In Hangzhou, a 72-year-old retiree named Mrs. Zhang kept 180,000 RMB sitting in a basic checking account for more than five years. She had a steady pension and no pressing need for all that cash. When asked why she avoided even the safest low-yield products her bank offered, she didn’t cite fear of losing money. She said the product terms confused her, she wasn’t sure whether her money would be locked up, and she’d rather not think about it.
That explanation, repeated by millions of older savers in China, points to a puzzle that traditional finance has trouble explaining. Elderly customers in China lose more than 50 billion RMB to fraud each year while simultaneously leaving vast sums in accounts that earn almost nothing. They appear both too trusting and too cautious at the same time. A new study in Computers in Human Behavior Reports proposes a single idea to make sense of this contradiction, and tests whether a generative AI tool can do something about it.
A new name for an old inertia
Bin Shi of Zhejiang Gongshang University and Wei Chen of Zhejiang Tailong Commercial Bank call the pattern a “cognitive liquidity trap.” The phrase borrows from the Keynesian idea that people sometimes hoard cash even when the returns to doing something else are clearly positive. In the authors’ version, the cost that keeps people from acting isn’t a fee or an interest rate. It’s mental effort: the work of reading product disclosures, parsing jargon, weighing risks, and tapping through an app.
Their argument is that when that mental cost exceeds the perceived benefit, inaction becomes a reasonable choice, not an irrational one. This reframing matters because most existing explanations for elderly underinvestment point to risk aversion, low financial literacy, or cognitive decline. Those explanations treat the person as the problem. The authors argue that the decision environment, which they call the “decision architecture,” is at least as much to blame, and is something that can be redesigned.
Measuring mental effort from calls and clicks
To test the idea, the researchers partnered with Zhejiang Tailong Commercial Bank, a regional lender based in Taizhou. They assembled 18 months of data on 600 customers aged 60 and over, drawing from three sources: anonymized transaction records, transcripts of customer service calls, and behavioral logs from the bank’s mobile app.
The centerpiece of the measurement work was an attempt to turn “cognitive load” into something quantifiable. The team constructed a composite index from three kinds of signals:
- Language cues in phone calls. A speech recognition system transcribed about 120 hours of customer service calls. The researchers then counted pauses longer than a second, filler words like “emmm,” repeated phrases, and hedging language such as “maybe” or “let me see.”
- App navigation patterns. From mobile app logs, they tracked how many clicks a user took to reach the wealth management page compared to the shortest possible path, how long they lingered on product pages, and how often they abandoned a session mid-transaction.
- Transaction timing. They measured how much time passed between first viewing a product and actually buying it, and how many calls to customer service occurred along the way.
These signals were combined statistically into a single score representing how much cognitive strain each customer seemed to be experiencing when dealing with the bank. In the baseline data, 81% of these customers’ assets sat in current accounts, and only 18% owned any wealth management product at all.
The experiment
Beginning in October 2024, the 600 customers were randomly split into two groups of 300 each, balanced on age and digital literacy. The control group received the bank’s standard services: generic product pitches by SMS and rule-based fraud alerts.
The treatment group received messages written by a fine-tuned large language model. When the system detected a high cognitive load score and noticed the customer hadn’t been investing, it would trigger a tailored message. These messages did three things. They translated jargon into plain language (“This product is like an upgraded version of a time deposit”). They anchored decisions to concrete life goals (“If you save 2,000 RMB monthly, you could accumulate 300,000 RMB in 10 years, enough to cover basic elderly care expenses”). And they borrowed trust from recognizable sources, such as referencing the customer’s named client manager or suggesting the customer check with a family member before proceeding.
Messages went out by push notification or voice message depending on the customer’s comfort with technology. All were reviewed by the bank’s compliance team.
What changed
Over the post-intervention period, the treatment group’s average cognitive load score dropped by roughly a quarter of a standard deviation, while the control group’s barely moved. More importantly, that shift tracked with real financial behavior.
Fraud-related transactions, defined as large transfers to unfamiliar accounts, fell 38% in the treatment group compared to controls. Adoption of low-risk wealth management products rose 29%, moving the share of customers holding such products from roughly 18% to 23%.
The authors illustrate the fraud result with the case of a 74-year-old customer named Mr. Lin, who had begun transferring 20,000 RMB in response to a call claiming he’d won a government elderly care subsidy. An AI-generated message pointed out that no such subsidy program existed in official announcements and suggested he call his son. He canceled the transfer. In a follow-up survey, he said: “Previously, bank messages felt too formal. This time, it sounded like a reminder from my family—I became immediately alert.”
The effect was largest where friction was largest
One of the more informative patterns in the data concerns who benefited. The researchers split customers into four quartiles based on their baseline cognitive load scores. In the lowest-friction quartile, the intervention had small and statistically insignificant effects. In the highest-friction quartile, fraud-related transactions dropped substantially more, and low-risk product adoption rose by about 9 percentage points, more than four times the effect in the lowest quartile.
The authors argue this monotonic pattern is hard to reconcile with a pure risk-aversion story, under which the most hesitant customers should resist all financial engagement regardless of messaging. It fits more naturally with the idea that those facing the highest mental costs have the most to gain when those costs fall.
A separate analysis asked how much of the intervention’s effect ran through the cognitive load reduction itself, as opposed to other channels such as simple persuasion. The authors estimate that changes in the cognitive load score accounted for roughly 76% of the fraud effect and 72% of the adoption effect, which they interpret as evidence that the AI worked by making decisions easier to think through, not by pushing people harder.
What it might mean for banks and regulators
The authors frame their results as a case for treating AI as a “cognitive scaffold” rather than an automation engine. In their view, the bulk of AI applications in consumer finance today either replace human labor (robo-advisors, chatbots) or tighten fraud detection. Few are designed specifically to lower the mental effort a customer has to spend to make a reasonable choice.
For banks serving aging populations, the practical suggestion is that product design and education are not the only levers. How information is delivered, whether it connects to concrete life goals, and whether it invokes trusted figures appear to matter as well. The authors also suggest that regulators could eventually incorporate cognitive-load measures into consumer protection standards, treating cognitive accessibility as a feature alongside safety and transparency.
Caveats
Several limits are worth keeping in mind. The study covers customers of a single regional bank in eastern China, so the specific mix of familial trust cues and messaging styles that worked here may not translate directly to other cultures or financial systems. The intervention lasted only a few months, leaving open the question of whether behavior persists once the AI nudges stop. And the cognitive load index, while built from observable signals, doesn’t capture how users themselves describe their experience of deciding.
There is also an ethical edge the authors acknowledge. A system sensitive enough to detect when an elderly customer is confused, and tailored enough to nudge them effectively, is a system that could also be misused. The study’s intervention was reviewed by compliance staff and framed as informational rather than directive, but the authors note that any broader deployment would need continuing oversight, particularly for differential effects across demographic groups.
