In brief
Chris Churchman of Goldman Sachs argues that useful AI products in finance come from reliability, not impressive demos. Because language models cannot tell a fact from a plausible guess, a firm has to build systems that tie every sentence to a checkable source. He says teams should build only what models cannot learn themselves, such as a firm's data, permissions and mandates, and should redesign work instead of just speeding up old steps.
The central tension is between a technology that improves so fast it makes yesterday's workarounds pointless and a business that must stand behind every answer. Churchman describes four years of rapid change, from prompts to autonomous agents. He also worries that handing reasoning to machines will weaken junior staff, so he wants products that strengthen human judgment. He remains optimistic that scaling will continue.
Facts
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Chris Churchman is head of Marquee, Goldman Sachs' digital platform for institutional and corporate clients, and co-chair of the firm's Global Banking and Markets AI Working Group.
“Chris is the head of Marquee, the firm's digital platform for institutional and corporate clients. And he's also co-chair of the firm's Global Banking and Markets AI Working Group.”
0:01:26 · Allison Nathan · Guest role · 2026
Churchman says Marquee is Goldman Sachs' platform for institutional and corporate clients and that its mission is to help them make decisions under great uncertainty.
“So Marquee is the firm's platform for institutional corporate clients. And really, our mission is to help our clients make decisions under great uncertainty.”
0:02:00 · Chris Churchman · Marquee mission
Churchman describes Market View as an ecosystem on Marquee, like a Pinterest for capital markets, where each widget is created by a domain expert from vendor data.
“We've got an ecosystem, fortunately, called Market View, which contains these, you can think of it like Pinterest for capital markets, where each widget is produced by an expert in their domain.”
0:03:11 · Chris Churchman · Market View
Marky AI, Goldman Sachs' generative AI assistant within Marquee, was available only internally at the time of recording.
“Marky AI, which is only available internally right now, you ask Marky a question, which is very targeted, is what you care about.”
0:04:10 · Chris Churchman · Marky AI · 2026
Churchman says Marquee includes a visual structuring tool, an intuitive interface for building derivatives that also shows scenarios and backtests, which he wants to wrap in an AI agent.
“we have something called visual structuring, right, which is an intuitive visual interface that comprise any derivative, for example.”
0:19:31 · Chris Churchman · Visual structuring
Churchman recalls that early ChatGPT-era models could handle only about 4,000 tokens, roughly 3,000 words, in their context window.
“And so you go, right, okay, it can only handle 4 ,000 tokens in the context window, so 3 ,000 words.”
0:06:26 · Chris Churchman · Context window · 2022
Churchman says context windows grew to 128,000 tokens a year after ChatGPT and then to a million tokens, making work on chunking documents largely pointless.
“And then a year later, the context window goes to 128 ,000 tokens. And you're like, oh, well, that was pointless. And then it goes to a million tokens.”
0:06:52 · Chris Churchman · Context window · 2023
Churchman says Google solved the 'needle in a haystack' retrieval problem when it raised its context limit to 1 million tokens.
“And so Google, when they raised the context limits and 1 million tokens, which was amazing, they solved the needle in the hate step problem as well.”
0:07:54 · Chris Churchman · Long context
Churchman cites a paper arguing that language models hallucinate because training and testing reward guessing over abstaining, like multiple-choice tests.
“there's a famous paper why LLMs hallucinate and it's because they're doing multiple choice tests and they're not rewarded for abstaining”
0:09:03 · Chris Churchman · Hallucination
Churchman describes context engineering as the era after prompt engineering, in which builders assemble memory, data, prompts and playbooks for a model.
“We've gone through context engineering, which is, okay, what if I can cobble together a bit of memory, a bit of data, a bit of prompts, a bit of playbooks.”
0:09:49 · Chris Churchman · Context engineering
Churchman predicts a next phase he calls mandate engineering, which concerns what AI agents may do, under whose authority and in whose name.
“like mandate engineering, which is like, what are AI agents allowed to do under the who's authority in whose name.”
0:10:51 · Chris Churchman · Mandate engineering
Churchman predicts AI agents will gain autonomy once an insurance or liability market exists around their actions.
“I think once there's an insurance market or a liability market around what these agents do, then I think you'll see the autonomy of them increase”
0:10:51 · Chris Churchman · Agent liability
Churchman says an AI demo is judged on its best day while a product is judged on its worst day, and that closing the gap is difficult work.
“a demo is judged on its best day, whereas a product is judged on its worst day, right?”
0:12:00 · Chris Churchman · Demo versus product
Churchman says reasoning models, function calling and the Model Context Protocol made Goldman's earlier reasoning scaffolds and custom tool connections pointless.
“all of those reasoning scaffolds we built, once reasoning models came out, totally pointless. All of those ways that we'd connected the LLM to our tools and our data, totally pointless once you had function calling and MCP.”
0:13:03 · Chris Churchman · Obsolete workarounds
Churchman says models do not natively know a firm's entitlements, mandates or how its data is connected, so these are what institutions should build.
“It doesn't know our entitlements. It doesn't know our mandates. It doesn't know the way that we connect our data together, et cetera.”
0:13:34 · Chris Churchman · Institutional knowledge
Churchman credits the 2017 transformer paper 'Attention Is All You Need' with letting every word in a context window attend to every other word.
“it's really the transformer paper in 2017, attention is all you need, which basically said that actually you can have every word pay attention to every other word in a context window”
0:21:26 · Chris Churchman · Transformer · 2017
Churchman says the Kaplan scaling laws and the Chinchilla paper showed that small experiments with AI models can predict results at much larger scale.
“But then the Kaplan scaling laws plus the chinchilla paper basically told you that these things, if you do a small experiment then to see whether it works you can then scale it up”
0:21:57 · Chris Churchman · Scaling laws
Churchman says the top priority of AI labs is to automate AI research and development, with running experiments the main bottleneck.
“importantly, in the labs, their number one priority above all else is to automate themselves, which is automate AI R &D.”
0:21:57 · Chris Churchman · AI R&D
This episode of Goldman Sachs Exchanges was recorded on Tuesday, August 11th, 2026.
“This episode of Goldman Sachs Exchanges was recorded on Tuesday, August 11th, 2026.”
0:25:03 · Allison Nathan · Recording date · 2026