USE CASE
Thematic investing for AI agents: turn a view into a portfolio
A person has a view: more AI, less oil, "something like Berkshire's book". Abacus Labs does not judge the view. It measures how much of the theme the account already holds, then returns portfolios that reach the target exposure with the rest of the risk, tax and turnover under control.
The problem
Themes are where people most want to act and where the record is weakest. Morningstar's thematic fund landscape report found that fewer than one in ten thematic funds both survived and beat a global equity index over the 15 years to mid 2024, and more than half of those open at the start had closed (Morningstar, 2025). An agent that answers "more AI" by buying a thematic ETF or eight tickers from memory makes two mistakes at once: it adds a forecast nobody asked it to make, and it ignores what the person already owns.
The second mistake is the fixable one. A broad index fund already holds the theme. Three funds can hold the same ten names. Sizing a tilt means looking through every holding, measuring the exposure that is already there, and adding only the difference, while keeping sector and factor risk from drifting with it. That takes holdings data and a risk model, not a language model's recall.
What the agent asks
USER I want more AI exposure but not all in.
agent -> analyze(account)
agent -> propose(account, "raise AI-linked exposure to 15%, keep overall risk about the same", options=3)
AGENT You already hold about 9% in AI-linked names through
your index fund. To reach 15%:
1 one semiconductor ETF +6% vol +1.1 pts tax $0
2 eight individual names +6% vol +1.4 pts tax $0
3 chips only, smaller tilt +3% vol +0.5 pts tax $0
USER Option 1. ETFs only.
agent -> commit(option_1)
USER Separately, can you make my stocks look more like
Berkshire's holdings?
agent -> propose(account, reference="berkshire_13f_latest",
"move toward reference, tax under 1000", options=2)
What Abacus Labs returns
{
"status": "ok",
"trades": [
{ "symbol": "SMH", "side": "buy", "qty": 31 },
{ "symbol": "VTI", "side": "sell", "qty": 9, "lots": ["lot_02"] }
],
"theme": {
"name": "ai_linked",
"definition": "operator_list_v3",
"exposure": { "before": 0.09, "after": 0.15, "target": 0.15 },
"overlap_with_existing_funds": 0.04
},
"risk": { "before": 0.17, "after": 0.18, "benchmark": 0.17,
"largest_added_exposure": { "technology": 0.06 } },
"tax": { "realized_gain": 310.00, "realized_loss": 0.00 },
"checks": { "cash": "pass", "whole_shares": "pass",
"wash_sale": "pass", "limits": "pass" },
"explain": "Buys a semiconductor ETF to lift AI-linked exposure from 9% to 15%. Trims one low-gain index lot to fund it. Adds about one point of volatility.",
"certificate": "cert_2026-11-06_7a1f"
}
How it works
A theme enters Abacus Labs as a definition: a list of names, an ETF's holdings, a reference portfolio such as a public 13F filing, or a list the operator's agent resolves from text. Abacus Labs looks through every fund in the account to measure current exposure to that definition. The target exposure becomes a constraint, and our point-in-time risk model holds the other factor and sector exposures near where they were, so "more AI" does not quietly become "more momentum and more size".
This is the portfolio optimization API doing what it does elsewhere, with the theme as one more constraint. Options differ in the vehicle and the size: one ETF, individual names, a smaller tilt. Each carries its volatility change, tracking error to the original account, overlap with existing funds and tax. Explore returns them in tens of milliseconds. Commit re-solves the chosen option with an exact mixed-integer solver and returns whole-share trades with a certificate.
The honest framing: the view is the person's. Abacus Labs implements it, measures the risk it adds and reports what it costs. It does not forecast whether the theme will pay.
AI-linked exposure, example account
Today, via index fund ######################## 9% 1 semiconductor ETF ######################################## 15% 2 eight names ######################################## 15% 3 chips only tilt ################################ 12%
Example account, illustrative numbers from the conversation above. Exposure is measured through fund holdings using the operator's theme definition.
Who uses this
- General assistants via MCP
- "I want some of X" is the most common investing request an assistant hears. The MCP server returns sized, checked options instead of a ticker list.
- Broker agents
- Turning a user's view into a basket that respects what the account already holds, with whole shares and wash-sale checks.
- Advisor tools
- Client values and exclusions, such as no fossil fuels, implemented as constraints with the tracking error shown.
The Personal plan is free: 3 portfolios and 300 credits a month, no card. Paid plans add portfolios, credits and client use.