Fluent Isn't Trustworthy: What the PensionBee Study Says About AI Financial Advice

Most people won't buy a product on Amazon without checking the reviews. They'll compare star ratings, scroll through photos, and read the one-star complaints. But when an AI tells them how to roll over a 401(k) or when to start taking Social Security, more than half of AI users say they'd act on it without checking.

I've read a lot of surveys about AI adoption this year, and most of them tell you what you'd expect: usage is up, trust is increasing, younger people are more fluent than older people. PensionBee's new research on AI and personal finance has all of that too, but buried in it is the number I'd put in the headline: 57% of the AI users surveyed say they'd act on an AI tool's financial advice without checking it first. You can read PensionBee's full report here. Put simply, the most dangerous thing about AI financial advice isn't that it's sometimes wrong. It's that it always sounds right.

I read the survey from the perspective of someone who is responsible for how one of these tools behaves. I'm the CEO of Datalign Advisory, an SEC-registered investment adviser that helps RIAs with organic growth, and we built HaloAI, a platform for consumers and financial advisors.

PensionBee, an online IRA provider, released the survey on September 17. It covers 1,000 U.S. adults who use AI chatbots for help with their personal finances. The research firm Attest surveyed them online July 18–21, 2026, using an opt-in sample balanced by age, gender, and region. For scale, about one in five U.S. adults have used an AI tool for personal finance information, according to the TIAA Institute-GFLEC Personal Finance Index.

The step that disappeared

That 57% figure says less about whether AI gives right answers and more about a step people are dropping: the pause to compare, check, or ask for a second opinion. That pause has protected people from bad financial decisions for as long as financial advice has existed. It's where mistakes get caught.

And these tools are good at changing minds, mostly with facts. MIT's David Rand and his colleagues showed this in two studies published in Science and Nature in December 2025, as New Scientist reported. In one, a six-minute conversation with an AI changed voters' views on legalizing psychedelics by about 10 points on a 100-point scale, more than twice as much as a video ad. In the other, a study of nearly 77,000 people in the U.K., the 19 AI models tested persuaded best when they used factual claims.

Those studies measured political views, not financial decisions. What worries me is the pairing: a tool that wins people over with "facts,"and users who have stopped asking whether those facts are actually correct.

Some of the decisions people hand over are the wrong ones to hand over

What surprised me most in the data wasn't the average. It was the breakdown by stakes. People are most willing to hand over the low-cost, easily reversible decisions: a little over a third would go ahead with an AI tool's suggestion on monthly savings amounts, and about a quarter on canceling a subscription or the order to pay off debts. Those are decisions you can undo in a week if they turn out badly.

Now look at the decisions that are hard to undo. Fifteen percent would act on an AI tool's suggested retirement age, 16% would make a major purchase, and 18% would change how their investments are allocated. Fewer people skip checking when the stakes are higher, but nearly one in five still would. A wrong answer here can cost years of investment growth or a retirement date you can't take back.

Every firm adding AI to financial advice, including ours, is responsible for that one in five. A firm that rushes a product out to keep up with competitors passes its mistakes to the people who won't check. Before putting AI in front of clients, a firm should know which questions the product will answer, which it will send to a person, and how someone can check where an answer came from.

Why wrong citations matter more than wrong answers

Nearly a quarter of respondents say AI has already given them wrong information about their financial decisions, and 5% found out only after acting on it. That figure will nab the headlines, but a number from outside the survey matters more to me. When the Tow Center at the Columbia Journalism Review tested eight AI search tools in March 2025, the tools misidentified the source of news excerpts more than 60% of the time.

A wrong answer is a bug. A wrong citation, delivered in the same confident tone as a correct one, is a fundamental design problem, because it removes the check a careful person would use: does this answer trace back to a source I can read and trust?

Better questions get better answers, but knowing what to ask takes financial literacy, and financial literacy is at its lowest point in the 10-year history of the TIAA Institute-GFLEC Personal Finance Index. Put those together and you get a tool that answers confidently, citations that often point to the wrong source, and users who, on average, lack the background to ask the question that would catch the error. I don't fault the users for this. They are the people this technology is supposed to help most.

Gen Z hands more decisions to AI and reports more bad advice

It's easy to assume older people are the most vulnerable here, and the study complicates that. Baby Boomers who use AI for financial questions are the most likely to say they've never been misled: 82% of them. Younger users are more comfortable sharing sensitive information and letting AI act for them: 66% of Gen Z would let AI act on their behalf, compared with 47% of Boomers. PensionBee's full report adds that 31% of Gen Z say AI has given them unsuitable financial advice, and 10% noticed only after acting on it.

You could read that as "young people trust more, so they're more at risk," and that's partly right. I'd put it differently: the people most comfortable handing off a decision often have the least experience of what a bad financial decision costs over time. Comfort with an AI tool and readiness to live with its mistakes are separate things.

More than half worry about privacy, and many share their income anyway

More than half (53%) of the AI users surveyed report moderate or extreme concern about privacy, and yet many share sensitive financial data with AI anyway. Nearly a third share budget details, close to 30% share their income, and a small number paste in account numbers or a Social Security number. (Let me be clear: don't do this. Just don't.)

People understand the tradeoff and make it on purpose, because a better answer needs more context. Companies can fix that, but only if they design for it from the start instead of treating data collection as a side effect of giving a good answer.

Fluency is not the same as trust

This is the idea underneath everything above. Modern models are extremely good at sounding right. That's what they're optimized for. But sounding right and being checkable are two different properties, and in finance only one of them protects the client.

There's a line you hear around MIT's CSAIL that I keep coming back to: you don't trust a system because it's smart, you trust it because you can check it. The PensionBee data describes a world where people trust the system because it's fluent. And nobody can check it.

Why we built HaloAI around friction, not just fluency

This is where I'll talk about our own work, because the PensionBee data is surprisingly close to the design documents we built HaloAI against nearly three years ago.

HaloAI's core design decision is that the depth of an answer should match the complexity and potential impact of the question. HaloAI will happily give you a fast, confident answer on the low-stakes, reversible questions: how much you might save by trimming a subscription, whether a smaller debt should be paid off before a larger one. But the moment a query touches something in the hard-to-reverse category the survey flags, like a retirement age, an allocation change, or a major purchase that affects a long-term plan, HaloAI uses a deeper mode grounded in a defined set of sources.

After we introduce someone to an advisory firm, their advisor receives a financial snapshot, the themes of the client's conversations with HaloAI, and the topics they want to go deeper on. The 57% number in the PensionBee study describes a world where the AI and the user are the only two parties in the room for a decision that used to require a third: someone with a duty of care. As the makers of HaloAI and an SEC-registered company, we have that duty of care, and HaloAI was purposely built to be the thoughtful third party in the room.

We separate answering from checking. For source-grounded questions, HaloAI generates its answer from a defined set of sources, including Datalign's own financial-education content, then checks the response against those sources before a person sees it. Our compliance team reviews all of the content HaloAI uses to provide educational guidance. HaloAI can still make mistakes, but the answer is traceable to material a user or reviewer can inspect. That is our response to the Tow Center finding: people shouldn't need specialized financial knowledge just to determine whether an AI answer is backed by the source it cites.

We treat privacy as a design constraint. The survey shows what people care about. 53% are moderately or extremely concerned about the information they share with AI, yet 32% have shared budget details, 29% income, 17% account balances, 4% account or credit card numbers, and 3% a government ID or Social Security number. Our rule: when we use external models, we send only the minimum amount of information a task requires, and we keep sensitive financial and personally identifying information out of that exchange. HaloAI encrypts data at rest and in transit, follows CCPA requirements, and shares information with the advisory firm a person is introduced to in line with our privacy policy. In that interaction, the outside model provider is not acting as the person's investment adviser. If someone is going to trust us with their account balances and income to get a better answer, the least we owe them is a system that treats that information as seriously as the assets themselves.

The standard financial AI should meet

I don't think the PensionBee data argues against AI in personal finance. Far from it. It argues against AI in personal finance that mistakes a confident answer for a trustworthy one. The people surveyed are right to want fast, clear answers to financial questions, and that demand isn't going away. But the simple truth is that the right answer isn't always the fast one.

So back to the one line. The most dangerous thing about AI financial advice isn't that it's sometimes wrong. It's that it always sounds right. Anyone building in this space has three jobs: make sure decisions that can't be undone still get a second look, make sure every claim traces back to a source, and treat the data people share, often reluctantly, with the care it deserves. That's the standard we hold HaloAI to, and it's the standard I think the whole industry should be measured against.