It’s March 2028. HR and Payroll have not disappeared.

A response to Citrini Research’s Global Intelligence Crisis memo

It’s March 2028. Every HR team delivered the workforce forecasts and finalized the performance cycle. Every payroll team completed their first run of the year with new social security thresholds and updated tax tables that came into effect on the 1st. Some CLA updates were confirmed in late December while a few only arrived in the first week of January. Annual statements were compiled and delivered. There were some minor adjustments in February. And now it’s back to the regular HR & payroll cadence.

Here’s what was different from 2026: AI functionality delivered the forecasts and created performance analyses. It also flagged most of the regulatory payroll changes. It read the legislation, identified the relevant changes, and created a summary of what needed updating and where. That part was genuinely impressive. It was faster and more thorough than any manual monitoring process. And then a payroll person reviewed it all, validated the interpretation against the specific context of her organization, made two corrections where AI got it wrong, signed off on the configuration changes, and ran the first payroll of the year. On time and correctly.

I am going to watch this happen in 2028, as I have watched the January HR & payroll activities happen every year since 2001: first in the ERP era, then through the long transition to SaaS, and now in whatever we’re calling this next phase. The work has changed. The responsibility hasn’t moved.

The memo that started a wave

Have you read the recent Citrini Research memo called The 2028 Global Intelligence Crisis? It imagines a June 2028 where all the above work has been swept away. Where AI has eliminated a huge portion of the workforce, where the mortgage market is cracking, where HR and payroll as we know them have been disrupted beyond recognition. It is a compelling piece of scenario thinking. It also made an impact: the market dropped, and software stocks lost 5% in one session. I read the memo carefully. And I think it was wrong in ways that matter, because the authors are looking at this space from the outside.

I’ve been on the inside since 2001. I’ve watched this industry move from local applications to on-premise ERP to shared ERP and then to SaaS. I’ve been in implementation projects, on vendor calls, and in boardrooms where these decisions are made. I’ve also spent the last several years as an independent advisor working across the landscape, with established vendors, with startups rebuilding HR & payroll from scratch, and with employers trying to make sense of both. And what I see in 2028 is not the catastrophe the Citrini memo imagined. It’s something more interesting, and considerably more nuanced, than that.

The assumption that was never argued

The entire Citrini Research scenario rests on a technology assumption that was stated rather than argued: that AI agents will simply work. Not just work in demos or on well-guarded tasks in controlled conditions. No, the assumption is that they work reliably, at scale, across the complexity of enterprise environments. But we know that HR & payroll systems are sometimes messy, underdocumented, supported by complex regulation, and have been accumulating edge cases for a decade or more.

By 2028, the assumption that AI agents work will have moved faster than expected in some areas and far slower in others. The gap between those two things is where the scenario falls apart: it only considers the upside. And even though the memo focuses on enterprise systems in general, I will use HR & payroll systems to illustrate this with examples in real life. Because that is where the changes happen.

Too fast to be true

Before we even get to whether AI technology works, there is a more fundamental problem with the 2028 timeline that the conversation conveniently ignored: enterprise HR and payroll systems take between 6 and 12 months to implement even when the technology is mature, well-understood, and has a decade of successful deployments behind it. Rollouts in global companies are multi-year programs with change management workstreams, parallel runs, data migration projects, and specialists whose entire job is managing the gap between how the system works and how the organization actually operates. Even the purchasing process can take months with all the governance, audit, data privacy and security questions that must be addressed. The idea that companies would simultaneously decide to standardize on a framework, rebuild on agentic AI, execute that transition successfully, and reach reliable production quality by 2028 requires a suspension of disbelief about how enterprise technology actually moves. Large organizations have never moved that fast. Not once in the history of this industry.

Now, you might argue that a new implementation isn’t necessary because Citrini Research also assumes that companies will not employ people anymore, or far fewer than they do today. They assume that the labor market will implode. They reason that when AI capabilities improve, companies need fewer workers, and so layoffs increase. And because unemployed people don’t spend money, company margins will erode, so they will invest even more in AI, which improves AI capabilities, and so on.

I have shown several times that the companies that are yelling loudly that white-collar work will be fully automated within the next 12-18 months, are also the ones that continue to hire people at scale. Their many vacancies (aka data points) tell a different story. Several people wrote rebuttals to the Citrini Research memo pointing out work is not a ‘zero sum’ game: that more technology doesn’t automatically mean fewer jobs. That so far, every major technological revolution has meant more jobs, not less. And while I do think that AI has the capability to augment workers, full replacement is not nearly as close as you might think. Let me illustrate that with an example.  

Jonathan Jebson published the results of an experiment where he built a custom GPT and ran it through 100 common compensation tasks. He simulated a company by uploading market data exports, offer data, and employee data. He then let the GPT run tasks like pulling market medians, comparing compa-ratios, and analyzing compensation positioning. The overall success rate was just 45%. It was also surprisingly slow: each prompt took between three and five minutes to complete. The outputs couldn’t be audited meaningfully. When he asked the system to explain results, it produced Python logs that no business leader can interpret. But more problematic, the moment he uploaded the data, it was frozen in time. That does not really work in an environment where offers change daily and market data moves constantly.

45% success on routine tasks, with no auditability, in a domain where a wrong number doesn’t only cause embarrassment but can also be a regulatory breach, lead to employee dissatisfaction or a fine. That outcome is surprisingly problematic. By 2028, the technology will have improved. This experiment was not a purpose-built system, and a more sophisticated architecture would perform better. But the structural problems, from data accuracy to auditability and accountability, don’t disappear because the models get smarter. They require deliberate architectural decisions that take time and expertise to get right.

The ‘jagged frontier’ problem

There’s a broader principle here that I’ve seen described as the ‘jagged frontier’ problem. AI doesn’t fail the way humans fail. It fails confidently, in unpredictable places, in ways that are hard to catch until something has already gone wrong. AI can excel in one thing and be incredibly dumb in other, often real-world problems (like the should I walk to the car wash? answer). A senior engineer put it plainly: even a perfect AI cannot know what was never communicated to it. And that means that for organizations, especially regulated ones, the problem of explainability comes into play.

Enterprise software is not just algorithms. It consists of years of edge cases, compliance workarounds, and business decisions baked into the architecture. The reason your overtime calculation looks the way it does isn’t arbitrary. It reflects a union negotiation, a ruling from the labor authority, and an exception added three years ago that nobody fully documented but everyone quietly understands and relies on. Strip away the code and that institutional knowledge is still there, still needed, and can only be explained by someone who has been paying attention.

And then there is the accountability question, which will remain unresolved in 2028. When an AI system makes a payroll error that costs an employee their mortgage payment, who is liable? The vendor? The employer? The model? This is not a philosophical question. It is a legal and commercial one that every risk-conscious organization must deal with. Legislation moves slow, but I do not expect that governments will allow organizations to point at AI and delegate responsibility. And that means a person must stay in the loop on decisions with that kind of consequences, because of exposure and liabilities.

The ERP era with better tools?

What surprised me most about the scenario was how completely it ignored a history that everyone in the tech industry has lived through. We have been here before.

In the ERP era, companies ran their own software, hosted locally, customized to their specific requirements. Every organization had its own HR system, its own payroll logic, its own integrations. And then, over about a decade, the industry moved away from that model, not because customization wasn’t valuable, but because the cost of ownership was punishing. You owned every bug, every upgrade cycle and every compliance update when the tax authority changed the rules on December 30th. And when it broke, you hired a consultancy that billed by the hour.

SaaS solved this by spreading that cost across clients. One system, continuously maintained by one company, with regulatory updates delivered in the background by a team whose entire job is to stay current with the legislation, so you don’t have to.

The Citrini Research scenario essentially proposes that we will abandon the shared software model and rebuild from scratch, but this time with AI. Every company with a few engineers and a code generation tool will build its own stack. But ask yourself a simple question: if you run a business, will you actually want to go back to owning your HR & payroll system?

I am a small business owner. My CRM connects to my bank, reconciles payments with outstanding invoices automatically, and handles my VAT submissions without me understanding the underlying rules. When those rules change, it just keeps working. I don’t have to worry about it, and that is precisely the point. My focus is on the work that generates revenue, because everything else is a cost I can’t justify.

And sure, I could code a CRM that is perfect for me. But while AI lowers the cost of building something, it does not eliminate the cost of owning it. Of having to adapt each time something changes somewhere down the line and keeping tracking of all that. The moment every company starts building its own, you haven’t killed SaaS, you’ve recreated the ERP era with better tooling and far worse institutional memory. You’re also creating a technical debt problem that compounds every quarter, because nobody will fully understand the code the agent wrote six months ago, including the agent. I’ve seen this movie before. We all have. We spent the past fifteen years getting out of it.

Or a fresh new start?

That said (and this matters!) not everyone starts from legacy. Some of the most interesting work I’ve seen in the last two years is coming from startups building new HR and payroll solutions on modern architectures, including AI, without the constraints of a decade of accumulated technical debt. They’re moving faster than incumbents can, they’re making design choices that simply weren’t available five years ago, and some of them are genuinely rethinking what the solution should be using an outcome perspective. This is real disruption, and it would be dishonest to wave it away.

There is also an alternative version of the disruption argument. It’s not that every company will build its own stack, it’s when one or two AI-native platforms achieve genuine enterprise-grade coverage and consolidate the market around them (basically a repeat of SaaS incumbents). This scenario is more plausible, and more interesting. And if it happens, it will reduce headcount. But the threshold for enterprise-grade reliability in payroll and HR is higher than almost anywhere else. Because it comes with auditability, regulatory coverage and accountability for errors with real financial consequences. The Citrini Research memo assumes that threshold has already been cleared. I don’t believe it has. And even if it does, the transition period, when a new platform is being implemented at scale, is precisely when experienced practitioners become most valuable. Someone has to manage the gap between what the system promises and what it must deliver in a specific organizational context. We know what that looks like. It’s every major SaaS migration in a shorter time.

But even these new platforms will discover that the complexity of the domain doesn’t disappear just because your stack is clean. The regulatory requirements will be the same. The edge cases don’t disappear. The moment you take on your first enterprise client, you start inheriting their history. The greenfield advantage is real and meaningful, but it is not a permanent exemption from the hard parts.

The labor shortage nobody mentioned

There is also an argument that runs in entirely the opposite direction to the worker displacement narrative, and it is one the memo ignored completely. Across the Global North, working age populations are shrinking. In my home country, the Netherlands, 60,000 people will retire each year, who can’t be replaced by young workers. The demographic pressure on organizations over the next decade is not an excess of labor that AI needs to replace but a shortage of labor that AI needs to help compensate for. The urgency driving AI adoption in HR is increasingly about capacity, not elimination. How do we do more with a workforce that is not growing? How do we retain institutional knowledge as experienced people retire? Those are the questions employers should actually ask. The answer is augmentation, not replacement. And that is a fundamentally different adoption dynamic from the one the Citrini scenario assumed.

Trust is not the same as friction

The claim in the memo that I found most revealing was this: “We had overestimated the value of human relationships. Turns out that a lot of what people called relationships was simply friction with a friendly face.”

I understand what they mean. Some of what passed for relationship value in certain industries really is just switching cost. But in HR and payroll, this is different. Trust is not friction. They are different things that sometimes look similar from outside the function.

When an HR professional calls an employee back the same day because a data entry looks wrong, that is not friction. That is accountability, the kind that an employee relies on and that an employer’s reputation is built on. When a benefits administrator walks a new hire through their health cover options in plain language because the plan documents are genuinely incomprehensible, that is not friction. That is the difference between a benefit that gets used and one that gets wasted. When a compensation analyst can stand in front of the CFO and explain precisely where a number came from and why, that is not friction. That is the kind of transparency that allows organizations to make decisions with confidence.

These functions require human judgment, human accountability, and the kind of trust that is earned through repeated reliable performance over time. You can demonstrate competence with a benchmark. You cannot benchmark your way into the room where the difficult conversation happens.

What actually changes (and what doesn’t)

I want to be clear: I am not arguing that 2028 looks the same as 2026. It doesn’t. AI will have changed work meaningfully, and I find the direction genuinely exciting.

A useful framework I’ve encountered for thinking about this comes from a conversation with Origin, a company that is building a new generation of benefits intelligence technology. They describe their architecture in three layers: the system of record, which remains the authoritative source of truth and changes relatively slowly; the system of insight, where AI is doing its most valuable work today; and the system of action, where insights are translated into decisions and outcomes.

The middle layer is where the transformation is most visible right now. Errors that used to reach employees are being caught in pre-processing. Anomaly detection has made compliance reviews faster and more reliable. Compensation teams can identify patterns in their data (e.g. flight risk, pay equity gaps, budget misses) that were theoretically available but practically invisible because nobody had the time to look. HR leaders are walking into board conversations with an analysis that in the past would have taken weeks to produce but can now be generated in real time.

This is not trivial. It is changing what it means to do this work well. The system of action layer, where AI moves from surfacing insights to actively recommending and in some cases executing decisions, is in its early development, and appropriately so. The questions of accountability, auditability, and human oversight are most acute here. But the trajectory is clear, and the professionals that build fluency with these tools today are positioning themselves well.

What AI has not done is replace the expertise required to act on what the system surfaces. AI can flag that something looks unusual in a particular cost center. It cannot determine whether that’s a data quality issue, a legitimate business change, or the early sign of something that needs an urgent conversation with a line manager. That determination requires someone who understands the context: the business, the people, the history. And that person’s value has, if anything, increased. Because now they’re spending less time on work that didn’t require their judgment, and more time on work that does.

This part is still yours

I want to end with something direct to the HR and payroll professionals reading this, because you are the audience I care about most.

For three years now, you have been bombarded with headlines telling you that your work is about to disappear. Some of those headlines came from people who have never run a payroll, never navigated a retro-active CLA update on a tight deadline, never taken the call from an employee who is frightened because something is wrong. These people look at what AI can do in a benchmark and assume it will translate directly into what AI can do in your environment. That translation is happening, but it is more deliberate, more structured, and more dependent on human expertise than the headlines suggested. And an HR AI plugin doesn’t change anything about that.

The regulatory complexity is not going away. The requirement for auditability is not going away. The trust that employees place in the people who manage the intersection of their work and their income is not going away. These are not weaknesses in your professional model. They are the reasons your professional model exists. This is also the reason the best new entrants in this space are building around them, not trying to eliminate them.

AI will continue to change what your day looks like. The insight layer will get richer. The action layer will mature. The tools will surface things you currently miss and accelerate work that currently slows you down. Engaging with that seriously, understanding which tasks to delegate, which to verify, and which to hold onto, is the professional challenge of this moment. The ones who do it well will be more effective than any previous generation in this field.

But the January run will still happen. The policy update will still need someone who understands it. The employee with a question will still need someone who can answer it with authority and care.

That part is still yours. And it is not a small thing.