In March 2024, the SEC charged two investment advisers for making false or misleading statements about their use of artificial intelligence, the agency’s first enforcement actions against what it called “AI washing.” The firms had marketed capabilities their models didn’t have. For anyone within a decade of retirement, the lesson is smaller and sharper: the number an AI planner hands you is a marketing artifact as much as a forecast, and the confidence around it is almost always overstated.
Monte Carlo simulation is the engine underneath most of these tools. Feed it your balances, your spending, and your horizon, and it returns a headline probability of success. The number looks like certainty, and it reads that way to almost everyone who sees it, but the math behind it is far softer than the display suggests.
Before the Plan Runs, the Assumptions Are Already Doing the Work
Every Monte Carlo engine starts with capital market assumptions: expected returns, volatility, correlations, inflation. Nudge the equity return assumption down by even a fraction of a percent and a confident-looking plan can slip several points lower without a single input from you changing. Two respected tools can look at the same household and produce meaningfully different answers because the plumbing underneath is different.
AI-branded planners tend to hide this. The interface shows a clean gauge and a confident sentence. The reader almost never sees which return assumptions were used, how many paths were run, or whether spending was modeled as flat, rising, or declining. A financial planner who works through those assumptions with you in plain language is doing something the software will not do on its own.
The Moment the Number Appears, It Gets Misread
A high success rate does not mean an equally high chance of a comfortable retirement. It means that in that share of the simulated paths, the portfolio held a positive balance at the end of the horizon. One dollar at year 30 counts as a win. Running out at 79 counts as a loss identical in weight to running out at 94.
The path matters more than the tally. Two plans with the same headline score can carry radically different real-world risk, with one failing gently in the final year and the other collapsing early with no time to recover. The success rate alone will not tell you which is which.
The AI Layer Adds Explanation and Removes Judgment
Large language models grafted onto retirement tools are good at explaining outputs and often unreliable at catching the things that actually break plans. Experts interviewed by CBS News have flagged Social Security claiming nuance, tax interactions, longevity risk, and long-term care costs as areas where general-purpose AI still gets confident answers wrong.
The risks that most often destroy a retirement don’t live inside the simulation at all. A few worth naming:
- Spending shocks. A roof, a lawsuit, a child moving home, a parent needing memory care. Standard runs assume smooth, inflation-adjusted spending for 30 years.
- Sequence of returns. A weak first five years of withdrawals damages a portfolio in a way a strong average return cannot repair.
- Behavioral drift. The plan assumes you’ll rebalance in a bear market, while the historical record shows a meaningful share of investors freeze or sell instead.
- Health and longevity. Planning to age 90 when a spouse lives to 97 is not a rounding error.
In the Years After the Number, the Plan Should Move
The single most useful thing a Monte Carlo output can do is get re-run. A Kitces analysis of whether these models are actually predictive found their real value is less in the initial percentage and more in how the score moves year over year as reality updates the inputs. A plan reviewed once at retirement and filed away is close to useless. A plan re-scored annually, with spending, markets, and health folded back in, becomes an early warning system.
That reframes what the score is for. The score works as a dashboard light, one that tells you whether to keep going, ease off spending for a year, or delay a discretionary purchase until the market cooperates.
Read the Output Like a Skeptic, Not a Customer
If you’re within ten years of retirement and running these tools on your own, a few habits help:
- Ask for the assumptions. Get the equity return, bond return, inflation, and volatility numbers in writing. If the tool won’t show them, that itself is the answer.
- Stress the spending line. Re-run the plan with spending materially higher for the first decade. If the score falls off a cliff, the original was fragile.
- Look at failure timing. Ask when the failing paths fail. Early failures matter more than late ones.
- Model the shocks. Add a five-year long-term care event in your late 80s and see what the plan does.
- Re-run every year. Treat the score as a living reading, not a verdict.
AI-driven planning tools will keep getting better at explaining themselves. They are nowhere near replacing the judgment call about what to do when the number moves. The score is the beginning of the conversation. Anyone selling it as the end of one is selling certainty that the math cannot deliver.
