Why AI Still Isn’t Ready for Full Autonomy in Investment Banking

“Why AI Models Suck at Investment Banking” main points:
The article argues that while AI models are impressively capable in broad tasks (poetry, text generation, code), they falter in investment banking largely because training data for specialized workflows is sparse and largely proprietary—there’s inadequate data for IPO execution, M&A diligence, spreadsheet work, etc. [1]. AI’s tendency to hallucinate and accumulate errors across sequential steps undermines its suitability for high-stakes financial work. Real-world ROI remains weak: Forbes cites a MIT study finding that 95% of generative AI pilots don’t move past proof-of-concept [1].

Corroborating and contrasting evidence:
Multiple sources back these claims. A study evaluating 22 models across basic finance tasks finds average accuracy below 50%, underscoring a substantial performance gap in entry-level banking functions [2]. A separate Forbes analysis notes that 95% of financial AI pilots fail to generate P&L impact, even as attention hypes up transformative potential; yet some firms (especially senior decision-makers) leveraging AI correctly outperform due to context and integration [3].
EY’s survey of large companies finds that early AI deployment has led to ~$4.4 billion in combined losses, driven by compliance failures, flawed outputs, and sustainability disruption—not job loss or legal issues [4].

Strategic implications:
— Investment banks and deal firms should temper expectations: AI is not ready to replace junior bankers but can assist in amplifying human expertise. Overpromising autonomy may erode trust internally and with clients.
— Focus should shift to acquiring domain-specific data, building feedback loops, and integrating human judgment where context matters. Tasks requiring precision should be adhered to hybrid models—AI + oversight—rather than full automation.
— Risk management and governance are essential: accuracy, auditability, bias, explainability, and regulatory compliance pose rising costs and institutional risk.
— Select pilots carefully: projects with clear ROI and high repeatability (e.g. document processing, memo generation, formatting, initial screening) may succeed; complex, multi-party workflows (e.g. fundraisings, complicated M&A) are less forgiving.

Open questions:
— What models or architectures can better handle small, private, and sequential datasets characteristic of investment banking workflows?
— How to rigorously measure and reduce compounding errors in multi-step AI tasks?
— What regulatory frameworks will emerge to govern AI failures, responsibility, and disclosure in high-stakes financial decisions?
— How can institutions source or create enough high-quality data without breaching confidentiality or incurring prohibitive costs?
— What training or organizational change is needed to bridge the intuition gap between technologists building AI and bankers using it?

Altogether, while the promise of AI in investment banking is real, current limitations are too material to ignore; strategic investment in infrastructure, data, and governance is required to turn potential into consistent performance.