How OpenAI Built ChatGPT for Bankers and Analysts - Joseph Kim, OpenAI

AI model capabilities are advancing rapidly, but value realized in industries like finance lags behind due to a 'capability overhang'; builders must create t...

Abstract

This synthesis examines the deployment of a finance-specific configuration of ChatGPT, built by OpenAI with design partners Morgan Stanley and Evercore, as a case study in closing the capability overhang - the gap between what frontier models can do and the value practitioners actually extract. The methodology combines qualitative product analysis with technical evidence on model scaling, data architecture, and adoption patterns. Key findings include a documented adoption inflection driven by interface changes rather than model upgrades, an architecture favoring proprietary data hosting and post-training over protocol-level access alone, and a threshold model explaining why capability gains appear discontinuous to end users. Practical implications include specific guidance on when to invest in post-training versus MCP-based tool access, and a forward-looking projection that finance-specific models will match top human expert performance by 2027.

1. Introduction

Frontier model capability has advanced on a trajectory that is, in retrospect, strikingly predictable: scaling laws forecast loss curves and evaluation performance before training runs complete, and compute requirements for fixed performance levels have fallen by roughly half every eight months. Yet the translation of this capability into realized professional value - particularly in regulated, document-heavy industries such as investment banking and equity research - has lagged conspicuously behind.

This asymmetry is termed the capability overhang: the distance between what a model is capable of and the value a user can actually obtain from it. The overhang is not primarily a modeling problem. It is a product problem, arising from mismatches in data access, interface design, and output format between general-purpose model deployments and domain-specific workflows.

This analysis addresses a central question: what determines whether advancing model capability converts into realized enterprise value, and what architectural and product choices close that gap fastest? The discussion is organized around a concrete case - a tailored ChatGPT deployment for financial professionals - and extends to broader frameworks governing capability growth, including the three drivers of model improvement (scaling compute, algorithmic improvement, and unhobbling) and a threshold phenomenon that explains why users often perceive progress as sudden rather than continuous.

2. Background and Related Work

The analysis rests on three interlocking frameworks. First, capability growth is attributed to three drivers: scaling compute, algorithmic improvements, and unhobbling - the extraction of more useful work from a fixed level of underlying model capability through techniques such as reinforcement learning from human feedback (RLHF) for instruction-following and tool use for search, code execution, and software interaction. Second, pre-training scaling laws have demonstrated predictive validity: GPT-4's final loss and coding-evaluation performance were forecast accurately before training concluded, and algorithmic efficiency gains have independently halved the compute required for a given performance level approximately every eight months.

Third, the threshold phenomenon describes a task-specific point of usefulness determined by the cost of checking and correcting model output. Below this threshold, improvements are difficult for users to appreciate because verification costs dominate perceived benefit; above it, the task becomes delegable and progress appears discontinuous. Because different tasks carry different verification costs, thresholds are crossed asynchronously across a domain, implying that product readiness must be maintained continuously rather than achieved once.

3. Core Analysis

3.1 The Capability Overhang and Evidence of Product-Driven Adoption

The clearest empirical evidence for the capability overhang comes from Codex usage data. Non-engineering usage share - spanning finance, legal, and recruiting - rose from approximately 0% to approximately 90% following the launch of the Codex desktop application, not following a new model release. This finding is significant because it isolates product experience, rather than underlying model capability, as the proximate driver of adoption. It suggests that in many professional domains, the binding constraint on value realization is not whether the model can perform the task but whether the surrounding product makes that performance accessible, trustworthy, and workflow-compatible.

3.2 Architecture of a Finance-Specific Deployment

The finance-tailored ChatGPT configuration was built around three pillars: financial data access, a finance-specific interface, and firm-specific output formatting. On the data side, OpenAI opted to license financial data directly and build proprietary hosting and search indices using the same infrastructure underlying its web index, rather than relying solely on existing MCP connectors, which exhibited latency and quality limitations. Notably, MCP is still used for model access to this data; the architectural change was in controlling hosting and indexing quality rather than abandoning the protocol.

A second architectural decision concerns the role of post-training relative to protocol-level access. MCP and skills provide instructions and access - analogous to handing an instruction manual to a junior analyst - whereas post-training builds the model's intrinsic familiarity with specific datasets, analogous to a senior analyst with two decades of experience who knows the data's quirks firsthand. This distinction has direct implementation consequences: tasks requiring nuanced judgment about data idiosyncrasies benefit more from post-training investment, while tasks requiring broad but shallow access benefit from MCP-based connectors.

The product layer itself - interactive financial charts, citation viewers, and firm-approved templates for Excel and PowerPoint output - was built using the same ChatGPT plugin platform available to external developers, including MCP tools and connectors, MCP apps, skills, and shared artifact tools. This is a notable design choice: rather than building a bespoke internal stack, the finance product was constructed on infrastructure external developers can also access, suggesting a belief that comparable domain-specific products are replicable by third parties.

3.3 Capability Trajectory and the Threshold Phenomenon in Practice

Mathematical benchmarks offer a concrete illustration of the pace of capability growth: a 2023 model scored 700/800 on SAT high school math; a 2025 model achieved International Mathematical Olympiad gold-medal status; an early-2026 internal model disproved a long-standing 1946 conjecture; and a subsequent internal model proposed a solution to the Navier-Stokes Millennium Prize problem, producing 719 manuscripts across 372 families of results. This trajectory was sustained despite earlier concerns about a data wall, addressed through synthetic data generation and reinforcement learning from verifiable rewards, and context-length limitations, addressed through sparse attention and compaction techniques.

Applied to finance, this trajectory is framed through a rising tide and waves model: general-purpose gains in math, coding, and reasoning lift financial capabilities broadly (the rising tide), while finance-specific post-training investment creates targeted performance peaks on top of those gains (the waves). Near-term investment is directed at financial reasoning, Excel modeling, and PowerPoint generation quality; longer-term investment targets judgment-intensive tasks such as earnings forecasting, commodity price assessment, and investment evaluation, with the explicit goal of matching best human expert performance by 2027.

4. Technical Insights

Several implementation-level findings merit attention for builders of domain-specific AI products. First, data infrastructure quality - latency, indexing fidelity, and search accuracy - functions as a binding constraint on usability independent of model capability; building proprietary hosting and indexing addressed limitations in existing MCP connectors without abandoning the protocol itself. Second, the choice between post-training and protocol-level access (MCP/skills) should be guided by task character: post-training is warranted where deep, idiosyncratic familiarity with a fixed dataset yields disproportionate value, while MCP-based access suffices for broader, instruction-following tasks. Third, output formatting - firm-approved Excel and PowerPoint templates, interactive charts, citation viewers - appears to be as consequential as underlying model quality in determining whether outputs are usable in regulated workflows. Fourth, the empirical link between the Codex desktop launch and adoption inflection indicates that measurement of product-level metrics (interface launches, workflow integration points) may be as informative as model benchmark tracking for forecasting value realization. A limitation of this evidence base is its reliance on internal, largely unaudited usage statistics and benchmark claims (e.g., the Navier-Stokes result) that are not independently verified in the source material.

5. Discussion

These findings suggest that capability and value realization are decoupled processes governed by different constraints: capability growth is driven by compute, algorithms, and unhobbling, while value realization is governed by product design and threshold crossing. This decoupling has implications beyond finance - any knowledge-intensive domain with high verification costs (legal, medical, scientific research) likely exhibits a similar overhang, where product investment, not model upgrades, is the near-term lever for adoption.

A further implication concerns the next unhobbling frontier: proactive, persistent agents that hold goals and work in the background, reducing the need for users to learn new prompting workflows. If realized, this would shift the interaction model from prompt-response toward something resembling delegation to a colleague, which may itself redefine what "threshold of usefulness" means for a given task.

Open questions remain regarding the generalizability of the 2027 timeline, the reproducibility of claimed mathematical breakthroughs, and whether architectures combining proprietary indexing with post-training will scale economically across other regulated industries with similarly strict data governance requirements.

6. Conclusion

This analysis has argued that the capability overhang, not model capability itself, is the primary near-term constraint on realized value from frontier models in finance. The case of a finance-tailored ChatGPT deployment illustrates that closing this gap requires proprietary data infrastructure, deliberate choices between post-training and protocol-level access, and domain-specific output formatting - built atop the same plugin platform available to external developers. Practically, builders should prioritize product experience and data quality alongside capability tracking, anticipating that financial tasks will cross usefulness thresholds asynchronously through 2027, with judgment-intensive tasks such as earnings forecasting representing the final frontier for parity with top human experts.

Sources

• How OpenAI Built ChatGPT for Bankers and Analysts - Joseph Kim, OpenAI - Original Creator (YouTube)
• Analysis and summary by Sean Weldon using AI-assisted research tools

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub