So here's my thesis: the future of AI in finance isn't the smartest model, it's the most auditable one.

There's a problem with AI agents in finance that doesn't get enough attention, and it's not the one most people expect. Building the agent is the easy part. Getting it through an audit is the hard part. This is one of the most important challenges facing anyone deploying AI in a regulated financial environment, and I think it deserves far more scrutiny than it's getting.

Here's why. AI models are probabilistic. Ask the same question twice and you can get two different answers, depending on the model version, the temperature setting, token limits, even how fast the API responds. Regulators don't work that way. They want to pull up a decision from six months ago, replay it, and see the exact same reasoning every time.

That chasm is real. And for financial advisors, it isn't abstract at all. Let's figure out how to cross it together.

The cost of being "usually right"

A few weeks ago I wrote about the PensionBee 2026 study on AI and personal finance. The numbers have stuck with me, and if you've suffered through a call with me in the last couple of weeks, I probably regurgitated the stats below at least a dozen times per hour.

Nearly six in ten Americans who use chatbots for financial advice — 57% — say they'd act on that guidance without checking it. Almost one in four, 23%, say a chatbot has already given them wrong information about their finances. And 5% only found out after they'd acted on it. Separately, a 2025 study by Columbia University’s Tow Center for Digital Journalism found that eight AI search tools misidentified the sources of news excerpts more than 60% of the time.

Put those together. You have a tool that's wrong often enough to matter, used by people who mostly don't double-check, on decisions that are hard to undo — how to invest, when to retire, whether to make a major purchase. This is where a technology shortfall becomes a point of failure as a fiduciary.

Every auditor will eventually ask the same questions. Why did your agent make this recommendation? Can you prove it would make the same one today? If your honest answer is "probably," you've already lost. Simply logging what the AI said doesn't get you there.

The field is starting to say this plainly. As one analysis of AI explainability in finance put it: "Chain-of-thought is not an audit trail. Attention weights are not explanations. Prompt-based reasoning is not reproducible across sessions."

I'd go a step further.

The mistake is treating the model as the place where all accountability must live. A model-generated explanation is not an audit trail — it's generated by the same probabilistic system as the answer it's supposed to justify.

Determinism where it counts

The answer, as I see it, is a validation layer that sits between the AI agent and production. It captures the agent's reasoning, checks it against deterministic rules, and writes an immutable audit trail. The agent can stay creative and probabilistic while it's being developed and tested. But anything that reaches production has a repeatable, checkable decision behind it.

In my view, any system like this has to deliver four things: reproducibility, explainability, auditability, and version control. I wouldn't accept a system that falls short on any one of them.

And in wealth management, that layer can't be something you bolt on after the fact. It has to be the foundation everything else is built on. That's the bet we made with Halo.

How Halo is built

Halo is extensible by design. We have specialized agents for retirement, taxes, mortgages, and investments, and they all plug into a shared core. That core is where compliance lives — not inside each individual agent. In practice, that means a few things.

First, every material claim Halo makes about a client's plan traces back to the client's actual accounts and documents. Not a model's general sense of how a 401(k) usually works. That's the direct answer to the 60% citation problem: don't build a system where the client's only defense is their own financial literacy.

Second, the rules that matter aren't left to the model's judgment. Suitability, concentration, and regulatory constraints are checked deterministically before anything reaches an advisor or a client. The model can propose. The rules decide whether the proposal moves forward.

Third, hard-to-reverse decisions slow down instead of speeding up. When a question touches a retirement date or an allocation change, Halo shows its work and routes the decision to a human advisor with real fiduciary accountability.

And fourth, every decision is recorded — the inputs, the model version, the rules applied, and the outcome. So when someone asks "why," the answer exists, and it's the same answer every time.

Why extensible means fast

This is the part people miss. Most teams think compliance slows you down. Done right, it's what lets you move quickly.

Because the guardrails live in Halo's core, every new agent inherits them on day one. When we build the next one, we're not rebuilding audit infrastructure from scratch. We're writing the domain logic and plugging it in.

The same is true for models. Frontier models are changing every few months. When a better one comes out, we can bring it in, version-stamp it, and test it against our rules and prior decisions before it ever touches a client.

That's the whole philosophy in seven words: probabilistic in the lab, deterministic in production.

There's an idea you hear a lot around MIT's CSAIL: you don't trust a system because it's smart, you trust it because you can check it. Halo is built around that idea. The intelligence can keep improving. The accountability stays fixed.

The right time for advisors

Here's what I keep coming back to. Clients are already using chatbots / agents for financial questions — this isn't going away. So advisors have a choice. They can ignore it and let clients take advice from tools that carry no duty of care. Or they can offer an AI experience that's actually accountable — fast enough to meet clients where they are, with a human and an audit trail behind every decision that counts.

The window for that choice is now, while client habits are still forming. Advisors bring the fiduciary judgment. Halo brings the speed, the traceability, and the record to prove it.

Geoffrey Moore wrote the book on crossing the chasm: the gap between the early adopters who'll try anything and the pragmatic majority who only buy what's proven and safe. AI in finance is sitting right at that gap. The early adopters are already in. But advisors are pragmatists, and they have a fiduciary duty to be. Moore's answer was the "whole product," everything a mainstream buyer needs to adopt with confidence. In regulated finance, auditability is the missing piece of the whole product. Build that layer once, make it solid, and build everything on top of it.

Because in the end, the future of AI in finance isn't the smartest model. It's the most auditable one. That's the standard we hold Halo to, and I think it's the one the whole industry should be measured against.