The Prompt Portfolio Paradox

The Ad That Started This

If you listen to podcasts about markets, you’ve heard the pitch. Several platforms now run ads on every finance podcast in your feed: “Type your investment thesis. Our AI builds your portfolio.” Clean energy leaders. AI infrastructure picks. Mars exploration stocks. You describe a conviction, and an AI evaluates thousands of companies, weights them, and hands you a custom index in thirty seconds. One click to buy.

It’s a compelling product. I heard the ad, thought that’s a nice feature, and started digging into how it actually works under the hood. Thirty minutes later I realized it’s not a feature. It’s a bug.

Not the stock selection. The stock selection might be brilliant. The bug is the tax structure wrapping it. A simulation using actual S&P 500 returns from 1994 to 2023 puts a number on the damage: $100K in a Vanguard S&P 500 ETF for 30 years grows to $1.54M after taxes. The same $100K in an AI-generated custom index reaches $1.24M.

The gap: $297,474. Not lost to bad stock picks. Lost to tax plumbing.

The prompt portfolio paradox: the most sophisticated stock selection tool available to retail investors sits inside the least tax-efficient structure. The smarter the personalization, the worse the plumbing. And this disadvantage compounds.

Strange Plumbing

To understand why, you need to know about a piece of financial infrastructure so obscure it makes compiler optimization look mainstream: the in-kind redemption.

When investors sell mutual fund shares, the fund manager sells securities to raise cash. If those securities have appreciated, the fund realizes capital gains and distributes them to all remaining shareholders at year-end. You pay taxes on someone else’s exit. It’s as if every time a neighbor sold their house, you owed taxes on the appreciation of yours.

ETFs fixed this in 1993.¹ When institutional players (authorized participants) redeem ETF shares, they exchange them for the underlying securities directly. No sale. No taxable event. No distribution. A structural innovation that costs nothing and requires no skill.

The in-kind redemption tax advantage was never deliberately legislated. It emerged from the interaction of two provisions written decades apart: Section 852(b)(6) of the Internal Revenue Code,¹⁰ added by the Tax Reform Act of 1969 to govern how regulated investment companies handle in-kind distributions, and the general non-recognition rules for corporate formations under Section 351.¹¹ When Nate Most designed the ETF creation/redemption mechanism at AMEX in 1992, these provisions together meant ETFs could distribute appreciated securities to authorized participants without recognizing taxable gains. The IRS confirmed this treatment in Revenue Ruling 2001-38.¹² Congress has never explicitly blessed the arrangement. It survives because closing it would disrupt trillions in ETF assets.

The plumbing premium: a structural advantage that accrues to any fund that can use in-kind redemptions. Free. Automatic. Invisible. And unavailable to anyone who owns individual stocks.

Personalization Breaks the Wrapper

AI-generated index products are clever.³ You describe an investment thesis in natural language. An AI evaluates thousands of stocks. The platform builds a weighted portfolio of individual stocks you own directly, typically with a low minimum and a 0.40 to 0.50% annual fee.

The personalization works. And personalization is exactly what breaks the tax wrapper.

The result: 0.49% in fees (16x VOO’s 0.03%), no tax-loss harvesting, no in-kind redemption mechanism, and zero structural tax advantages. The stock selection has to outrun a handicap that compounds every year.

Say one of your AI picks doubles. You want to trim the position or exit entirely. That’s a taxable sale at your full rate. In an ETF holding the same stock, an authorized participant redeems shares in kind: the appreciated shares leave the fund, no sale occurs, no capital gains distribute. You notice nothing. The AI gave you a good pick, and the structure ensures you pay full freight on the gains.

The Simulation

I built a simulation to measure the structural cost. Actual S&P 500 index returns from 1994 to 2023. The mutual fund and AI index use real capital gains distribution rates from FXAIX and S&P 500 reconstitution data, respectively. Direct indexing uses a tax-loss harvesting schedule calibrated from academic literature. Four vehicles, each starting with $100,000.

By year 15, the AI index trails by six figures. By year 30, the gap is catastrophic. The gap accelerates because fees and taxes eat principal that would otherwise grow.

After full liquidation (all vehicles pay 15% LTCG on remaining unrealized gains):

High fees, forced capital gains, no tax shield.⁸ Any one alone is survivable. All three at once are devastating. If you wanted to guarantee the worst possible tax outcome from an index, you’d design this. The forces don’t add. They multiply.

Assumptions for reproducibility: actual S&P 500 total returns (1994 to 2023), 15% LTCG / 35% STCG tax rates. MF capital gains distributions from real VFINX data (1994 to 1999) and FXAIX data (2014 to 2019); all other years 0% (loss carryforwards). AI capital gains rates derived from S&P 500 constituent turnover (~4%/yr) multiplied by embedded gain fraction. DI tax-loss harvesting rates calibrated from academic estimates (1 to 2% early, decaying to ~0.2%; spikes in crash years). ETF distributions at 0.05% of NAV. All vehicles fully liquidated at year 30. No additional contributions. Full simulation code on GitHub.

The Direct Indexing Tradeoff

Schwab charges 0.40% annually for direct indexing, 13x the ETF’s fee. It finishes at $1,421,058: $120K below the ETF and $101K below the mutual fund.

The 0.40% management fee eats most of the TLH benefit over 30 years. Tax-loss harvesting still works (those crash-year spikes are real savings), but it can’t overcome the fee drag over three decades. Direct indexing is most valuable if you have large taxable gains elsewhere to offset, high marginal tax rates, and a portfolio above Schwab and Fidelity’s $100K minimums. Without those, a plain ETF wins.

The paradox cuts deeper at the top. Family offices with $50M+ negotiate custom ETF creation baskets with authorized participants, seeding new ETFs with appreciated securities from their existing portfolio without triggering a taxable event.⁹ When they exit, they redeem in-kind. No sale. No tax. Personalized stock selection with in-kind redemption benefits. The investors who need tax efficiency the least have the most access to it. The plumbing premium scales with wealth. Like most structural features of the financial system.

The Alpha Objection

AI-generated indexes aren’t sold as index replication. They’re sold as alpha. If the AI consistently beats the S&P 500 by 2% annually, the structural drag is worth paying.

AI stock selection is fundamentally different from human fund management. A human analyst covers 30 stocks and relies on intuition shaped by career incentives. An AI evaluates thousands simultaneously, processes earnings calls in seconds, and has no career risk from contrarian positions. The alpha mechanism is different in kind, not degree.

But over the 15 years ending 2023, roughly 90% of large-cap US funds underperformed the S&P 500 after fees.⁵ Those funds also employed smart people with proprietary data. The 0.49% fee means the AI needs at least 0.46% annual alpha just to match the ETF’s cost structure, before overcoming any tax drag. The tax drag compounds: the AI index needs roughly 0.8% annual alpha to break even with the ETF.

Structure alone costs $297K. If you believe your AI index will overcome that through superior stock selection, you’re betting on AI capability, not fund structure. That bet might pay off. But know the headwind.

When Structure Doesn’t Matter

This analysis applies to taxable brokerage accounts, where roughly 30% of US retail assets sit. In the rest, fund structure is irrelevant:

• Small portfolios. Below $100K, direct indexing is unavailable and the fee differences compound slowly. ETFs win by default.
• Short time horizons. Under 5 years, the spread between vehicles is modest. The compounding benefit of tax deferral needs decades to become dramatic.
• Tax-advantaged accounts (401k, IRA, HSA, Roth). No capital gains tax. No distributions to worry about. No turnover drag. The 0.49% fee is the only cost, and it buys you something a plain index fund cannot: a portfolio built around a specific thesis. Inside a Roth IRA, only fees matter.

What’s More Interesting

The Missing Pieces

AI-generated indexes are the mutual funds of 2026: a good idea trapped inside infrastructure that hasn’t caught up. Mutual funds needed the ETF wrapper to become tax-efficient. AI indexes need their own structural innovation.

The infrastructure is already there. AI index investors own individual stocks. That’s the same foundation direct indexing uses. It could support strategies that none of these platforms currently offer:

• Tax-loss harvesting. Individual stock ownership already supports it. When Apple drops 8% in a week but the AI’s thesis hasn’t changed, sell for the loss, buy a correlated substitute, keep the thesis intact. Schwab and Wealthfront proved this works at scale. Nobody offers it for AI-generated portfolios.
• Covered calls. You own the shares. Sell calls on positions where the AI’s conviction is cooling or where implied volatility makes the premium attractive. The income could offset part of the management fee. Harvested losses could offset the short-term tax on premiums. An ETF can’t do this at the position level.
• Portfolio hedging. Protective puts on concentrated positions. Collars that cap upside in exchange for downside protection. The AI already evaluates the risk profile of every holding. Position-level hedges become possible. Pooled funds can’t offer this.
• Automated exit strategy. The simulation liquidates everything at year 30, triggering 15% LTCG on all remaining gains. A managed exit unwinds positions over years: specific lot identification, pairing sales with harvested losses, timing exits around low-income years. The liquidation tax bill shrinks from a cliff to a slope.

The Training Problem

This is the interesting AI problem. Not stock selection. The model has to balance tax efficiency, income generation from options, and downside protection simultaneously, shaped by each user’s risk tolerance, income needs, tax bracket, and time horizon. Every trade is a decision with tax consequences, income tradeoffs, and hedging implications. The action space is large, the feedback is delayed (tax outcomes materialize at year-end), and the objective varies per user.

There’s reason to think the training problem is tractable. Tax rules are deterministic: given a sequence of trades, the tax outcome is computable without human labels. Decades of historical stock prices, options chains, and index reconstitution data are public. You could build a simulator and generate millions of portfolio trajectories with known rewards. Offline RL (Decision Transformer,¹³ Conservative Q-Learning¹⁴) has shown promise learning policies from historical trajectories in other domains. Model-based RL¹⁵ could plan within such a simulator, testing tax-loss harvesting schedules and covered call strategies across market regimes.

Whether any of this survives contact with real markets is an open question. Simulators miss liquidity constraints, regime shifts, and the feedback loops from crowded trades¹⁷. Personalization adds another layer: how much does this user value income over tax savings over downside protection? Tuning that weighting requires RLHF¹⁶ data from real users, which means building a platform with enough scale to collect meaningful preferences. You’d start with sensible defaults and hope enough users stick around to generate the preference data for personalization.

The simulation suggests how much room there is. The mutual fund trails the ETF by just $19K over 30 years. Direct indexing trails by $120K with TLH and a 0.40% fee. An AI index with a 0.10% fee, TLH, covered call income, and a managed exit might land in that range: close enough to the ETF that even modest alpha could make the product net-positive. The breakeven would drop from 0.8% annual alpha to something lower, though how much lower depends on execution details nobody’s tested yet. On the regulatory side, Rule 6c-11⁶ standardized custom baskets for ETF creation/redemption, and the SEC has since approved non-transparent active ETFs (Precidian, Fidelity, T. Rowe Price) that disclose holdings quarterly instead of daily. The precedent exists for active strategies inside tax-efficient wrappers, though whether the SEC would allow a non-transparent active ETF with daily trading and in-kind redemptions is an open question.

Conclusion

The current offerings charge 0.49% for what amounts to an API call that picks stocks and hands you the tax bill. The picks could be brilliant. On a risk-adjusted basis, they need to have a 0.8% annual alpha to just to break even with a plain ETF. The interesting product might be not the stock picker, but the agent that manages what happens after the picks. Harvesting losses, generating income from options, hedging concentration, unwinding positions tax-efficiently over decades. Stock selection for a trade was the easy part, executing it and managing the lifecycle is the hard part.

Bibliography