Goldman Sachs Exchanges
2026-08-31 · Hosted by Allison Nathan · Goldman Sachs
Executive Summary
Goldman Sachs Exchanges hosted Chris Churchman, head of Marquee and co-chair of the firm's Global Making and Markets AI working group, for a deep dive on how the firm is embedding generative AI into its institutional platform. Churchman described Marquee AI as a system that takes a client's natural-language question, decomposes it into research topics, pulls from research, trading-floor commentary, and market data widgets, and assembles a fully sourced, auditable answer — with every sentence grounded (traceable) to a specific document or calculation.
Key Stories & Changes
1. Marquee AI's grounding architecture to fight hallucination
Every AI-generated sentence in Marquee AI is grounded to a specific data source or auditable calculation — a deliberately harder standard than typical retail AI tools that merely cite generic sources
Churchman noted that even web-grounded models, when challenged, admitted they are "better at sounding thorough than being thorough"
Cited a research finding that LLMs hallucinate partly because they are never rewarded for abstaining from an answer during training/testing — guessing is statistically optimal
Marquee AI is currently available only internally at Goldman Sachs, not yet extended to clients
2. Four eras of AI engineering in under four years
Prompt engineering → context engineering (stitching together memory, data, prompts) → agentic engineering (self-checking loops that catch errors) → environment engineering (securing agent access to tools/data)
Churchman predicts the next phase will be "mandate engineering" — defining what AI agents are authorized to do and under whose authority, eventually enabled by an insurance/liability market around agent actions
Key lesson: "build for the model you will have at launch, not the model you have when you start development" — RAG pipelines and vector databases built for small context windows became obsolete once context windows expanded to millions of tokens
3. Automation vs. reconception mindsets for AI products
Automation mindset: replace a human cognitive bottleneck in an existing process with AI, get the same answer faster — most firms are doing this today, but it "locks in legacy" processes
Reconception mindset: assume intelligence is abundant and elastic, then ask what questions could never be answered before due to human bandwidth constraints
Example given: Goldman's "visual structuring" derivative-pricing tool being wrapped in an AI agent so users only need to express intent, rather than learn the tool
4. Belief in continued AI scaling
Churchman cited the 2017 "Attention Is All You Need" transformer paper and subsequent scaling-law research as justification for continued heavy AI infrastructure spending
Argued AI labs' top priority is automating AI R&D itself — if scaling laws hold for small experiments, they can be trusted to hold at 10 orders of magnitude larger scale
Named two missing pieces: self-learning/continual memory (catastrophic forgetting is unsolved) and a "native form factor" for AI, comparing the current era to early television replaying radio shows
Trends Identified
1. From automating tasks to redesigning workflows
Most institutions remain in an "automation mindset," using AI to speed up existing processes without questioning whether those processes still make sense. Churchman's framing suggests the next competitive wave will favor firms that reconceive workflows entirely around abundant intelligence rather than layering AI onto legacy systems.
2. Grounding and provenance as the real bottleneck
The technical frontier in enterprise AI has shifted from raw model capability to trust infrastructure — ensuring every output is auditable and sourced. This is especially critical for regulated financial institutions where a firm must "stand behind" the analysis it generates, unlike consumer AI tools that can hedge with vague citations.
3. Institutional AI differentiation shifts to proprietary data and entitlements
As frontier models absorb nearly all general-purpose capability, Churchman argues the only durable edge left for enterprises is what a model cannot learn on its own — internal entitlements, mandates, and proprietary data connections — meaning firms should stop building around model deficiencies that will soon be solved.
4. Concern over cognitive atrophy and apprenticeship culture
As AI absorbs more reasoning tasks, Churchman warned of a "cognitive atrophy" risk where junior professionals lose the tacit, hands-on learning that built senior expertise historically. He framed preserving apprenticeship culture as a deliberate design choice, not something that happens by default. ---
Sentiment Analysis
Overall Market Sentiment: Cautiously Optimistic
The discussion reflected strong belief in AI's technological trajectory tempered by candid acknowledgment of its current limitations and risks around trust, reasoning atrophy, and unresolved architecture questions.
Risk Factors Highlighted
Hallucination is structurally unsolvable: Models cannot distinguish between a learned fact and a plausible-sounding extrapolation, and are not trained to abstain from guessing.
Cognitive atrophy: Over-reliance on AI reasoning could erode professionals' ability to reason from first principles, mirroring historical loss of skills like way-finding or memorization.
Loss of tacit/apprenticeship knowledge: Automating junior-level tasks risks losing the informal, hands-on learning that produces senior expertise.
Unresolved "mandate" and liability questions: No insurance or liability framework yet exists for what AI agents are authorized to do, limiting how much autonomy institutions can safely grant them.
Betting against a fast-moving model landscape: Building product infrastructure around a model's current deficiencies (e.g. small context windows) risks wasted engineering effort once models improve.
Security and entitlement risks in agentic environments: Churchman referenced "recent headlines" about AI agent security, underscoring risk in ensuring agents operate only with properly entitled data access.
No native AI form factor yet identified: The industry is still in an analogous "command line" phase, creating uncertainty about how AI products will ultimately be consumed.
This episode was covered in today's [The Market Signal — 2026-08-31](https://marketsignal.beehiiv.com/p/the-market-signal-2026-08-31), a cross-source synthesis of multiple podcast reports.