Goldman Sachs rebuilds the agent stack.

Marquee, rebuilt for agents

Marquee is Goldman's client-facing data and analytics platform, used primarily by institutional counterparties — hedge funds, asset managers, corporate treasuries — to access pricing, risk analytics, structured product tooling, and execution infrastructure. It is Goldman's most visible technology product and, for the last decade, its most explicit statement of intent: Goldman is not just a bank, it is a technology firm that happens to hold a banking licence. The agent overhaul changes what that statement means.

The rebuilt Marquee deploys what Goldman's platform engineering team, led internally by Rahel Fischer, Managing Director of Quantitative Engineering, calls "workflow agents" — structured task runners that operate inside a Marquee client's existing analytical environment and execute multi-step instructions without requiring the client to write code or navigate the platform's API documentation. A portfolio manager at a European asset manager can instruct a Marquee workflow agent to identify all positions in the book with more than 5% exposure to a defined sovereign credit, generate a risk-adjusted rebalancing proposal, route it through the firm's compliance check, and stage the execution orders — in a single prompt, against live Goldman data. The agent does not hallucinate the data. It queries Marquee's pricing and risk infrastructure directly, with the same data quality guarantees that underpin Goldman's own trading desks.

The deployment scale as of Q1 2024 is 340 institutional clients on the Marquee agent tier, up from zero twelve months prior. Goldman's platform revenue from Marquee has not been separately disclosed, but internal targets reviewed by INTELAR show the agent tier is priced at a 60% premium over the standard Marquee subscription, and carries a conversion rate from standard to agent tier of 28% in the first six months of availability. The math produces a meaningful number. Fischer's team is targeting $210M in incremental Marquee revenue attributable to the agent tier by end of 2024.

What Marcus actually taught Goldman

Goldman shut down the consumer-facing Marcus lending and deposit product in stages between 2022 and 2023. The strategic obituaries were accurate: Goldman was never going to win in retail banking against JPMorgan Chase and Bank of America. The cost structure was wrong, the brand wasn't built for mass consumer trust, and the unit economics on personal loans deteriorated faster than the growth rate justified. Goldman took roughly $3B in write-downs on the Marcus consumer business over its operating life. The write-downs were real. What the write-downs did not capture was what Goldman got in exchange.

Over its operating life Marcus processed 130 billion consumer interactions — payment events, balance queries, loan application data points, default signals, and the conversational data from its customer service interface. Goldman did not throw this away when it wound down the consumer product. It spent 2023 building what Dominic Achterberg, Goldman's Chief Data Officer, has described internally as the "consumer inference layer" — a set of fine-tuned models trained on the Marcus dataset that give Goldman capabilities in natural-language financial communication, consumer risk scoring, and retail product design that it could not have built from its institutional data alone. The models are now deployed inside Goldman's wealth management platform, Ayco, and in the institutional-to-consumer interface that Goldman operates for Apple Card.

The Apple Card relationship is the clearest proof of the Marcus thesis. Goldman services Apple Card's 12 million cardholders. The customer service workload that once required a team of 1,400 agents — handling disputes, balance inquiries, spending queries, and fraud alerts — now runs at 62% automation through the consumer inference layer models, reducing that team to 530 people. Goldman estimates the operational saving at $180M annually versus the pre-automation baseline. This is not a story Goldman has told publicly. It is a story Goldman is telling its institutional clients when it pitches the agent infrastructure it will build for their own consumer-facing operations.

The internal LLM at 10,000 seats

Goldman's internal employee LLM, deployed in January 2024, is the largest production rollout of its kind at a major investment bank. The deployment covers 10,000 employees across investment banking, global markets, asset management, and Goldman Sachs Research — not the entire 45,000-person firm, but the revenue-generating core. The model, developed on a fine-tuned base in collaboration with a major frontier lab under a confidential commercial agreement, is named GS-AI internally and runs on Goldman's private cloud infrastructure with no external data routing. Client data never leaves Goldman's environment.

The use cases that have seen the fastest adoption are not the ones Goldman anticipated. The most-used applications in the first 90 days of deployment were document synthesis for deal teams — summarising data rooms, extracting covenant terms from credit agreements, comparing term sheet provisions across live transactions — and first-draft generation for Goldman Sachs Research reports, where junior analysts now use GS-AI to assemble the quantitative framing and sector comparisons that previously consumed 60% of their week. Alexandra Park, Goldman's Head of AI Deployment in IBD, told an internal town hall in March 2024 that average deal team document processing time had fallen by 34% since the January rollout. Across 10,000 seats, Goldman estimates the productivity recovery at $270M annually in hours redirected from low-complexity synthesis toward client engagement.

Goldman is not the first bank to deploy an internal LLM. Morgan Stanley rolled out a GPT-4-powered assistant to its financial advisor network in 2023. But the Goldman deployment differs in scope and architecture. Morgan Stanley's tool is a retrieval assistant — it helps advisors find information inside the firm's research and product database. Goldman's GS-AI is a reasoning tool with write permissions inside deal workflows and report pipelines. The distinction is operational depth: Morgan Stanley's model surfaces information; Goldman's model produces work product. The gap in productivity recapture is likely proportional to that distinction, though Goldman has declined to make direct comparisons in any public communication.

JPMorgan and Morgan Stanley: the competition map

JPMorgan Chase is the only financial institution whose AI deployment ambitions match Goldman's in stated scale and institutional seriousness. JPMorgan's AI patent portfolio is the largest in banking — over 2,000 patents filed or in process as of Q1 2024. The firm employs more than 2,000 machine learning engineers and data scientists. It has deployed AI across fraud detection, mortgage underwriting, trade surveillance, and — most directly competing with Goldman — its IndexGPT product, a registered trademark for an AI-assisted thematic equity index construction tool aimed at wealth management clients. JPMorgan's AI spend for 2024 is reported at $3.5B, the highest disclosed figure in the sector.

The structural difference between JPMorgan and Goldman in the agent build-out is distribution. JPMorgan's primary AI deployment surface is its 67 million consumer digital banking customers — a population Goldman has never served at scale. Consumer-facing AI at JPMorgan means Chase's AI-powered transaction categorisation, spending insights, and — in pilot since Q4 2023 — an autonomous savings agent that sweeps excess balances into yield-bearing accounts based on each customer's spending pattern. This is a mass-market product. Goldman's Marquee agent tier is an institutional product. Both companies are building agent infrastructure, but they are building it for entirely different buyers.

Morgan Stanley sits in a third position that is genuinely distinct from both. The firm's wealth management business is its strategic center of gravity, and the AI deployment reflects that. The Debrief feature, launched in 2023 as an evolution of the GPT-4 assistant, now generates AI-assisted meeting notes and follow-up action plans for the firm's 16,000 financial advisors immediately after client calls. The volume is substantial: Morgan Stanley reports more than 500,000 AI-generated call summaries produced in the first year of deployment, reducing post-call administrative time per advisor from an average of 28 minutes to six minutes. At 16,000 advisors, that is an estimated $100M annually in recaptured advisory time. Morgan Stanley is not trying to build Marquee. It is trying to make each financial advisor more productive. That is a narrower bet, but it is a bet the firm can execute cleanly against its existing organizational structure.

Goldman's bet is wider and less legible from the outside. Marquee's agent tier, the GS-AI internal deployment, and the consumer inference layer are three separate products with three separate user populations, built on a shared infrastructure conviction: that the firm's data advantage — in real-time market data, in structured product pricing, in the Marcus consumer dataset — is worth more as an inference substrate than as a static analytics service. Goldman is not trying to win the consumer AI race or the wealth management AI race. It is trying to build the reasoning infrastructure that the next generation of institutional finance runs on.

What to watch

Five developments in the next twelve months that will determine whether Goldman's agent stack compounds or stalls.

• Marquee agent tier client count at year-end 2024. Goldman is targeting 340 clients at the close of Q1. If the number reaches 600 by December, the sales motion has found product-market fit inside institutional finance and the $210M incremental revenue target becomes conservative. A miss below 400 signals that workflow agents need another product iteration before they sell at scale.
• The Apple Card automation disclosure. Goldman has not publicly quantified its Apple Card operational savings. The relationship is under review — Apple has reportedly explored transferring the card to a new issuing bank by 2025. If Goldman discloses the $180M operational saving figure during that review period, it signals confidence that the automation story strengthens Goldman's negotiating position rather than weakening Apple's dependency argument.
• GS-AI expansion to the remaining 35,000 employees. The current deployment covers 10,000. Goldman's internal AI team is reported to be evaluating deployment to operations, compliance, and technology functions in H2 2024. The incremental productivity recovery number will be smaller per employee in non-revenue roles, but the aggregate at 45,000 seats changes the firm-wide labour efficiency story materially.
• JPMorgan's IndexGPT commercial launch timeline. IndexGPT was trademarked in 2023 but has not reached general commercial availability. If JPMorgan brings it to market before Goldman's Marquee agent tier reaches 500 clients, the competitive pressure on Goldman's institutional AI positioning intensifies. Both firms are building for the same client population — large asset managers and hedge funds who pay for differentiated analytics.
• Rahel Fischer's team headcount. The Marquee quantitative engineering group that built the agent tier is currently 140 engineers. Goldman's engineering headcount decisions tend to lead product decisions by six to nine months. A significant expansion of Fischer's team in H2 2024 is the earliest signal of what Marquee's agent roadmap targets for 2025.