USE CASE
Concentrated stock de-risking for AI agents
Half the account in one stock is the most common risk problem an agent meets. Abacus Labs returns the ways out: trim and pay tax, trim using losses elsewhere, or keep the position and diversify around it, each with the tax bill, the risk left and a checked trade list.
The problem
Single stocks are a lottery with a long tail. Bessembinder's study of every US common stock from 1926 to 2016 found that 4% of companies accounted for all of the net wealth created over Treasury bills, and 58% of stocks returned less than Treasury bills over their lifetimes (Bessembinder, Journal of Financial Economics, 2018). Vanguard treats 10% to 20% of a portfolio in one name as concentrated (Vanguard Advisors, 2023).
Knowing that is easy. Acting on it is lot arithmetic. Shares from vesting arrive in many lots, some short-term and some long-term, at very different cost bases. The gain on a sale depends on which lots go. Losses elsewhere in the account can offset it. An employee purchase plan buying the same stock inside 30 days of a loss sale is a wash. And whatever is left after the trim has to be measured, because 35% in one name is still a lot of risk.
A language model will say "diversify". It cannot compute the gain per lot, the tax across options, or how much risk actually remains.
What the agent asks
USER Half my account is NVDA. Is that bad?
agent -> analyze(account)
agent -> propose(account, "reduce NVDA concentration", options=3)
AGENT NVDA is 52% of your account and 71% of its risk.
A 30% drop in NVDA takes 16% off your total.
Three ways out:
1 trim to 25% tax $2,100 risk -41%
2 trim to 35% using losses tax $0 risk -27%
3 keep it, add uncorrelated tax $0 risk -33%
USER Could I cap this year's gain at $5,000?
agent -> propose(option_1, change={ "max_tax": 5000 })
USER Do 2 now.
agent -> commit(option_2)
What Abacus Labs returns
{
"status": "ok",
"trades": [
{ "symbol": "NVDA", "side": "sell", "qty": 120, "lots": ["lot_19", "lot_22"] },
{ "symbol": "INTC", "side": "sell", "qty": 400, "lots": ["lot_03"] },
{ "symbol": "VTI", "side": "buy", "qty": 38 }
],
"concentration": {
"symbol": "NVDA",
"weight": { "before": 0.52, "after": 0.35 },
"risk_share": { "before": 0.71, "after": 0.49 }
},
"tax": { "realized_gain": 11940.00, "realized_loss": -11960.00, "net": -20.00 },
"risk": { "before": 0.31, "after": 0.22, "benchmark": 0.17 },
"checks": { "cash": "pass", "whole_shares": "pass",
"wash_sale": "pass", "limits": "pass" },
"explain": "Trims NVDA to 35% using Intel losses, so the sale costs no tax. The two NVDA lots sold are the highest-basis long-term lots.",
"certificate": "cert_2026-11-06_8c1e"
}
How it works
Abacus Labs measures concentration two ways: the position's weight, and its share of portfolio risk from our point-in-time risk model. The second number is usually larger and is the one that matters. Each option is a different objective under the same rules: a target weight with a tax budget, a target weight with net tax near zero using loss lots elsewhere, or an unchanged position with added names chosen to move little with it.
Lot selection is explicit. Long-term, high-basis lots go first unless the agent says otherwise. Planned buys are checked against 30 days of sales in every linked account, including an employee purchase plan. Whole shares, position limits and a cash floor hold exactly because commit re-solves the chosen option with an exact mixed-integer solver. A gain cap scales the trim back to the largest step that fits, and the response says how far from the target it stops.
Explore returns options in tens of milliseconds, so "what if I cap this year's gain" is a follow-up question, not a new project. Abacus Labs does not say whether the stock will go up. It says what each exit costs and what risk remains.
US stocks by lifetime outcome, 1926 to 2016
Below Treasury bills ######################################## 58% Above, modestly ########################## 38% All net wealth creation ### 4%
Source: Bessembinder, Do Stocks Outperform Treasury Bills? (2018), about 26,000 CRSP stocks, 1926 to 2016. Not an Abacus Labs result.
Who uses this
- Broker agents
- Employees with vested stock asking what to do with it. The agent calls the portfolio API and shows three exits with tax, instead of one answer.
- Advisor tools
- Multi-year sell-down plans with a gain budget per year, and a direct indexing sleeve built around the position.
- General assistants via MCP
- The first analyze call on a linked account often surfaces concentration. The next call shows the options.
The Personal plan is free: 3 portfolios and 300 credits a month, no card. Paid plans add portfolios, credits and client use.