I Asked AI to Build Me An Early Retirement Plan. Here is what I Got.

 

Mo Shouman​

​

I Asked AI to Build Me An Early Retirement Plan. Here is what I Got. This is the 28th month for Wealth Bytes and I’m writing this issue on a beach in Bali. Three weeks in, working around four hours a day, which is about as close as I've come to actually honouring a promise I made myself a few years ago: six to eight weeks a year working from somewhere I actually want to be, not a five-day holiday bolted onto a full client load. This year I'm ahead of schedule for once.

​

It's also, as it happens, the year I was named one of the Top 50 Influential Financial Advisers in Australia. A genuinely humbling nod after more than 20 years in this business. I mention both because they're connected. Stability, then clarity, then options, then the freedom to choose when, where and how you work - or don't - is the entire point of the Confident Choice System, and it felt a bit hollow building it for clients for two decades without ever proving I could live inside it myself.

 

The extra headspace this trip freed up is also where this issue came from. A few weeks before I left, I'd been wanting to properly test something instead of just talking about it: what actually happens if you hand a real client profile to an AI tool and ask it to build the retirement plan?

 

So I sat down and pretended to be one of my own clients.

 

I opened Claude. One of the better AI tools on the market right now and gave it a five-line prompt: 46- year-old software sales director at Microsoft, Sydney, married with two kids, income $400k–$600k, an investment property with a mortgage, cash sitting in an offset account. Then I asked the question I get asked more than any other in my career: how much do I need to retire early?

 

What came back was three pages. Tables. Four retirement ages. Capital targets to the thousand dollars. It knew the preservation age. It knew Division 293 applied. Genuinely, some of it was better than what you'd get from a generalist adviser who's never worked with a tech professional's pay structure.

 

Then I checked the maths.

 

What It Got Right

Fair's fair. The preservation age was correct. It split the strategy into money needed before 60 (outside super, taxed on the way) and money needed after 60 (inside an account-based pension). That's the right architecture for early retirement in Australia, and close to the one I use myself. It correctly flagged that Division 293 applies. And it picked up that $400,000 sitting in "offset" was actually parked against the wrong loan. If I'd stopped watching there, I'd have told you to cancel your next adviser meeting.

 

I didn't stop there. I looked deeper and here what I found.

 

Catch #1: The Foundation Number Was Never Asked For

The plan built a full table of "super needed at each retirement age", $1.89m, $2.05m, $2.3m, $2.49m, with no idea what this client's actual super balance was. It was never in the prompt. Further down, buried, it noted that retiring at 55 "requires you to already hold $1.2–1.3 million" in super. An assumption, not a fact, and one that moves the retirement date by up to five years depending on whether it's true. The single biggest number in the entire plan was invented, then softened into a caveat instead of a stop sign or a follow up question to understand my personal circumstances.

 

Catch #2: The $41,000 Tax Saving That Was Actually $22,100

This is the one I want you to sit with. The plan recommended sweeping five years of unused concessional cap into a single $130,000 contribution, taxed at 15%, for an "immediate tax saving" of $41,000. Two paragraphs later, the same response correctly stated that Division 293 pushes this client's effective contributions tax to 30%, not 15%. Run $130,000 at his 47% marginal rate versus 30% - not 15% - and the saving is $22,100. Eighty-eight per cent less than headlined. The model contradicted its own maths and never went back to fix the number.

 

It gets worse: the strategy likely doesn't exist for him at all. His employer's compulsory super, on an income in this bracket, already consumes his entire concessional cap. There's nothing unused to carry forward. And even if there were, carry-forward contributions require a total super balance under $500,000 at the prior 30 June. The same response had already assumed he's sitting on $1.2–1.3 million. Both can't be true. The strategy it valued at $41,000 was, in its own numbers, unavailable to him.

 

Catch #3: Reading From Last Year's Rulebook

The plan used a $2 million transfer balance cap. It's been $2.1 million since 1 July this year which makes the couple's combined figure $4.2 million, not $4 million. Not fatal on its own. But nothing in the output carried a date, a source, or a "current as at" line. It read exactly as confidently as if it were current.

 

Catch #4: The Surplus That Wasn't There

This is the load-bearing error. The client's stated $150,000 was his target retirement spend after he stops working. The model reused it as his current spend, a different number entirely, and never asked which one it meant. Run it properly: on $500,000 gross he nets roughly $299,000. Take out mortgage interest and principal (around $110,000, on the model's own 6% assumption) and actual living costs of $150,000, and the real surplus is closer to $37,000 a year - not $150,000. Add two children in Sydney private schools, which the same response separately priced at $30,000–$45,000 each, and that surplus turns negative. The "nine years to clear the mortgage" conclusion - the spine of the whole plan - was built on a number out by a factor of four.

 

Catch #5: The Instruction That Could Cost Him His Deductions

The offset arithmetic itself was right: moving $400,000 out of the wrong loan saves roughly $11,280 a year. Buried at the very end of that section was one line "redraw and offset mechanics differ… sequencing matters." That sentence is carrying the entire risk. If that $400,000 sits in a genuine offset account, moving it is mechanical. If it's actually in a redraw facility, pulling it out to pay down a private home loan can permanently taint the deductibility of the investment portion of the debt sometimes impossible to unwind. The model can't see loan documents. It doesn't know which one this is. It gave the instruction anyway, attached an attractive number to it, and hedged in one sentence at the end. The conversation that should have happened instead - debt recycling, converting non-deductible debt into deductible debt over time - never came up. It was never asked.

 

What It Never Asked

He's a software sales director. That means base plus variable plus equity. Almost certainly RSUs, quite possibly an ESPP, very likely a concentration problem in a single stock. The word "RSU" did not appear once in three pages of output for the single biggest structural feature of this man's financial life. It never asked about his super balance, the $2 million foundation of the whole plan. Never asked about his spouse, her income, her super despite assuming a full couple's transfer balance cap for a household it knew nothing about. It invented a 6% interest rate for both loans, when I have clients on investment loans locked well under that today. It never asked what he spends, only what he'd like to spend one day. And it built a nineyear plan resting entirely on this one income continuing uninterrupted, for two children and $1.85 million of debt, without a single word about income protection or what happens if a sales quarter goes badly.

 

I counted at least eight material assumptions made about a stranger's life. Then the result was presented as a plan.

 

What AI Genuinely Can't Do

The maths, mostly, wasn't the problem. I checked it, and outside the errors above it largely held up. The real gap is structural, and it's bigger than most people think.

 

It cannot make anyone act. That response contained roughly a dozen action items. Everyone is a suggestion. Nobody rings him in March to check whether he moved the offset. A suggestion is not a commitment, a meeting and full follow up system with humans behind it is.

 

It cannot pre-commit you to anything. AI can tell someone to sell a fixed percentage of every RSU vest from here on. It cannot make that rule survive the month the share price jumps 40% and every instinct says "just this once, hold." The only version of a selling rule that actually works is one locked in before the emotion shows up. It’s called a standing instruction, agreed to in a calm month, that executes automatically in whatever month isn't calm. That's the exact mechanism behind the RSU strategy I built with a client last year: a set number of shares sold every quarter, tax planned for in advance, no discretion left for the day the share price is doing something exciting. AI can hand you the rule. It cannot lock you into keeping it.

 

It cannot coordinate other humans. Nearly everything in that plan needs an accountant, a broker and a super fund to act in the right sequence. Get the offset move wrong and you contaminate a deduction permanently. Get a contribution wrong and you breach a cap. Writing the plan is not the hard part making sure three or four different professionals do their part before 30 June is.

 

It cannot live outside a spreadsheet. The plan assumed nine years of clean, optimal decisions in a row - every RSU sold on schedule, every dollar directed exactly where the model said, no year where a bad quarter, a redundancy scare, or a once-in-a-decade family trip changes the sequence. Real financial plans don't get lived out on a spreadsheet. They get lived out by people who occasionally redraw $15,000 for a trip they decided they needed or hold onto shares a month longer than the rule said because letting go felt harder than the model assumed. A plan that only survives if you behave like a spreadsheet for nine straight years isn't a plan - it's a best-case scenario dressed up as one.

 

It cannot absorb liability or be accountable. There's no engagement letter behind that response, no best interests duty, no professional indemnity policy, and nobody whose name is attached if the offset instruction turns out to be sitting on a redraw facility instead. A recommendation that costs its author nothing when it's wrong isn't the same category of thing as advice, no matter how confidently it's delivered. And the last line of the response was a general advice warning: "prepared without taking into account your objectives, financial situation or needs." Read that against everything above it - his age, income, family, debts, and a firm answer on whether he can retire at 55. That warning and the three pages sitting above it can't both be true.

 

It cannot protect you from unknowns. Nothing in the prompt mentioned redundancy risk, because nobody's prompt ever does You don't think to ask an AI to plan for the layoff you don't know is coming. I wrote about this back in August 2025: two rounds of cuts at Microsoft, five months apart, same broader business. The clients who came through that cleanly weren't the ones with the best spreadsheet. They were the ones who'd already had the redundancy conversation with someone who'd watched it happen before. That kind of pattern recognition comes from sitting across from hundreds of tech professionals over two decades not from a single prompt in a single chat window.

 

The Bottom Line

Go and use it. I do, most days. It's a genuinely good place to get oriented, and to ask the questions you're embarrassed to ask a person. Just know the difference between what you're getting and what you think you're getting: information delivered with total confidence, built on assumptions you can't see, using thresholds that may already be out of date, from something that will never ask a follow-up question and will never be accountable for the answer. That's not nothing. It isn't advice.

 

If you're a tech professional on a high income with equity compensation and a mortgage, and you've got a nagging feeling you should be further ahead than you are - bring me whatever an AI tool has already told you. I'll tell you, honestly, what's solid and what isn't, and whether I'm the right person to take it further.

 

One condition, same as always: come ready to actually do something with it. The Confident Choice System only works for people who are ready to act, not people who are still thinking about it.

 

This is general information only and does not take into account your personal objectives, financial situation, or needs. Please consider whether it's appropriate for you before acting on it.