The Inverse AI Trade

The search pattern

Screening 10,000+ earnings calls for structural fear

The screener behind this list ran two layers. A quantitative filter pulled 6-month, 1-year, and 3-year price deltas plus short-interest data from the Massive API to find equities under sustained selling pressure. Then a transcript-search layer read earnings calls for one specific conversational pattern — not companies hyping AI, but analysts pushing management to defend the core business against AI displacement, and whether the defense landed:

# The transcript signal, per company (grammar-constrained JSON, local LLM):
SCHEMA = {"ai_threat_raised":     "did an analyst explicitly push management on AI"
                                  " replacing the CORE business (not AI adoption hype)?",
          "management_defense":   "summarize the defense offered",
          "analyst_convinced":    "did follow-ups accept the defense, or keep pressing?"}

ask_transcript(ticker, quarter, question=(
    "Ignore generic AI enthusiasm. Find exchanges where an analyst or investor "
    "presses management to defend the core business against AI displacement. "
    "Did management's answer neutralize the skepticism?"), schema=SCHEMA)

# Screener result across all 19 flagged names: analyst_convinced = FALSE, every time.

The core finding: across every flagged company, not one management team convinced analysts that AI does not pose a structural threat. The strongest combined signals (price pressure + direct replacement questioning) sat in freight matching (YMM), legal research (RELX), middle-market credit (PNNT), and government procurement (TYL).

The flagged 19

Four waves of expected displacement

Wave 1: Immediate content, freelance & moderation disruption

Management teams here openly admit low-value transactional volume is evaporating - the screener's one group with tangible revenue degradation, not just narrative risk.

Wave 2: Ad tech, media distribution & platforms

Traffic and attention re-routing through AI answer layers instead of the open web these businesses monetize.

Wave 3: Professional services, legal & human capital

High-margin expertise businesses where the deliverable is text, judgment, or a match - exactly what LLMs claim to commoditize.

Wave 4: Logistics, enterprise software & financial systems

Second-order exposure: procurement freezes, portfolio credit risk, and platform intermediaries that AI agents could route around.

Expansion

Related companies, by what they actually sell

Each flagged ticker was pushed through the Atlas embedding corpus — 4,560 companies positioned by LLM-read 10-K product descriptions — to find its nearest product-space neighbors:

# Expand each flagged ticker through the product-embedding corpus:
sims = X @ X[idx[t]]            # cosine vs all 4,560 product-description vectors
neighbors = [tickers[i] for i in np.argsort(-sims)[1:7]]   # top 6, self excluded

Red chips above are flagged names showing up as each other's neighbors (Fiverr↔Upwork style confirmation). More interesting are the unflagged neighbors that appear next to multiple flagged companies — same product exposure, no analyst interrogation yet:

ticker	company	neighbor of
WMG	Warner Music Group Corp.	GETY, TME
HUYA	HUYA Inc.	IQ, TME
STGW	Stagwell Inc.	UPWK, NRDS
RGP	Resources Connection, Inc.	UPWK, KFY
NOW	ServiceNow, Inc.	TASK, KFY
NNI	Nelnet, Inc.	NRDS, TYL
IX	ORIX Corporation	RELX, KFY
ALIT	Alight, Inc.	TYL, CNNE

Live pricing

How the market is treating them

Daily closes via the Massive API (same serverless proxy as the Atlas map), normalized to the start of the window. The screener's windows were 6M / 1Y / 3Y:

If the hype bursts: the ranking

The inverse-AI screen, sorted by drawdown over the selected window, across the flagged 19 plus their top product-space neighbors. Read it with the screener's own caveat: Wave 1 names already report real revenue degradation — a burst in AI hype doesn't refill an evaporated market, so deep drawdowns there can be value traps. Waves 2–4 are largely narrative pricing: if agentic displacement stalls, the most-written-down names with intact revenue have the most torque. Unflagged neighbors with similar products and shallow drawdowns show what the market charges for the same exposure without the AI discount.

This page is an educational example of chaining a transcript screener, an embedding corpus, and market data. It is not investment advice or a recommendation; the flagged list and its framing come from an external screener and are reproduced as-is for illustration.