Several young accountants I mentor have asked me how artificial intelligence might affect their careers. I wrote a short piece about that recently and admitted I did not have the reassuring answers I usually have. That is still true. But there is one answer I do have, and it comes out of work I do myself.
What I use it for
I use artificial intelligence for my own financial planning, and not for simple tasks.
The problem is not complicated because of any single piece. It is complicated because the pieces arrive on a schedule, and once they arrive they cannot be turned off.
Annuity income is already coming in, and part of what I project is where those annuities break even and what they look like well beyond that point. Social Security is deliberately delayed, which means another increase at age 70. Then required minimum distributions begin, at age 73 for those born from 1951 through 1959 and at age 75 for those born in 1960 or later, under the SECURE 2.0 Act. Stack those on top of one another and taxable income steps upward on a known timetable, whether the money is needed that year or not.
That rising income is the pressure. Income-related Medicare surcharges apply above published thresholds, and crossing one costs a full year of higher premiums. So the planning question is not what income is today. It is what income will be forced to be later, and what can be done now, while there is still room, to keep those years from being pushed across an expensive line.
Roth conversions are the lever. Converting now means paying tax voluntarily today to shrink the balance that will generate mandatory withdrawals later. Each year the exercise is to convert as much as possible up to a ceiling, and the ceiling has to account for tax already owed on everything else.
One refinement I would pass along: set that ceiling below the surcharge threshold rather than at it. Converting right to the line leaves no margin. If an unexpected expense forces a larger withdrawal later in the year, the line has already been crossed. Building in headroom costs a little conversion capacity and buys the ability to absorb a surprise.
Work like this once took me a very long time, because a multi-year projection built by hand has to be rebuilt every time one assumption changes. Now I build the model once, change a variable, and see the effect across every future year in minutes. The speed comes only after the foundation is right. Getting the structure correct takes real effort and real checking. After that, testing a different assumption is quick.
I use AI as a research, drafting, and analytical aid, but I remain responsible for reviewing the underlying assumptions, source material, calculations, and conclusions.
Where knowing the rules still matters
Here is a small example that cost me nothing only because I caught it.
Social Security is deliberately delayed in my plan, which means benefits begin partway through a year rather than in January. The model assumed a full year of benefits in that first year. It did this for my wife’s benefit as well as my own, and it did it silently. Nothing in the output indicated an assumption had been made.
That single assumption does more damage than it appears to. An overstated first year of benefits raises total income for that year, which changes how much of the benefit is taxable, which changes how much room is left for a Roth conversion. One quiet assumption moves several numbers downstream, and every year after it inherits the distortion.
The correction was simple once identified. Benefits begin in the month selected on the application, not automatically at the beginning of the year, and the model has to be told so explicitly. Even that has a second layer. Each month’s benefit is paid the following month, so the first year brings one fewer payment than the months of entitlement.
In fairness, the fault may not have been entirely the tool’s. I was still learning how to use it, and I may simply not have asked the question precisely enough. But that is the lesson rather than an excuse for it. The output was clean and confident and wrong, and it was wrong in a way that only someone who knew to look would have caught.
This is one example among many. Social Security alone carries a set of rules deep enough that the timing question is nowhere near the most complicated one. Every additional rule is another place a model can make a reasonable-sounding assumption and never mention it.
That is the whole point. My value in this work lies in knowing which questions the tool will not think to raise. My example comes from tax planning, but the same is true in audit, financial reporting, and every other area of the profession.
Where my confidence ends
I developed that knowledge the slow way. I prepared schedules, tested transactions, reconciled accounts, researched guidance, and produced first drafts that someone more senior then took apart. The work was tedious and often felt like a poor use of a trained person’s time. It was also how judgment got built. I learned what a reasonable answer looked like by producing several thousand of them, including the wrong ones.
That work is precisely what is being automated now, and the shift is likely to accelerate.
I do not expect the accounting profession to disappear.
The aggregate outlook is not alarming. The World Economic Forum’s Future of Jobs Report 2025 projects that technological, economic, demographic, and green-transition shifts together could create 170 million jobs globally by 2030 while displacing 92 million, a net gain of 78 million. OECD research points the same direction: no clear evidence yet that AI is reducing the overall number of jobs, but occupations at the highest risk of automation account for roughly a quarter of employment across its member countries, and that risk becomes real if the transition is handled badly.
But those projections answer a different question than the one I am asking. They tell us how many jobs may exist. They do not tell us how the people in them will learn to do the work well.
That is the gap I cannot close with reassurance. If the entry-level work that built professional judgment is largely automated, and judgment is still required to validate what the machine produces, the profession has a sequencing problem. We are removing the training ground while keeping the requirement it trained people for. A firm cannot eliminate the work through which judgment developed and then expect experienced judgment to appear on schedule five years later.
Firms will need to build deliberate training, review, rotation, and supervised decision-making opportunities that replace some of the learning once embedded in routine work.
What I tell them
I still cannot promise anyone their job will look the same in five years, or that the skills they build now will not themselves be disrupted. I am not going to pretend otherwise.
What I can say is that the ability to produce an answer is becoming common, and the ability to know whether an answer is right is not. That second capability is built, not downloaded. It comes from understanding the work well enough that a wrong result feels wrong before it can be proven wrong. Nothing I have seen suggests it is about to lose its value.
So I encourage the young professionals who ask me to stay curious, keep learning, protect their financial flexibility, and keep building that second capability.
The information in this article reflects my personal experiences and observations over nearly forty years in corporate finance and is intended for educational and informational purposes only. It should not be considered personalized financial, investment, tax, or legal advice. Every individual’s circumstances are different, and readers should consult qualified professionals before making decisions based on their specific situation.
Jimmy T. Singh, CPA, is an independent financial consultant and author of A Life Well Lived: From a Small Village to a Life of Purpose, Success, and Love and The Wealth You Build: A Lifetime of Financial Wisdom. He served as Corporate Controller and writes about financial independence and purposeful living.
← All articles