Reframing AI in Investing: Humans Still Call the Shots

AI should be an assistant, not a scapegoat. The Wealth Management piece, “Advisors Should Tread Carefully in Using AI for Investment,” gets one big thing right: caution beats hype. But it treats “tread carefully” like a user tip, not a governance mandate. Telling advisors to be careful without asking who owns the mistake is how you end up with a lot of fingers pointing at a model and no one signing the loss letter.

Let’s start with the part the article handles reasonably well: skepticism about black boxes and overreliance. That’s fine as far as it goes. But “don’t rely too much on AI” is about as actionable as “don’t eat too much sugar.” It’s technically sound and practically useless unless you define limits, controls, and consequences.

The real tension isn’t about curiosity; it’s about fiduciary liability. Advisors can’t outsource duty of care to a vendor’s model and call it modernization. When an AI-driven allocation underperforms or misfires, someone still has to sit across from the client, explain what happened, and document why that decision met a standard of care. If the answer boils down to “the software said so,” that’s not a defense — that’s an admission you handed judgment to a tool you didn’t control.

I’ve sat in model risk meetings where a pretty dashboard and a three-page methodology deck were somehow treated as due diligence. Back at Goldman, I learned quickly that if you can’t see the inputs, assumptions, and edge cases, you don’t actually understand the output — you’re just renting conviction from someone else’s code. Wealth advisors using AI are in the same position now: if you can’t explain in writing how the model behaves in different environments, you’re not supervising it; you’re hoping it’s right.

That’s the gap in the Wealth Management piece: it nods at caution, but barely touches the machinery of governance. You need version control on models, clear data provenance, change logs, and documented human sign-offs for material portfolio shifts. You need a written policy for how and when AI outputs can be overridden — not as an afterthought, but as a core part of the investment process. Without that, “tread carefully” is just risk marketing.

The column also soft-pedals how correlated model risk actually becomes once vendors get traction. When a few big providers push similar models into dozens or hundreds of advisory shops, the industry starts to move like a synchronized swimmer. It’s not one advisor making a bad call; it’s the same call echoing through a lot of supposedly “independent” portfolios. That’s not individual incompetence, that’s crowding in code form.

And then there’s the data problem. AI trained on curated historical sets will be exquisitely tuned to the past and reliably startled by regime shifts. Models extrapolate; markets ambush. You can chase interpretability, but interpretability doesn’t grant immunity. It buys you a postmortem, not a shield. Right now, too many people treat “we can explain the model” as a green light, when it should just be table stakes before you even start talking about deployment and stress scenarios.

The Wealth Management article briefly touches on client-facing benefits, but glosses over the slow erosion AI can cause in the advisor-client relationship. Clients are not paying for access to a screen. They’re paying for someone who can absorb that screen’s output, contextualize it, and sometimes ignore it. If the pitch effectively becomes “our algorithm is smarter than the next firm’s algorithm,” then fee compression is inevitable. Machines sell scale. Humans are supposed to sell discernment. Mix them carelessly and you’ll end up justifying your fees against a cheaper robo-advisor that looks uncomfortably similar to your own tech stack.

There’s also a missed competitive angle: governance itself becomes a differentiator. The winners won’t be the firms with the fanciest AI branding; they’ll be the ones that can walk a regulator or a skeptical client through their process step by step. Think of how some managers now market their compliance infrastructure and risk teams as features, not overhead. The same thing will happen with AI: the advantage goes to the advisor who can say, “Here’s exactly how this tool fits into our process, where humans intervene, and how we document all of it.”

History has seen this movie before. Quant funds didn’t blow up because math was evil; they blew up when everyone chased the same signals with similar models, and risk controls lagged the enthusiasm. AI in wealth management is just a friendlier interface on the same old problem: complex tools used without equally complex oversight. The labels have changed; the structural vulnerability hasn’t.

Now for the standard counter-argument: AI can scale advice, lower costs, and detect patterns no human will ever see. All true. But speed and scale without matching governance just accelerate the path from small error to large problem. You can absolutely use AI to reduce operational drag and research time while still building audit trails and escalation points. What you can’t do is let “efficiency” become the excuse to thin compliance and pretend nothing fundamental has changed. The math doesn’t lie: once you embed AI into day-to-day decisions, your risk profile changes, whether you acknowledge it or not.

So what should advisors actually insist on?

• Vendor agreements that spell out data sources, update cycles, and ownership of model changes.
• Independent backtesting and stress testing done by the firm, not just consumed from marketing decks.
• Human-approval gates for any AI-triggered move beyond a defined threshold.
• Plain-language disclosures to clients about when AI is used, what it does, and what it doesn’t do.

Wealth Management is right to tell advisors to tread carefully with AI; the piece just understates how much that carefulness has to be engineered, not wished into existence. The advisors who treat “tread carefully” as an operating blueprint, not a headline, will be the ones still explaining their process with a straight face when the first big AI-driven misstep hits the industry.