- 1. OpenAI seeks to automate junior bankers’ grunt work
- 2. OpenAI launches web browser built with ChatGPT
- 3. Users can now tailor Claude capabilities to specific tasks
- 4. Adobe rolls out AI foundry service for enterprise-grade custom models
- 5. Anthropic launches Claude Code on the web
- 6. Google and Yale use AI model to uncover new cancer therapy target
- Recommended reading
The CFO AI Digest: October 21, 2025
AI behemoths continue to expand their reach, integrating AI deeper into the everyday lives of users. In big news for finance workers, OpenAI reveals an internal initiative aimed to replace entry-level banking jobs through automation.
Let’s get into it:
1. OpenAI seeks to automate junior bankers’ grunt work
OpenAI is running an internal project, code-named Project Mercury, focused on replacing the entry-level tasks of junior investment bankers, according to Bloomberg. The company has hired more than 100 ex-bankers from firms like JPMorgan and Goldman Sachs, for $150/hour, to train its AI by writing prompts and building models for time-consuming work such as IPOs and restructurings.
CFO takeaway: Citi mandated AI prompt training across its workforce earlier this month, signaling that bankers must learn to use AI tools. OpenAI takes that a step further by training AI to eventually altogether replace many of the tasks junior bankers perform. This raises questions about the traditional postgrad path in finance. As Matt Levine, a Bloomberg columnist and ex-Goldman investment banker put it, “It’s hard to know exactly what will happen to an industry built on an apprenticeship model when it no longer needs apprentices.”
2. OpenAI launches web browser built with ChatGPT
OpenAI has released ChatGPT Atlas, a web browser with ChatGPT integrated into the interface. The goal is to have the browser function as a “super-assistant” that can use memory (optional and user-controlled) to recall browsing context, personalize suggestions, and help users complete tasks without switching tabs. The agent mode allows ChatGPT to take actions inside the browser, such as booking appointments, making purchases, or compiling documents. Atlas is available for macOS users across all plan tiers, with versions for Windows, iOS, and Android coming soon.
CFO takeaway: Following the Apps SDK launch that brought third-party apps into ChatGPT at DevDay, Atlas brings ChatGPT into the browser itself. As Head of ChatGPT Nick Turley noted, “ChatGPT started as a chat window and is now becoming an operating system that helps you get things done across the web and in more parts of your life.” This move can help OpenAI monetize web browsing, poses another threat to Google's search dominance, and significantly expands its addressable market for subscriptions and enterprise licensing.
3. Users can now tailor Claude capabilities to specific tasks
Anthropic has introduced Agent Skills, a feature that lets users create specialized task modules to improve Claude's performance on specific tasks. Skills are folders that contain instructions, code, and resources that Claude automatically loads when relevant. Claude can then generate Excel spreadsheets with formulas, PowerPoint presentations in line with brand guidelines, or fillable PDFs. Users can build custom Skills once and deploy them across Claude apps, Claude Code, and the API. Anthropic also has prebuilt Skills for common tasks like creating and editing documents.
CFO takeaway: Agent Skills positions Claude as configurable enterprise infrastructure, with embedded automation for repeatable, rules-based work. While most tools rely on prompting and oversight, Claude allows teams to encode institutional knowledge into reusable modules that it activates automatically when relevant. This build-once, deploy-everywhere model also reduces duplication across departments.
4. Adobe rolls out AI foundry service for enterprise-grade custom models
Adobe has launched AI Foundry, a service that lets enterprises build custom generative AI models trained on their own branding and intellectual property. The models are built on Adobe’s Firefly family, trained entirely on licensed data, and can generate text, images, video, and 3D scenes. With a usage-based pricing model, Foundry enables enterprises to repurpose branded content across seasons and languages.
CFO takeaway: AI Foundry targets an enterprise pain point around content localization: adapting campaigns across markets and formats eats significant agency and production budget. Custom models trained on brand assets can generate variations at scale. Adobe's pitch: IP-safe models with brand customization — or what SVP of design Eric Snowden has called "an amazing blend of trust, quality, and customization that is truly unique."
5. Anthropic launches Claude Code on the web
Anthropic has introduced Claude Code on the web, a browser-based tool that lets developers assign and run multiple coding tasks in parallel using Anthropic’s cloud infrastructure. Now in beta for Pro and Max subscribers, it supports GitHub integration, real-time progress tracking, and automatic pull request generation. Each task runs in a secure, isolated environment that restricts file system and network access so Claude can interact only with authorized resources. The tool is especially well-suited for bug fixes, backend updates, and repetitive dev work.
CFO takeaway: By moving from a terminal-based tool to the browser, Claude Code is addressing engineering resource constraints. Teams can now more easily run multiple coding sessions in parallel instead of having engineers work sequentially through maintenance backlogs. This frees up developer capacity for higher-value work like product feature development. It also compresses delivery timelines without adding headcount.
6. Google and Yale use AI model to uncover new cancer therapy target
Google DeepMind and Yale have released C2S-Scale 27B, a model built on Google's open-source Gemma framework, designed for single-cell analysis. The model generated a novel hypothesis: that the drug silmitasertib, when combined with low-dose interferon, could significantly boost the visibility of tumor cells to the immune system. The prediction was later validated in lab experiments, marking a breakthrough in how large-scale biological models can surface previously unknown therapeutic pathways.
CFO takeaway: This marks AI’s shift from pattern recognition to hypothesis generation in drug discovery. For biopharma CFOs, the economics of early-stage drug discovery could change quickly: AI can now run high-throughput virtual screens to surface promising candidates before expensive lab work begins. The bottleneck shifts from generating hypotheses (now AI-accelerated) to experimental validation (still rate-limited by biology).
See you next week.
Recommended reading
- How AI Is Transforming Data Centers and Ramping Up Power Demand (Goldman Sachs)
- ‘Of course it’s a bubble’: AI start-up valuations soar in investor frenzy (Financial Times)
- Is the Flurry of Circular AI Deals a Win-Win—or Sign of a Bubble? (Wall Street Journal)
