Goldman Sachs Silicon Valley AI Survey: Data Moats and Workflow Control Define Corporate Competitive Advantages
nashnova research
Goldman Sachs' third annual Silicon Valley AI field survey concludes that proprietary data uniqueness, workflow control, and monetization capability will decide which companies win as AI reshapes information and business services.
Why do companies with exclusive data benefit first?
AI models keep improving, but they cannot generate the external information they need on their own — whoever holds unique, continuously updated, hard-to-replicate datasets has a moat.
Moody's demonstrated a case: insurance submissions in PDFs, spreadsheets, and free-form documents were standardized for underwriting in seconds — manual processing typically takes hours or days. This means → Moody's proprietary data, entity-relationship graphs, and catastrophe models form a moat generic AI cannot replicate.
The counter-example is FactSet (FDS): Daloopa showed how public company filings can automate financial data collection and model maintenance — information FactSet can access, but so can everyone else. In plain terms = if the data you sell can be scraped from public sources, that business is in danger in the AI era.
What does the "spectrum" of data uniqueness look like?
S&P Global (SPGI) holds ratings, commodity benchmarks, indices, private-market pricing, and other multiple differentiated data lines — Goldman assigns a "positive/mixed" read, noting asset diversity but uneven defensibility across lines.
Gartner (IT) faces a tougher position: Harvey and Simile demonstrated that AI can turn standardized research content into reusable agentic workflows. This means → when research reports themselves can be "decomposed, reassembled, and redistributed" by AI, the scarcity of Gartner's research and training materials declines.
This reflects a broader rule: the data moat question is not "do you have data?" but "can AI reconstruct your data from public sources?"
Why does "who controls the workflow" matter more than "who has the data"?
Goldman observed a key AI evolution: from retrieving information to drafting documents, coordinating tasks, and completing complex professional workflows. In plain terms = AI is no longer just the assistant that looks things up — it is starting to do the work itself.
Thomson Reuters (TRI) combines Westlaw legal content with CoCounsel, extending into legal research, drafting, and document review — but the risk is clear: if Harvey or Clio control the client interface, Thomson Reuters could become a back-end content supplier.
Clio acquired vLex in 2025 for $1 billion, gaining a library spanning over 110 countries and more than 1 billion legal documents. This means → the strategic value of global legal content assets is being repriced, and the contest centers on "who owns the platform where clients get work done."
Which companies have moats AI cannot easily bypass?
FICO and Equifax (EFX) received the clearest positive read. Stanford researchers noted that the main barrier to deploying AI agents in production is not model capability but unresolved accountability mechanisms — a framework closely aligned with FICO's strength in decision governance.
Equifax's The Work Number — built on employer-contributed employment and income records — is the hardest-to-replicate differentiated asset. This means → as AI agents evolve from "generating recommendations" to "initiating decisions," verification demand will occur earlier and more frequently, lifting Equifax's data value.
Verisk (VRSK) faces pressure: Corgi demonstrated how an AI-native insurer can automate underwriting without heavy reliance on Verisk's data, challenging the assumption that traditional insurance data becomes harder to replace in the AI era.
Why are physical services and education more resilient?
Robert Half (RHI) faces the largest impact: AI automation of data collection, model maintenance, and structured knowledge processing will directly shrink demand for accounting, finance, administrative, and junior tech roles.
ManpowerGroup (MAN) gains relative protection from a higher share of blue-collar placements, though its Experis brand still faces automation pressure in routine IT support. In plain terms = blue-collar jobs are safe for now, but "sitting at a computer doing standardized work" is the first category hit.
McGraw-Hill (MH) holds over 100 million paid course licenses and 7.5 million early AI-product users, with a potential licensing premium of 5% to 15%; ADT, Cintas (CTAS), and Rollins (ROL) receive positive reads because their core business is physical activity — AI adds value but cannot replace the on-site service itself.
Where does the money come from? How many monetization paths are there?
Goldman identifies three revenue paths: usage and transaction pricing, product premiums and outcome pricing, and infrastructure demand. This means → future pricing moves away from "one seat per person per year" toward "how many times AI called your data and how many decisions it made."
Iron Mountain (IRM) has a more direct opportunity: Vercel and ClickHouse noted that agentic applications generate request volumes, token consumption, and telemetry data far exceeding traditional software, directly driving data-center leasing demand.
Goldman also warns: FactSet, Gartner, Robert Half, and Pitney Bowes (PBI) face a "disruption-before-monetization" timing risk — existing revenue may erode faster than new AI products can generate income. This reflects the most dangerous dynamic in the AI transition: not "can you build a new product?" but "can old revenue hold until the new product ships?"
市场有风险，内容仅供研究参考，不构成投资建议。