The numbers tell a stark story. S&P Global reports that 42% of companies abandon AI initiatives before production, a stark rise from 17% the previous year. This is compounded by BCG showing 74% struggle to achieve value, and Gartner predicting 50% of generative AI projects will be abandoned at the pilot stage. 

While an impressive 78% of organisations use AI in at least one function, a mere 1% believe they're at maturity. In capital markets, these aren't just disappointing statistics. They are signals of a fundamental mismatch between what generic AI offers and what financial enterprises actually need. 

The Specific Challenges of Capital Markets 

ChatGPT impresses in demos and pilots show promise, but the gaps become clear when applied to complex financial workflows. The question isn't whether generic AI is powerful; it's whether power without precision matters when the stakes are regulatory compliance and capital allocation.

Capital markets face challenges that make generic AI insufficient in four critical ways: 

1. The High Cost of Error 

Consider when an analyst asks for Q3 exposure to the tech sector. The AI returns $50M, a number that looks reasonable, passes the gut check, and gets used in risk committee presentations. The actual exposure, however, was $200M. The query calculated this exposure using the wrong methodology: the wrong time period, the wrong aggregation logic, or the wrong entity resolution. 

In marketing, a wrong recommendation might mean a poor performing campaign. In capital markets, it means compliance violations, misallocated capital, and regulatory scrutiny. Generic AI simply wasn't built with these stakes in mind. 

2. The Structured and Unstructured Data Split 

Financial decisions require a synthesis of both structured data (from databases, trading systems, and CRMs) and unstructured data (like research reports, contracts, and earnings transcripts). Although 90% of enterprise data is unstructured and grows 3x faster, less than half is used in generative AI today. 

Generic AI treats these as separate problems: query databases with one approach, search documents with another. Financial workflows, however, don't work that way. You need portfolio positions from your database and the latest analyst report on those holdings while understanding what might change your calculation methodology. Fragmenting these inputs means fragmenting your answer. 

3. The "Looks Right But Is Wrong" Problem 

This is the killer problem. A portfolio manager might ask an agent to calculate weekly returns for a basket of securities. The agent generates a 40-line SQL query with multiple joins and proper syntax, which executes without errors and populates a chart. Everything looks correct. 

Except, the agent used an average to aggregate daily returns over the week when it should have used a sum. Or perhaps it matched "Apple" to the wrong entity in your CRM because five companies have "Apple" in their name. It might even pull a research report from 2022 instead of 2024 simply because the older document had higher cosine similarity to the query - similarity is not relevance, but generic AI doesn't know that. 

It's impossible to eyeball these errors. The output appears valid and the logic seems sound, but the foundation is wrong, and every decision built on it compounds the mistake. 

4. The Lack of Auditability 

When auditors ask how you calculated exposure, "ChatGPT generated it" isn't an answer. They need to see the full lineage: Who asked what question? What data sources were accessed? Which queries were executed? What reports were retrieved? What logic was applied? 

Generic AI provides lacklustre auditability. Even when logs exist, reconstructing the decision path is difficult, as you can't easily verify which SQL ran, which documents were fetched, or why the system chose one calculation method over another. Trusting answers without a complete audit trail isn't just risky; it is not enough in regulated markets. 

The Limits of Generic Tools 

Over 80% of financial organisations are integrating AI, but "integrating AI" and "having AI that works in production" are two very different things. This is evidenced by 60% of AI leaders citing legacy integration and compliance as primary barriers, and 95% struggling with the hybrid multi-cloud environments that house sensitive data. 

ChatGPT, Claude, and Gemini are remarkable tools that democratised AI and showed what's possible. They were, however, built for general use cases, not capital markets. They don't understand entity resolution in financial databases. They can't distinguish between similarly named securities. They don't know your firm's calculation methodologies or data science standards, and they can't provide the audit trail regulators require. 

The architecture that works for writing emails or summarising articles simply doesn't extend to querying complex financial databases or synthesising structured and unstructured data with regulatory precision. 

The Reality of Production 

Most firms are still trying to make off-the-shelf AI work, with teams adding prompts, building wrappers, and implementing guardrails. The hope is that enough engineering around generic AI will eventually make it work for financial workflows. 

But prompt engineering can't solve architectural limitations. You can't prompt your way to entity resolution. You can't add guardrails that teach an AI your firm's specific calculation rules. You can't wrapper your way to full auditability. 

The firms that are actually in production (not just running pilots, but deploying AI at scale in capital markets) built differently from the start. They recognised that finance needs AI with domain knowledge embedded at every layer, not bolted on through prompts. 

The choice is now simple: either attempt to retrofit general-purpose tools with ever-more complex wrappers and prompt engineering, or acknowledge the need for purpose-built architecture.

The failure of generic AI is not a technological fault, but an architectural imperative. To move beyond promising pilots and achieve compliance and precision at scale, financial intelligence must be embedded from the ground up. This is the foundation of the Domain Knowledge Stack, which we will explore in detail in Part 2.