering Memorandum”), have not been registered under the Securities Act of 1933, as amended (the “Securities Act“), nor any state’s securities laws, and are sold for investment only pursuant to an exemption from registration with the SEC and in compliance with any applicable state or other securities laws. Interests are subject to restrictions on transferability and resale and may not be transferred or resold except as permitted under the Securities Act and applicable state securities laws. Investors should be aware that they could be required to bear the financial risks of this investment for an indefinite period of time.
All information contained in or derived from this letter is proprietary to the Fund and the General Partner.
A prospective investor should only commit to an investment in the Fund if such prospective investor understands the nature of the investment and can bear the economic risk of such investment. THE FUND IS SPECULATIVE AND INVOLVES A HIGH DEGREE OF RISK. AS A RESULT, AN INVESTOR COULD LOSE ALL OR A SUBSTANTIAL AMOUNT OF ITS INVESTMENT. In addition, the Fund’s fees and expenses may offset its profits. There are restrictions on withdrawing and transferring interests from the Fund. In making an investment decision, you must rely on your own examination of the Fund and the terms of the Offering Memorandum and such other information provided by the General Partner to you and your tax, legal, accounting or other advisors. The information herein is not intended to provide, and should not be relied upon for, accounting, legal, or tax advice or investment recommendations. You should consult your tax, legal, accounting or other advisors about the matters discussed herein.
PAST PERFORMANCE IS NOT INDICATIVE OR A GUARANTEE OF FUTURE RESULTS. This document may present past performance data regarding prior/other investments, funds, and/or trading accounts managed by the General Partner and/or the Principal. This is presented solely for explanatory purposes. The Fund may face risks not previously experienced or anticipated by the General Partner and/or Principal, and therefore, prospective investors should evaluate the Fund on their own merits.
Certain information contained in this document constitutes “forward-looking statements” which can be identified by use of forward-looking terminology such as “may,” “will,” “target,” “should,” “expect,” “attempt,” “anticipate,” “project,” “estimate,” “intend,” “seek,” “continue,” or “believe” or the negatives thereof or other variations thereon or comparable terminology. Due to the various risks and uncertainties, actual events or results in the actual performance of the Fund may differ materially from those reflected or contemplated in such forward-looking statements. The General Partner is the source for all graphs and charts, unless otherwise noted.
This document may also present “sample holdings” or “case studies” of a type of asset(s) the Fund may invest in or are expected to be typical of its holdings. Such “sample holdings” are not currently holdings of the Fund and are presented solely for explanatory purposes. Prospective investors should not assume that such “sample holdings” will actually be purchased by the Fund when determining whether to make an investment in the Fund.
