Bank of America’s AI Warning: The Model Is Not the Product. The Audit Trail Is.

Posted on by

(SeaPRwire) –

By: Oliver Hawthorne

The AI debate has a dirty secret. Most companies do not have an AI problem. They have a decision problem. At the inaugural AIQ Summit on Thursday, two executives made that clear. Hari Gopalkrishnan leads technology and information at Bank of America. Sally Moore is chief client officer and co-head of market intelligence at S&P Global. They did not pitch faster model deployment. They pitched trust. Regulated industries cannot treat AI as a lab toy. A wrong answer can hurt a customer. It can trigger a compliance review. It can stain a brand for years. Gopalkrishnan said the starting point is not the model. It is the client need. Does AI actually answer the problem? That question sounds simple. It is not. Many boards are under pressure to show AI progress. They greenlight pilots before defining the problem. Vendors encourage that rush. The result is a sophisticated system that solves nothing. Or worse, it creates a new risk surface. Moore framed the other half of the anxiety. Data is not a footnote. It is the foundation. The discussion followed this week’s release of the AIQ 75. It is an annual ranking of 500 companies generating measurable impact from AI. ServiceNow was part of that release. That ranking signals a shift. The market is moving from AI theater to AI evidence. The core contradiction is brutal. AI promises scale. Regulated work demands accountability. Those two forces do not naturally align. Governance is not a brake on AI success. It is the price of entry.

Bank of America evaluates AI implementations across 16 risk areas. Those include privacy, bias, workforce implications and intellectual property. That is not a legal checklist. It is a design constraint. Gopalkrishnan said one of the biggest mistakes is rushing to AI when deterministic models can do the job just as well. That line should be printed on every AI budget request. The most expensive mistake may not be moving too slowly. It may be funding a sophisticated AI solution where simpler technology would yield a more reliable, lower-risk return. When AI is warranted, governance must be continuous. Set guardrails. Evaluate results. Monitor systems for unintended consequences. A fraud model that disadvantages certain demographic groups is not just a tech bug. A chatbot that does not adequately serve customers with particular accents is not just a product gap. Both create customer, compliance and reputational risks. S&P Global starts with data. Moore described work with a Tier 1 bank. The company combined proprietary intelligence, subject-matter expertise and implementation support. S&P Global helped the bank reduce its time to market by roughly sixfold. It also helped improve accuracy in work drawing on multiple data sets from about 60% to 98%. Those numbers are impressive. They are also incomplete. Accuracy alone is not enough. Moore said you need to ensure you can source the original IP that informs some of those decisions. In high-stakes applications, the ability to trace an AI-generated output back to the underlying data and intellectual property is essential. That is the line between a demo and a deployable system. A bank cannot tell a regulator that the model “seemed right.” It must show the source. It must show the reasoning path. It must show who owns what. The AIQ 75 ranking rewards measurable impact. That metric will separate vendors selling magic from vendors selling proof.

The commercial loop is already changing. Model builders sell capability. Data providers sell provenance. Governance platforms sell control. Banks and insurers buy risk reduction. The buyer with the most to lose will set the terms. In high-stakes markets, the winning stack will not be the one with the largest parameter count. It will be the one with the cleanest audit trail. That means proprietary data has a new premium. Subject-matter experts have a new premium. Implementation support has a new premium. A generic AI wrapper has almost no moat. It can be copied. It can be swapped. It can be regulated into irrelevance. The endgame is a two-tier AI market. Low-stakes workflows will get cheap, fast and disposable AI. High-stakes workflows will get slow, expensive and traceable AI. Banks will not deploy black boxes for credit, fraud or customer service. They will demand deterministic fallbacks. They will demand provenance. They will demand monitoring after launch. Vendors that treat governance as paperwork will lose. Vendors that treat governance as product architecture will win. The AIQ Summit conversation points to a harder truth. The companies that succeed with AI are not the ones moving fastest. They are the ones that know when not to use it. They know when a simpler model is safer. They know when a human must stay in the loop. They know when data rights are more important than model novelty. That is not a sexy message. It is the only message that survives contact with a regulator. If your AI pitch cannot survive a compliance review, you are not selling innovation. You are selling a future incident report.

Author bio: Oliver Hawthorne, a principal correspondent at an international technology review covering enterprise AI, governance, and data economics.