By David Trainer, CEO, New Constructs
Most trading desks these days are running some version of AI—screening ideas, summarizing filings, drafting the first pass of a model. Fewer of them can tell you, with a straight face, whether what came back is actually right.
A 2026 HSBC survey of nearly 10,000 affluent and high-net-worth investors put a number on that unease: 73% now use AI for finance and investment decisions, but only 12% said it was the most influential factor in their last investment call. People are reaching for the tool constantly and trusting it selectively. That’s not a contradiction—it’s the right instinct, and trading desks would do well to hold onto it.
I’ve spent close to three decades reading 10-Ks and picking apart the accounting behind reported earnings, and the pattern here is a familiar one. The tools that sound most confident are rarely the ones that have actually done the work. That was true of glossy sell-side research a generation ago, and it’s just as true of a chatbot summarizing a footnote today.
Sounding Right Isn’t the Same as Being Right
Here’s the uncomfortable part. General-purpose AI models are built to sound helpful, not to be correct. Ask one about a company’s margins and it will hand back a clean, confident paragraph regardless of whether it caught the off-balance-sheet lease, misread a footnote, or quietly folded a one-time gain into “core” earnings. The prose doesn’t hedge just because the analysis underneath it is shaky.
That’s tolerable when you’re brainstorming. It’s a real problem once the output feeds a screen, a signal, or a risk model that trades without a human re-reading the filing. One bad number doesn’t stay in an analyst’s notebook anymore. It moves through every process downstream that assumes the input was already checked—and the biggest failures rarely come from an obvious hallucination. More often it’s something small: a liability moved off the balance sheet, a non-recurring charge treated as recurring, an adjustment nobody bothered to question because the answer arrived so quickly and so cleanly.
The Money at Stake Isn’t Small
This isn’t a hypothetical worry about some future AI bubble. Reuters’ coverage of Barclays and BofA Securities forecasts found that AI-related spending is a primary driver behind a projected 30% jump in net investment-grade bond issuance in 2026, as hyperscalers borrow heavily to fund the data centers behind all this new AI capacity.
Which sets up an odd loop: the same AI buildout that’s reshaping credit and equity markets is also producing the research tools desks now use to size up that buildout. If a tool can’t show its work, a desk ends up pricing an AI-driven market using AI-generated analysis it has no real way to check. That’s not a comfortable spot to be in when a client, or a regulator, asks how you got there.
Regulators Have Already Started Asking
FINRA’s 2026 Annual Regulatory Oversight Report puts generative AI governance squarely on the list of things examiners are watching, and it’s specific about it: vendor due diligence, documented model risk, an ability to explain how a tool arrived at its conclusion. Firms that can’t answer those questions cleanly should expect the exam to run long.
Desks adopted these tools because they’re fast. Keeping them will depend on whether they can be defended—and that’s a different bar, one that rarely shows up in a product demo.
Four Questions Worth Asking Before You Trust the Output
Not every AI research tool deserves the black-box label, but plenty earn it. Here’s a short list of questions to help flush out bad AI:
- Can it point to the primary source? A number should trace back to a filing or a trade record, not stop at a tidy summary.
- Does it admit what it doesn’t know? Missing or conflicting data should get flagged, not quietly papered over with a plausible-sounding guess.
- Does the answer hold up if you ask it differently? Rephrasing a question shouldn’t change the conclusion you get back.
- If it’s wrong, who’s accountable? Ideally a person—not “the model.”
Trust Is the Scarce Resource Now
Speed stopped being the differentiator a while ago. Every desk has access to tools that can spit out an idea in seconds. The real value-add comes from being able to show a risk committee, a client, or an examiner exactly how you developed that idea. As AI keeps working its way into surveillance and pre-trade analytics, auditability of ideas and data only gets more expensive to ignore. The firms that come out ahead of the next round of scrutiny won’t be the ones that moved fastest. They’ll be the ones that could always show their work.
The Next Big Thing In AI: Data Moat
Across social media, there’s a new buzzword: “data moat”. As Deepseek and Kimi reveal that the AI models have pricing power, people are looking for where the value-add exists in the AI supply chain. THE VALUE IS IN PROPRIETARY DATA. If an AI has it, it can’t show all of it, but it can provide examples or proof points.
Traditional asset managers can no longer rely solely on marketing scale and distribution to win. They must have proprietary data and rigorous fundamental research to survive. The same applies to AI models.
About the Author
David Trainer, the CEO of New Constructs, is a Wall Street veteran and corporate finance expert with more than 25 years of experience in fundamental analysis, valuation, and financial statement research. For over two decades, he’s challenged conventional investment research by pairing forensic accounting with artificial intelligence to get at the real economics behind public companies—pushing investors to look past accounting distortions, market narratives, and headline metrics toward what a business is actually worth.
As founder and chief executive of an independent equity research firm, David built his practice on a simple premise: AI-driven analysis is only as trustworthy as the data and methodology underneath it. Before starting the firm, he spent more than six years on Wall Street, including roles at Credit Suisse First Boston and Epoch Partners, where he developed proprietary valuation frameworks and led efforts to apply economic earnings analysis across industries. A former member of the Financial Accounting Standards Board (FASB) Investor Advisory Committee, he’s also the author of Modern Tools for Valuation (Wiley Finance). His research has been recognized by Harvard Business School, MIT Sloan, and Ernst & Young, and he’s a frequent commentator on market trends, valuation, AI in finance, and investment risk across leading financial media.
