Tax-Loss Harvesting at Scale: AI Tools for Advisors

Tax-loss harvesting (TLH) is one of those workflows where the math is undeniable but the operational cost has historically been high. Net of the wash-sale rule, of cost-basis tracking, of timing, a daily scan catches losses that a quarterly pass misses, because prices can recover between one quarter-end and the next.

The major TAMP and robo platforms (Schwab, Betterment, Vanguard, Wealthfront) deliver daily TLH at scale. For RIAs not on those platforms — or running custom portfolios — AI-augmented TLH is the way to match that capability without licensing the whole stack.

The math behind continuous TLH

Vendors including Vanguard, Wealthfront and Betterment have published their own estimates of what continuous harvesting adds to after-tax returns. The estimates differ widely, and each rests on its own assumptions about market volatility, tax bracket, cash flows and portfolio turnover. Read those assumptions before you repeat any figure. The direction is consistent: more in volatile years, less in calm ones.

To estimate the benefit for a client, apply the losses a daily scan would have harvested to that client's actual marginal rate and gains picture. Multiplied across an RIA's taxable book, the firm-level number can be significant.

This is real money that's leaking when TLH is done manually.

What's actually hard about TLH

Three things make TLH operationally expensive:

1. Wash-sale rule

You can't realize a loss on a security and buy back a "substantially identical" security within 30 days before or after. Substantially identical is a regulatory grey area for ETFs (VTI and ITOT? Probably substantially identical. VTI and VXUS? Not.). A conservative policy is the safer default. AI doesn't fix the regulatory grey area but does help apply the firm's policy consistently across thousands of positions.

Wash-sale also applies across accounts. Selling at a loss in your taxable account and buying the same security in your IRA within 30 days triggers a permanent disallowance, not just a deferral (Rev. Rul. 2008-5). A manual process that looks at one account at a time cannot see it.

2. Cost-basis tracking

Loss harvesting requires lot-level cost basis. Most custodians provide this; some don't, or provide it inconsistently across account types. Tax-lot accounting (specific identification vs. average cost vs. FIFO) compounds the complexity.

3. Replacement security selection

You sell a position for a loss. You need to replace it (or stay out for 31 days). The replacement should track the original closely enough to keep portfolio exposure intact but not be "substantially identical" under the wash-sale rule.

Pre-mapped replacement pairs (e.g., VTI ↔ ITOT for total market, IVV ↔ VOO for S&P 500) is the standard playbook. AI doesn't invent these — but it does maintain them, apply them consistently, and watch for new replacement opportunities as new ETFs launch.

Where AI specifically helps

Five places:

1. Daily scan vs quarterly project. A scan that runs daily across every taxable lot in every household identifies harvestable losses the day they appear. A quarterly pass sees only what is still at a loss on the day it runs.

2. Cross-account wash-sale checking. When the AI sees the full household — taxable + IRA + spouse + joint — it can check wash-sale exposure across all accounts before recommending a sale. Critical for clients with concentrated households.

3. Replacement security application. Maintain firm-approved replacement pairs in code. AI applies them consistently and flags edge cases (e.g., what if the replacement is also at a loss?).

4. Realized-loss budgeting. Set an annual realized-loss target per client ($3,000 against ordinary income under IRC section 1211(b), more if there are gains to offset). AI can pace harvesting throughout the year to hit the budget without over-realizing.

5. Year-end optimization. December gets crowded. AI prioritizes harvests by impact and processes them in a controlled cadence rather than the December 31 fire drill.

Build vs buy in 2026

Three options:

Option A: Use a TAMP that includes TLH. Schwab, Envestnet, Pontera, and others offer this. Simple. Constraint: you're using their model portfolios.

Option B: License a TLH-specific platform. Tools like Vise, Smartleaf, and others overlay TLH on whatever custodian you use. Pricing is usually per account, so check each vendor's current pricing. Works well if you have a unified account base.

Option C: Custom AI build. Sits inside your existing infrastructure, applies your specific wash-sale policy and replacement pairs, integrates with your custodian feeds. Higher upfront cost, lower per-account cost at scale, full firm control.

The right choice scales with firm size and customization needs. A firm on model portfolios is usually well served by A or B. A firm running custom portfolios across several custodians is where C earns a look. Compare the options on the after-tax value you estimate for your own book, not on a vendor's average.

What goes wrong

Three common deployment errors:

Aggressive substantially-identical interpretation. Some early TLH tools interpret "substantially identical" loosely (treating any S&P 500 tracker as different from any other). Risky. Conservative interpretation: pre-vetted replacement pairs only.

Cross-account blindness. If the system only sees the taxable account, it misses wash-sale exposure in the household's IRA or in the spouse's accounts. Make sure the data layer is at the household level, not the account level.

No advisor approval gate for high-impact harvests. A $50k harvest in a $1M account is material. The system should pause for advisor review on high-impact harvests, not auto-execute. Reg BI implications.

Compliance considerations

TLH itself is generally fiduciary-aligned (it benefits the client) but the execution layer raises questions:

• Best-execution obligations under FINRA Rule 5310 apply to trade execution
• Reg BI applies to the recommendation to harvest (technically a recommendation to sell)
• Fee disclosure for AI-augmented TLH services, described as part of the service offering
• Recordkeeping for the decision-making process — what trigger fired, what alternatives were considered, why this trade

These are not blockers, just requirements. Build the audit trail from day one.

How to size the return for your firm

The value arrives in two places, and both can be estimated from your own book:

• The after-tax value to clients, which depends on how much of the book is taxable, how volatile the holdings are, and how often losses were harvested before
• The capacity recovered from automating what was a quarterly manual effort, which the team's own time records will show

The first is the one clients notice. The second is the one the firm banks.

Custom build cost depends on the number of custodian feeds, how many replacement pairs and policies the firm maintains, and whether trades route back for execution or stop at a recommendation. Get a written scope and quote, and budget separately for recurring compute and support. Payback is the one-time cost divided by the monthly value of the two numbers above.

If you run a large taxable book without continuous TLH, it belongs on the shortlist for a first project. If you want to scope it for your firm, that's a 30-minute call.