BusinessMarkets, money & strategy
OpenAI's Joseph Kim says most of ChatGPT for Financial Services is replicable, with licensed data and integrations as the edge
Joseph Kim, OpenAI's head of financial services, reportedly says roughly 95% of ChatGPT for Financial Services is standard engineering a startup could replicate. In his view, the real advantage comes from licensed financial data, workflow interfaces and integrations.
From the Latent Space podcast episode of October 31, 2026 and cited sources: Kim reportedly describes most of the product as ordinary engineering built on widely available plugin building blocks. [2] [10]
SOURCES[2] BigGo Finance: Kim on replicability finance.biggo.com[10] Fireside Alpha: ChatGPT for Financial Services firesidealpha.substack.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Kim reportedly describes most of the product as ordinary engineering built on widely available plugin building blocks. [2] [10]
- Original source
- Latent Space podcast episode, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 31, 2026 podcast episode published
- Covered
- Oct 31 edition
- Evidence
- 13 sources across 13 sites; no transcript
- 1⚡ Building ChatGPT for Finance — Joseph Kim, OpenAI's Head…youtube.com
- 2BigGo Finance: Kim on replicabilityfinance.biggo.com
- 3CNBC: OpenAI targets work of Wall Street junior bankerscnbc.com
- 4Bloomberg: Anthropic courts finance industrybloomberg.com
- 5StartupHub.ai: OpenAI is pitching ChatGPT to bankersstartuphub.ai
- 6TechCrunch: ChatGPT for personal financetechcrunch.com
RisingHow it’s measured
YouTube video · 1.9K views · 11 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
Kim reportedly describes most of the product as ordinary engineering built on widely available plugin building blocks. He places its differentiation in licensed data, user interface and integrations. OpenAI licenses, hosts and indexes much of the bundled data itself, which Kim says also lets it post-train models on that data. [2] [10]
The product launched for investment banking and equity research on GPT-6 Astra and was developed with Morgan Stanley and Evercore. It comes with Daloopa, PitchBook and LSEG News data and more than 50 MCP connectors. It produces cited research, Excel models and decks in a firm's own template. OpenAI's vendor-reported OfficeQA Pro score of 69.9% has not been independently replicated. [3] [7] [8]
Kim reportedly treats connector reliability and evaluation as open problems. He says some large-provider connectors failed more than 40% of the time, while OpenAI's published baselines are 5–10%. He also calls existing finance benchmarks saturated or flawed, citing wrong answer keys and poorly specified prompts. [2] [8]
the product bundles licensed data and over 50 connectors, and Kim describes flawed benchmarks. Implication: advantage in finance AI may depend on data licensing and connector reliability, so buyers should measure error rates on their own workflows.
I've finished gathering sources and am writing the dossier now.
Executive brief
OpenAI's Head of Financial Services, Joseph Kim, reportedly says that about 95% of ChatGPT for Financial Services is ordinary engineering that any startup could have built on available plugin primitives. In his account, the product's advantage comes from licensed data, UI and integrations (BigGo Finance). The product launched on September 10, 2026. It runs on GPT-6 Astra, has built-in data from Daloopa, PitchBook and LSEG News, and was designed with Morgan Stanley and Evercore (CNBC). The commentary under review is Kim explaining how the product was built. No public pricing, seat counts or audited accuracy figures exist.
What changed and event timeline
Anthropic enters finance first
Anthropic launched Claude for Financial Services, connecting Claude to FactSet, PitchBook and Morningstar data. This set the competitive baseline OpenAI later answered ().
To 2026-06: Internal Codex use surges
Kim says OpenAI's internal Codex use by finance, legal and recruiting staff went from near zero to roughly 90% after the February Codex desktop app release ().
Consumer finance first
OpenAI launched personal-finance features in ChatGPT that let users connect bank accounts ().
Wall Street product launches
ChatGPT for Financial Services launched for investment banking and equity research, built on GPT-6 Astra. OpenAI declined to name customers (;).
Kim's AI Engineer talk covered
Coverage of the interview noted that Kim gave no paid seat counts, pricing or audited error rates for generated models and decks ().
Kim's Latent Space episode
A secondary write-up reports Kim's claims on replicability, connector failures and broken finance benchmarks ().
Capabilities and access
- Model: GPT-6 Astra, with high, medium and low effort settings that change token use and cost (Fortune).
- Outputs: cited multi-source research, Excel models, PowerPoint decks in the firm's own template, and web dashboards. Fortune saw sample decks take about 10 minutes.
- Data: bundled Daloopa, PitchBook and LSEG News. Sign-in access to S&P Capital IQ, LSEG, MSCI, Factiva and Moody's.
Read the full section
- Model: GPT-6 Astra, with high, medium and low effort settings that change token use and cost (Fortune).
- Outputs: cited multi-source research, Excel models, PowerPoint decks in the firm's own template, and web dashboards. Fortune saw sample decks take about 10 minutes.
- Data: bundled Daloopa, PitchBook and LSEG News. Sign-in access to S&P Capital IQ, LSEG, MSCI, Factiva and Moody's. More than 50 connectors via MCP, OpenAI's chosen way of plugging in outside data tools (VentureBeat).
- Access: enterprise accounts at "eligible institutions" only. No public pricing (OpenAI).
Technical analysis for researchers and developers
- Data layer: OpenAI licenses, hosts and indexes much of the bundled data itself.
- Three ways to reach data: bundled datasets, the customer's existing vendor logins, and MCP connectors.
- Governance: template administration for Office files, information barriers between workspaces, SAML SSO, SCIM provisioning and role-based access.
Read the full section
- Data layer: OpenAI licenses, hosts and indexes much of the bundled data itself. Kim says this also lets OpenAI post-train its models on that data (Fireside Alpha).
- Three ways to reach data: bundled datasets, the customer's existing vendor logins, and MCP connectors. OpenAI reports tuning the connectors to cut errors, for example FactSet from 9.59% to 6.45% and S&P Global from 6.84% to 2.66% (VentureBeat).
- Governance: template administration for Office files, information barriers between workspaces, SAML SSO, SCIM provisioning and role-based access.
- Evaluation: Kim calls existing finance benchmarks saturated or "essentially broken," citing wrong answer keys and poorly specified prompts (BigGo Finance). No public eval harness exists, so outside researchers cannot reproduce results.
Claims and evidence
- OfficeQA Pro benchmark: GPT-6 Astra scored 69.9%, against 60.2% for GPT-5.6 Sol and 62.4% for Claude Fable 5.1.
- Kim reportedly saw some large-provider connectors fail more than 40% of the time (BigGo Finance). OpenAI's published baselines are 5–10%.
- The "90%" figure is reported inconsistently. One write-up applies it to internal Codex use (StartupHub.ai), another to "finance adoption" (BigGo Finance).
Read the full section
- OfficeQA Pro benchmark: GPT-6 Astra scored 69.9%, against 60.2% for GPT-5.6 Sol and 62.4% for Claude Fable 5.1. These are vendor figures with no independent replication, and VentureBeat noted the release lacked enough detail to validate them (VentureBeat).
- Connector error rates conflict. Kim reportedly saw some large-provider connectors fail more than 40% of the time (BigGo Finance). OpenAI's published baselines are 5–10%. The two figures may measure different things.
- The "90%" figure is reported inconsistently. One write-up applies it to internal Codex use (StartupHub.ai), another to "finance adoption" (BigGo Finance).
- Enterprise revenue: CFO Sarah Friar said OpenAI's enterprise business now earns more than its consumer business (CNBC).
Context and prior work
- Anthropic's finance offering came 14 months earlier and also relied on third-party data connectors (Bloomberg).
- OpenAI's VP of Product, Nick Turley, pitched OpenAI's product as "the canonical product" for the industry (Fortune).
- Kim previously led applied AI at the finance AI startup Rogo, and before that worked on Gemini's reinforcement-learning-from-feedback research at Google (Rogo).
Read the full section
- Anthropic's finance offering came 14 months earlier and also relied on third-party data connectors (Bloomberg).
- OpenAI's VP of Product, Nick Turley, pitched OpenAI's product as "the canonical product" for the industry (Fortune).
- Kim previously led applied AI at the finance AI startup Rogo, and before that worked on Gemini's reinforcement-learning-from-feedback research at Google (Rogo).
- OpenAI says it plans similar packages for other sectors (Enterprise DNA).
Limitations, safety and contested findings
- Goldman Sachs' Chris Churchman warned that automating junior bankers' training tasks risks "cognitive atrophy" in the next generation (CNBC).
- Kim reportedly says hallucination now comes up less often and that judgment and taste are the bigger gaps. These are claims, not measurements.
- Bundling more than 50 connectors means buyers inherit many vendors they never vetted (Invela).
Read the full section
- Goldman Sachs' Chris Churchman warned that automating junior bankers' training tasks risks "cognitive atrophy" in the next generation (CNBC).
- Kim reportedly says hallucination now comes up less often and that judgment and taste are the bigger gaps. He expects agent oversight to copy banking's analyst-to-managing-director review chain (BigGo Finance). These are claims, not measurements.
- Bundling more than 50 connectors means buyers inherit many vendors they never vetted (Invela).
- Accuracy on live deals has not been audited.
Business and practitioner implications
- Startups: if Kim is right that about 95% of the product is reproducible, their advantage is licensed data, workflow UI and distribution, not model access.
- Banks: ask for error rates on your own workflows before rollout. Public benchmarks and vendor connector figures are weak evidence.
- Data vendors: bundling by OpenAI moves negotiating power toward model platforms that license and post-train on financial data.
Read the full section
- Startups: if Kim is right that about 95% of the product is reproducible, their advantage is licensed data, workflow UI and distribution, not model access.
- Banks: ask for error rates on your own workflows before rollout. Public benchmarks and vendor connector figures are weak evidence. Also plan how analysts will keep building skills.
- Data vendors: bundling by OpenAI moves negotiating power toward model platforms that license and post-train on financial data.
- Developers: connector reliability is a first-order engineering problem. Measure it per provider rather than assuming MCP connectors work as advertised.
The source trail.
Sources (13)
⚡ Building ChatGPT for Finance — Joseph Kim, OpenAI's Head of Financial Services
Metadata only. Transcript unavailable: Audio transcription returned HTTP 429.
www.youtube.com