For the last several years, the loudest conversation around artificial intelligence has focused on job replacement. AI was going to eliminate roles, reduce headcount, and allow businesses to produce the same amount of work with far fewer people.
But that is not exactly what I see happening.
Recently, I came across a discussion on LinkedIn that described something I think a lot of professional firms are beginning to experience. The argument was simple: AI can draft more memos, which means someone now has to review more memos.
The bottleneck did not disappear. It moved.
Before AI, drafting was often the constraint. A professional had to research the issue, develop the reasoning, write the analysis, and prepare the final work product. AI can compress much of that work dramatically, but it does not eliminate the need for someone to determine whether the output is correct, appropriate, and right for the client.
What becomes scarce is no longer drafting capacity. It is judgment.
And judgment does not scale simply because you added another AI subscription.
The Work Did Not Disappear. It Changed.
This is one of the most important lessons we have learned as we have built AI workflows and an expanding skill library for tax and accounting professionals.
You can absolutely get more work done with AI, but more output creates more things that require attention. There is more analysis to review, more client situations to consider, more opportunities to evaluate, and ultimately more decisions to make.
That creates a very different kind of bottleneck.
The LinkedIn discussion made another point that I think is particularly important: reviewing AI-generated work can actually be more mentally demanding than creating the work yourself.
When you create an analysis from scratch, you understand how you arrived at the conclusion. You know which assumptions were questionable, where the reasoning became difficult, and where you need to slow down.
AI-generated work removes many of those natural signals. The output may be beautifully written. The citations may look appropriate. The structure may be clean and convincing.
The conclusion can still be wrong.
That is what makes review so important, and it is also what makes blind approval so dangerous.
Read:Â Why AI Without Context and Guardrails Will Always Hit a Ceiling in Tax & Accounting Firms
The Human Harness
I recently heard the term “human harness,” and I think it describes where this next phase of AI is heading.
The goal should not simply be to put a human “in the loop.” That phrase is becoming too broad to be useful. The more important question is how we design the entire AI workflow so the human spends time on the decisions that actually require human judgment.
A poorly designed workflow simply shifts labor around. AI creates something, the professional reads everything, checks the research, reconstructs the reasoning, looks through the client’s history, determines whether anything important was missed, and only then makes a decision.
Yes, AI saved drafting time. But much of that time may have simply been transferred into review.
That is why I believe the real opportunity is not merely AI-assisted work.
It is AI-first workflow design.
Final Answer vs. Decision Package
One of the principles we have increasingly used in designing AI skills is that AI should not simply hand a professional a polished final answer. It should prepare the professional to make a better decision faster.
Traditional AI Output
AI-First Decision Package
Produces a polished answer
Shows the answer and the reasoning behind it
Summarizes research
Surfaces the most relevant authority
Drafts a recommendation
Identifies assumptions and risk
Uses the prompt as context
Pulls in client history and intelligence
Leaves the reviewer to find weaknesses
Flags inconsistencies and missing information
Optimizes for output
Optimizes for judgment
That distinction matters.
Instead of starting with a polished document and trying to reverse-engineer how the machine got there, the professional is given something closer to a decision package.
The AI can perform the initial research, surface relevant authority, pull together the client’s history and current facts, identify assumptions, highlight inconsistencies, compare the situation with prior years, and flag areas where confidence is lower or professional review is especially important.
The machine does what machines are good at: finding, organizing, comparing, summarizing, and preparing information.
The professional does what experienced professionals are good at: interpreting context, weighing risk, understanding the client, and making the judgment call.
That, to me, is the human harness.
Read: The Real AI Race in Tax & Accounting Isn’t About Prompts
Why “Adding AI” to Software Often Does Not Solve the Problem
This is also why I think a lot of software companies are going to struggle with their current approach to AI.
Many products have simply added an AI layer to an existing workflow. There is now a button that writes something, a button that summarizes something, or a button that produces a memo.
Those features can certainly be useful, but they do not necessarily change the economics of the workflow.
If AI allows a tax professional to generate five times as much analysis, but that professional must manually review five times as much analysis, the larger productivity problem has not been solved.
One step got faster.
The constraint moved somewhere else.
An AI-first application approaches the problem differently. Instead of asking, “Where can we add AI?” it asks, “What should the machine do, what should the human do, and how should the system operate between those two points?”
That distinction is going to become increasingly important.
Autonomous Does Not Mean Unsupervised
I also believe autonomous AI skills will become a major part of professional services, but autonomous does not mean uncontrolled.
There are many tasks where AI should be able to operate with very little human involvement, including collecting documents, organizing information, running initial analysis, monitoring for changes, preparing research, identifying opportunities, and drafting routine communications.
The key is that the level of human involvement should increase as risk, ambiguity, and consequence increase.
The best workflows will probably operate on a spectrum. Some work can be completed automatically. Some can be prepared and staged for quick approval. Other work should escalate immediately to an experienced professional.
That is much more sophisticated than simply placing a chatbot inside a software application, and it is much closer to how I believe professional firms will ultimately use AI.
Read: AI Can Make You Superhuman. But It Can’t Give You Judgment.
Being Busy Is Not Necessarily Evidence That AI Failed
There was one line in the LinkedIn discussion that stuck with me. The argument was essentially that if a team adopted AI and suddenly has enormous amounts of free time, you should probably ask whether the work is actually being reviewed.
That may be a little overstated, but the underlying point is important.
AI adoption should create leverage, but leverage does not always mean people immediately work fewer hours. Sometimes it means the organization can finally do things it previously did not have the capacity to do.
A firm might review more clients, find more planning opportunities, communicate more proactively, deliver more advisory work, conduct deeper analysis, respond faster, or serve more clients without sacrificing quality.
That is still productivity.
It is simply a different version of productivity than the original “AI will eliminate half the jobs” narrative.
Human Judgment May Become More Valuable, Not Less
For tax and accounting professionals, I think this is ultimately good news.
AI is going to become extraordinarily good at producing work, which means simply producing work will become less valuable.
Judgment becomes more valuable.
Knowing which question to ask matters more. Understanding a client’s complete situation matters more. Recognizing when something does not make sense matters more. Knowing when the technically correct answer is not the right recommendation for the client matters more.
That is where experienced professionals have an enormous advantage.
The firms that win with AI will not be the ones that remove humans from every workflow. They will be the ones that become very deliberate about where humans belong and how AI prepares them to make better decisions faster.
And I think this is where the AI conversation is beginning to change.
For the last few years, much of the focus has been on prompting: how to ask AI better questions, how to write better prompts, how to get better outputs.
That was an important first step.
But the next phase is not about prompts.
It is about system design.
Which work should AI handle autonomously? Which decisions require a professional? What information should be assembled before the human ever sees the issue? What should trigger escalation? Where should confidence thresholds sit? How do research, client intelligence, workflow, and human judgment work together?
Those are much bigger questions than prompt engineering.
AI can create the leverage. Autonomous skills can perform more of the repeatable work. Client intelligence can provide context. Research can be assembled almost instantly.
But eventually, someone still has to make the call.
The future of AI in tax and accounting is not about getting humans out of the loop.
It is about designing a much better system around human judgment.
