AI Agents Reshape Tax Advisory Professional Judgment
The tax advisor’s core promise has always been a simple one: I will see what you cannot see. For decades, that promise rested on a foundation of accumulated experience, the kind that comes from reviewing thousands of balance sheets and recognizing the pattern that signals trouble before the client ever senses it. The professional’s value was measured in the quality of that trained eye, in the ability to ask the right question at the right moment. That model is now being tested in ways that go far beyond faster email drafting or more efficient document summarization, because the question is no longer whether machines can assist human judgment, but whether they can eventually replace the conditions that made that judgment necessary in the first place.
What happens to a profession when the preparatory work that once demanded years of apprenticeship becomes automated? When the routine examination of financial records happens continuously, invisibly, and without the need for a human to open the file? The German tax advisory network HSP GRUPPE has spent the past year confronting exactly these questions, not as a theoretical exercise but as an operational reality. With 84% of employees actively using ChatGPT Enterprise on a weekly basis and 98.6% reporting higher productivity, the organization has moved past the experimentation phase into something more consequential: a fundamental restructuring of what professional work actually consists of. [1]
The Disappearing First Draft
Consider what happens when a professional encounters a complex tax question. Traditionally, the process involved a period of immersion — reading relevant statutes, reviewing comparable cases, constructing a preliminary analysis, and only then engaging in the actual intellectual work of forming a judgment. That immersion period was not merely preparation; it was the forge in which professional intuition was shaped. The hours spent wrestling with ambiguous language and contradictory precedents were what allowed a seasoned advisor to recognize, almost instinctively, when something was off.
HSP Managing Partner Frank Heibel describes using ChatGPT as a “technical sparring partner” for working through complex tax questions and refining client communications. [1] The phrase is revealing, because a sparring partner does not replace the fighter; the fighter still throws the punches. But the nature of the training changes fundamentally. When the machine handles the initial engagement — when it processes the statutes, organizes the precedents, and presents a structured analysis — the professional’s role shifts from constructing the argument to evaluating it. The evaluator’s skill set is different from the constructor’s, and it is not clear that the former can be developed through the same kind of practice as the latter.
The distinction matters because of what it implies for the next generation of professionals. The apprentice who never spends hundreds of hours wrestling with raw material may never develop the deep pattern recognition that distinguished the master. When Partner Magdalene Posnak reduced the time required to evaluate multiple real estate investments from nine hours to around two, she freed up capacity for client advisory work. [1] But she also compressed the very process through which her own expertise was originally formed. The question that emerges is whether the next generation of advisors, raised on AI-assisted analysis from the start, will ever develop the same depth of judgment — or whether they will become skilled evaluators of machine output, competent but somehow shallower in their understanding.
The Orchestrator’s Dilemma
The deeper transformation, however, is not about individual tasks but about the entire architecture of professional work. HSP is exploring how ChatGPT Work could continuously review bookkeeping throughout the year, identify missing information, and proactively request documents from clients. The goal is that “much of the preparation had already happened before an accountant even opens the file.” This represents a shift from AI assisting humans to AI orchestrating work across entire processes — a shift that raises uncomfortable questions about where human judgment actually adds value.
Year-end accounting serves as a particularly sharp illustration. Currently, accountants often discover missing information only when they begin preparing annual accounts months after bookkeeping has been completed. The discovery triggers a scramble: requests for documents, phone calls to clients, delays, and frustration on both sides. The new approach would eliminate much of that scramble by continuously monitoring the bookkeeping and flagging gaps in real time. The accountant would open a file that is already substantially prepared, with only the genuinely complex judgment calls remaining.
On the surface, this seems unambiguously beneficial. Who would defend the inefficiency of discovering missing information months late? But the orchestration model also removes a specific kind of professional experience — the experience of encountering the unexpected, of discovering that the client’s records tell a different story than expected, of having to dig deeper because something does not add up. The accountant who only encounters pre-processed information may never develop the instinct for what to look for in the first place. The ability to recognize anomalies is built through exposure to anomalies, and the new model systematically reduces that exposure.
The Capacity Paradox
HSP’s own numbers reveal the tension. The organization identified 88 opportunities for automation, but Managing Director Carsten Schulz suggests that “that’s already thinking too small.” The real question, he argues, is not how to automate today’s processes but how to redesign them entirely. This ambition is reflected in the capacity numbers: a conservative internal scenario estimates approximately 40,000 hours of additional annual capacity, with around 28,000 hours potentially supporting billable specialist work and roughly 12,000 hours supporting administration, client service, and support. Based on conservative hourly rates, HSP estimates a theoretical annual revenue potential of approximately 3.8 million euros — explicitly labeled as a capacity scenario, not realized or guaranteed revenue.
The paradox emerges when considering what that capacity is actually for. HSP is clear that the economic benefit is not about reducing headcount; the firms in the network are already operating at capacity and managing substantial backlogs. Time saved can be redirected toward additional client work, shorter turnaround times, more advisory capacity, and stronger client service. But this assumes that the demand for advisory services will continue to grow at a pace that absorbs the newly freed capacity. What if the same technology that creates the capacity also reduces the need for it? What if clients, armed with their own AI tools, require less hand-holding on routine matters?
The question is not hypothetical. As AI tools become more accessible to end users, the boundary between what clients can do themselves and what requires professional intervention will shift. The advisory services that HSP hopes to expand into may themselves become candidates for automation in a subsequent wave. The capacity that AI creates today may be the capacity that makes the next round of automation possible — a treadmill effect that keeps professionals running just to stay in place, while the actual value of their work is progressively commoditized.
The Judgment That Remains
None of this is to suggest that professional judgment is about to become obsolete. The legal and tax professions are built on accountability structures that require a named human to take responsibility for advice given. HSP is explicit that “professional review and final responsibility always remain with the relevant tax, legal, or accounting professional.” The question is not whether the human will be in the loop, but what kind of loop it will be. If the human’s role is reduced to approving machine-generated recommendations, the judgment required is of a different order than the judgment required to construct the analysis from scratch.
The distinction between constructing and approving is subtle but consequential. The constructor must engage with ambiguity, weigh competing interpretations, and make choices under uncertainty. The approver faces a different cognitive task: evaluating whether a proposed solution is acceptable, without having gone through the reasoning that produced it. Research in decision-making suggests that this kind of evaluation is susceptible to its own biases — automation bias, complacency, and the tendency to accept machine output that appears reasonable without deep scrutiny. The approver may be more efficient, but may also be more vulnerable to errors that the constructor would have caught through the sheer effort of working through the problem.
There is also the question of what happens when the machine is wrong in ways that are not obvious. The tax code is not static; it changes, and the interpretations of it shift with new rulings and new regulatory guidance. AI systems trained on historical data may not capture the full nuance of current law, and the professional who has not engaged deeply with the underlying material may not recognize the gap. The result could be a new kind of professional malpractice, one that arises not from ignorance but from misplaced trust in a system that appears competent but is not.
The Redesigned Workflow
HSP’s exploration of ChatGPT Work for continuous bookkeeping review illustrates the potential and the peril of the orchestration model. The system would not merely assist with individual tasks but would manage an entire process over time, identifying missing information, requesting documents, and preparing the groundwork for the accountant’s review. The accountant would then focus on the genuinely complex judgment calls, the ones that require human expertise.
This is the vision that Schulz articulates: “Once you understand how everything fits together, AI opens the door to a completely new world.” The new world is one in which professionals spend less time on preparation and more time on application. But it is also a world in which the preparation itself — the grinding, repetitive, often tedious work that once served as the foundation of professional development — is no longer part of the professional’s experience. The apprentice who never learns to prepare a file from scratch may never understand what goes into a well-prepared file, and may therefore be unable to recognize when the machine’s preparation is inadequate.
The risk is not that AI will make professionals obsolete in a dramatic, headline-grabbing way. The risk is more subtle: that the profession will gradually hollow out from within, as the tasks that once built expertise are automated away, leaving a generation of professionals who are skilled at managing AI systems but less skilled at the underlying substance of their work. The judgment that remains will be real, but it will rest on a thinner foundation than the judgment of previous generations.
The Question of Purpose
HSP reports that 79.7% of employees experienced better client service and 78.1% reported higher job satisfaction. The job satisfaction number is particularly interesting, because it suggests that employees are not feeling threatened by the technology but are rather finding their work more meaningful. This may be because the AI is handling the parts of the job that were least satisfying — the repetitive preparation, the document chasing, the administrative overhead — and leaving the parts that drew people to the profession in the first place: the intellectual challenge, the client relationships, the sense of making a difference.
But the satisfaction may also be a temporary phenomenon, a reflection of the novelty of the technology and the relief of having the drudgery removed. The deeper question is whether the profession can sustain that satisfaction as the technology matures and the boundaries of what is automated continue to expand. When the AI handles not just the preparation but also the analysis, and when the professional’s role is reduced to approving recommendations, will the job satisfaction persist? Or will the profession find that it has automated away not just the tedious parts but also the meaningful parts?
The question is not whether AI will make human judgment superfluous in some absolute sense. The question is more specific: which parts of human judgment are actually valuable, and which parts are merely artifacts of a particular way of organizing work? If the value of the tax professional lies in the application of expertise to client situations, then the automation of preparation and analysis may be a net positive. If the value lies in the depth of understanding that comes from doing the preparation and analysis oneself, then the automation may be a net negative, even if the immediate metrics look good.
The Uncomfortable Conclusion
The HSP case suggests that the answer is not clear-cut. The organization has achieved remarkable adoption rates and productivity gains, and its employees report higher satisfaction and better client service. The capacity numbers are real, and the potential for redirecting that capacity toward advisory work is genuine. But the case also reveals the uncomfortable conclusion that the goal of professional work — the exercise of judgment — may be undermined by the very tools that make it more efficient.
The professional who no longer needs to spend nine hours evaluating real estate investments may use the saved time for client advisory, but the professional who has never spent nine hours evaluating real estate investments may not have the depth of understanding to provide meaningful advisory. The judgment that remains is real, but it is a different kind of judgment — more evaluative, more strategic, less grounded in the granular details of the work. Whether that judgment is sufficient, and whether it can be developed in professionals who have never done the granular work, is an open question.
The answer may be that the profession will need to find new ways to develop professional judgment, ways that do not rely on the repetitive preparation that AI now handles. This could involve more case-based learning, more simulation, more deliberate practice on synthetic problems. Or it could involve a more deliberate cultivation of the evaluative skills that remain — the ability to question machine output, to recognize when the AI has missed something, to know when to override the recommendation. These skills are not the same as the skills of the traditional professional, but they may be equally valuable.
The End of the Beginning
What is clear is that the transformation is not about individual features or tools. Schulz’s refusal to name a favorite AI feature is instructive: “Individual features don’t matter. Once you understand how everything fits together, AI opens the door to a completely new world.” The new world is one in which the professional’s role is fundamentally different — less about constructing analyses and more about orchestrating outcomes, less about processing information and more about applying judgment to situations that the machine cannot fully handle.
Whether that new world is better than the old one depends on what the profession values. If it values efficiency, capacity, and client service, the new world is clearly superior. If it values the depth of understanding that comes from years of grinding through the details, the new world is more ambiguous. The answer is probably both: the profession will gain in some dimensions and lose in others, and the net effect will depend on how it adapts.
The uncomfortable truth is that the question of whether AI makes human judgment superfluous cannot be answered in the abstract. It can only be answered in practice, through the experience of professionals who are living through the transformation. HSP’s experience suggests that the answer is not yet clear. The organization is already preparing for the next phase with ChatGPT Work, piloting the platform with a small group of developers and administrators before expanding across the wider workspace. The goal is to understand how agentic AI can safely automate complex workflows while balancing governance, quality, and cost. The question of whether the human professional remains essential to those workflows is not yet resolved.
The story of HSP GRUPPE is a story of a profession in transition, caught between the efficiency gains of automation and the preservation of the judgment that defines its purpose. The numbers are impressive: 84% weekly active usage, 98.6% higher productivity, 500,000 conversations in six months, an estimated 40,000 hours of additional annual capacity. But the numbers do not capture the more subtle transformation — the shift in what it means to be a professional, the change in how expertise is developed and applied, the redefinition of the boundary between human and machine judgment. That boundary is not fixed; it is being negotiated in real time, in firms like HSP, by professionals who are discovering that the tools that make them more productive are also changing what they are productive at.
The question that remains is whether the goal the profession is pursuing is the right one. If the goal is simply to process information faster and serve more clients, the AI revolution is an unambiguous success. If the goal is to preserve and deepen the judgment that distinguishes the professional from the machine, the revolution is more ambiguous. The answer will emerge over time, as the next generation of professionals — raised on AI-assisted work from the start — reveals whether they have developed the depth of understanding that their predecessors had, or whether they have become something different: skilled orchestrators of intelligent systems, competent and efficient, but somehow less than the professionals they replaced. The door to a completely new world is open, and what lies on the other side will be defined by the choices professionals make today.
Sources
1. HSP GRUPPE