An executive had 5,000 incentive stock options. The exercise price was $20 a share. The stock was trading at $80. She had been told for years that the smart move was to exercise and hold: buy the shares at $20, start the clock, and if she held them at least a year, the eventual gain would be taxed at long-term capital gains rates instead of ordinary income rates.
She wanted to check the logic before she acted. So she described the situation to an AI tool and asked whether she should exercise now to start the holding period.
The answer was good. It explained the difference between long-term capital gains rates and ordinary income rates. It confirmed that holding the shares for the required period would qualify the gain for the lower rate. It walked through the math on the spread and noted that exercising sooner started the clock sooner. The reasoning was clean and correct. She exercised all 5,000 options.
The spread was $300,000. She did not sell any shares, so none of it was cash. It was paper gain on stock she now owned. In April, her accountant delivered the part of the picture the AI never mentioned. The exercise had triggered the alternative minimum tax. The bill was more than $90,000, due in cash, on a gain she had not sold.
The advice was right about the part it could see
This is the detail that makes the story worth telling. The AI was not wrong. Long-term capital gains treatment is a real and valuable benefit. Exercising to start the holding clock is a legitimate strategy that good advisors recommend constantly. Every individual thing the AI said was accurate.
The problem lived in what the answer left out. For regular income tax, exercising an ISO and holding the shares is not a taxable event. The alternative minimum tax treats it differently. The $300,000 spread between the $20 exercise price and the $80 market price becomes a preference item that counts as income for AMT purposes in the year of exercise, whether or not a single share is sold. That is the rule that generated the $90,000 bill. The AI said nothing about it.
A missing consequence is the hardest kind of error to catch, because there is nothing on the page to question. The reader sees a confident, accurate explanation and has no way to sense the weight of what was never written down.
It agreed because she wanted a yes
She came to the question with a plan she already liked. People had been recommending exercise-and-hold for years. She was not really asking whether to do it. She was asking the AI to confirm that she should, and an AI is built to confirm.
We covered the mechanism behind this in another piece. These tools are trained to produce the answers people prefer, and people prefer agreement. Hand one a plan you are leaning toward, framed as a question, and it will tend to build the case for it. That is what happened here. She presented a reasonable strategy and asked for a yes, and the AI gave her an articulate, well-supported yes. The agreement felt like validation. It was the tool doing what it was designed to do.
The one input that would have changed the answer
Whether to exercise ISOs, and how many, is a tax question before it is anything else. Answering it correctly requires three things the AI was never given: her total income for the year, any AMT credit carryforwards she had from prior exercises, and her full regular tax picture. Without those, the AMT exposure cannot be calculated. With them, it can be calculated precisely.
An accountant holding that information would have started somewhere different. The first question is not whether to exercise. It is how many options she could exercise this year before the AMT bill begins, the figure planners sometimes call the AMT budget. The answer might have been 2,000 shares this year and the rest next year. It might have been a recommendation to exercise in a year when her other income made the AMT math work in her favor. It might have surfaced a carryforward credit that offset part of the liability. Every one of those paths needed numbers the AI did not have and did not ask for. It answered the general question well and stayed silent on the specific one that controlled the outcome, because it had no idea the specific one existed.
Why this is worse than a bad email
If you ask an AI to draft an email and the draft is wrong, you fix it and send it. The cost of a bad answer is an edit.
Some financial decisions do not allow an edit. The exercise was one of them. Once she bought the shares, the AMT preference item existed for that tax year, and there was no taking it back. The same finality applies to other moves that look routine in the moment. Surrendering a cash-value life insurance policy is permanent. Completing a 1035 exchange into a new policy cannot be reversed once it closes. A property sale structured and signed is done. With decisions like these, an agreeable wrong answer does not cost you an edit. It costs you the full consequence, and the consequence stands.
This is the rule worth carrying: the more permanent the decision, the more expensive a confident wrong answer becomes. The exercise of an option and the surrender of a policy are exactly the decisions where people most want a quick second opinion, and exactly the decisions where a quick agreeable answer does the most damage.
The validation looked identical to real analysis
Stand where she stood that day. She had a thoughtful question, a sound strategy, and an answer that addressed her question accurately and confidently. There was no error to spot. There was no warning sign. The answer that walked her into a $90,000 bill was indistinguishable, in the moment, from the answer a great advisor would have started with before turning to the numbers that mattered.
That is the trap in a single sentence. A polished validation of a decision you already favor reads exactly like rigorous analysis, right up until the consequence the analysis would have caught arrives in the mail.
The seam
Exercising an ISO is a portfolio decision and a tax decision at the same moment. The portfolio logic, lock in the lower rate, start the clock, reduce concentration over time, is real, and we have written about it in the context of concentrated stock. The tax logic, how much can be exercised this year without triggering AMT, is equally real. Both calculations have to happen together, before the shares are bought, by someone who can see both at once.
The AI saw one. It saw the portfolio logic, explained it accurately, and had no view of the tax calculation running alongside it. The value was never in knowing what an ISO is. The AI knows what an ISO is, and so did she. The value was in the person who runs the AMT math against an actual tax return before the irreversible thing is done.
Before the next irreversible decision
The question to sit with is not whether the AI gave a good answer. It often will, about the part it can see. The question is who checked the numbers the AI never had, before the decision became permanent. For anything you cannot undo, that question is the whole game.
