AI in Finance: When “Advice” Becomes Delegation

AI in Finance: When “Advice” Becomes Delegation
From information requests to delegated authority in money decisions

Introduction: The new question isn’t what people ask AI—it’s who they trust

If you’ve ever asked ChatGPT or Gemini, “Is this a good credit option?” or “How should I compare these investments?”, you’re part of a rapidly growing behavior shift: using AI not just for facts, but to support financial judgment. New research from the paper “From Information to Delegation: Mapping Human-AI Financial Decision Making” digs into exactly how that support shows up in real-world conversations—and where it doesn’t yet go.

The core idea is simple but surprisingly important: instead of only measuring the topics people talk about with AI (investing, payments, taxes, etc.), the researchers measure intent and decision authority. In other words, they ask: What are people trying to accomplish, and how much are they handing over to the AI to shape or execute the decision? That “delegation” lens is the heart of the study.

Using 1.5 million real interactions (across ChatGPT and Gemini) from 6,304 users in the United States and India, the authors find something practical: financial services are already a major AI use case—about half of users engaged in finance conversations during the observation window. But even more interesting, delegation of actual financial execution is rare. People overwhelmingly use AI to retrieve information and shape decisions, while leaving the final action to themselves.

Why This Matters: “Decision-support” is the risk zone policymakers missed

This research is significant right now because the AI conversation is already embedded in consumer financial behavior—but not in the way regulators and product designers often assume. Many debates focus on whether AI systems can autonomously execute decisions (the “agent” storyline). This paper suggests we should also worry about the middle ground: AI that influences judgment at scale.

A real scenario you can map to the study

Imagine someone repeatedly using AI to troubleshoot a payment problem, compare credit products, or decide between investment options. Even if the person never lets AI place orders or transfer money automatically, the AI’s role can still strongly shape outcomes—like steering what they consider “reasonable,” which trade-offs they prioritize, or which risks they underestimate. According to the paper’s framework, that’s decision authority Level 2 (“Shape”), and it’s the dominant pattern.

How this builds on previous AI research

Prior work often analyzes AI usage by topic (what people ask) or assesses the quality of AI-generated advice in isolation. This study pushes beyond that by treating conversations as behavioral episodes and measuring delegation explicitly. If earlier research helped us understand what questions people ask, this work helps us understand how responsibility is distributed. That’s a meaningful upgrade in both product evaluation and consumer protection.

What the researchers actually measured: Intent + delegated decision authority

The paper’s contribution is a behavioral measurement framework that combines two complementary dimensions:

1. Behavioral intent: what the user is trying to accomplish in the finance domain
2. Decision authority (DA): how much decision-making power the user assigns to the AI

Think of it like the difference between:
- asking AI to read you a map (information),
- asking AI to suggest the route and trade-offs (shaping),
- and asking AI to drive the car (acting/execution).

Decision authority levels: a simple “who’s in charge” ladder

The authors define three levels:

• Level 1: Inform — AI provides information/explanations
• Level 2: Shape — AI influences choices/strategy (advice, comparison, optimization)
• Level 3: Act — AI executes/autonomously acts for the user

This structure lets them move from “AI is used for finance” to “AI is used as a decision-support tool vs. a decision executor.”

The finance conversation dataset: real chats, cleaned into sub-decisions

The study analyzes chat histories from August–October 2025 from:
- 2,499 US users
- 3,805 India users
recruited via an opt-in process (voluntary participation) by a third-party data provider.

They collected:
- 1.5 million real-world interactions total
- 291.9k “subchats” for the US
- 213.4k “subchats” for India

Why “subchats”? Because users often keep discussing multiple unrelated topics in one session. The researchers split long sessions into smaller units so a “finance decision episode” is easier to classify. They also:
- translate non-English India conversations into English (~144k non-English messages) using Google Translate
- use a fine-tuned classifier to detect whether a subchat is finance-related with Accuracy = 96.5 and F1 = 97.3
- then tag financial product/service categories using a MECE taxonomy (Mutually Exclusive, Collectively Exhaustive)

Intent and authority classification is designed to handle imbalance

One challenge in this kind of dataset is that rare “high-authority” behaviors are, well, rare. The authors report that at the time of data collection, conversations involving delegated execution were under 1% in volume.

To cope, they:
- fine-tune a long-context model for intent labeling
- add 600 synthetic examples generated with GPT 5.5 (with higher reasoning effort) to increase representation of higher-agentic intents

Where conversational AI is used in finance today: The biggest domains and regional differences

The first major result is that finance isn’t niche anymore—it’s already mainstream.

Finance conversations are widespread in both countries

During the 3-month period:
- US: 50.0% of users engaged in at least one finance-related conversation
- India: 44.3% of users did the same

On a conversation-share basis:
- US: finance accounted for 5.1% of subchats
- India: finance accounted for 6.3% of subchats

So yes, people talk about money with AI a lot. And the paper shows it’s not marginal—more like a routine companion for everyday financial thinking.

Different countries use AI for different financial slices

Across both markets, the largest share of finance conversations falls into categories like:
- Retail Banking & Credit
- Payments & Transfers
- Investments

But the mix differs by country. The paper reports:

• US: Retail Banking & Credit dominates, especially Credit Instruments & Monitoring and Deposit Products
• India: Investments are a larger focus, especially Liquid Securities and Alternative Assets
• Insurance: plays a substantially larger role in the US than India (the paper doesn’t claim causality, but points to US market maturity and product diversity)

Here’s the practical interpretation: even when the “AI finance” headline is the same, users’ lived financial context differs—and their AI queries mirror that.

How the conversation topics connect to real decisions

The authors go beyond listing categories by describing what those conversations look like in practice:
- In the US, Retail Banking & Credit discussions often cover credit score improvement, managing payments, and choosing products
- In India, Investment conversations lean toward portfolio management, investment strategies, and trading decisions
- Insurance conversations in the US include information navigation that likely reflects more diverse offerings

This matters because it reframes “AI finance usage” as an ongoing chain of micro-decisions—not just isolated Q&A.

What users want from AI, and how much they delegate: Inform and Shape dominate

Now the headline: even with “agentic AI” getting lots of hype, most users aren’t delegating financial execution.

Decision authority distribution (the money reality check)

Across both countries, the paper finds:

• Level 1 (Inform):
  • US: 63.5% of finance subchats
  • India: 72.1% of finance subchats
• Level 2 (Shape):
  • US: 58.6%
  • India: 49.7%
• Level 3 (Act):
  • US: 0.3%
  • India: 0.1%

A key nuance: this is intent-at-subchat level and the model can assign multiple intent labels in a conversation span. Still, the pattern is crystal clear: AI is mostly used to inform and influence, not to act.

What “Shape” looks like in actual finance workflows

The researchers list the major Shape-level intents as including:
- Financial Problem Resolution
- Product & Strategy Comparison
- Product & Strategy Optimisation
- Personal Financial Analysis

So users aren’t just asking “What is X?” They’re asking:
- “How do these options compare?”
- “What’s the best strategy given my situation?”
- “Can you help me analyze this decision?”

“Act” is basically absent—and mostly around tracking

What about execution? Level 3 (“Act”) is rare, and the paper notes it’s almost all:
- instruction-led budgeting and tracking tasks

They report close to no evidence of users delegating autonomous financial decisions to AI. That’s a meaningful behavioral baseline for future research as agentic systems improve.

Decision authority varies by finance domain

This isn’t one-size-fits-all. The authors find different behavioral profiles by domain, for example:
- Optimization is concentrated in Investments
- Problem resolution is centered in Retail Banking & Payments

So if you’re building or regulating AI tools, the risk profile isn’t only “Is it in finance?” It’s “Which kind of finance?” and “What role is the AI playing?”

The concrete activities behind the labels: Topics that show up in both countries

To translate intent+authority into something tangible, the authors also run topic modeling on the finance subchats (using BERTopic). They identify 170 topics overall.

They show the top topics (10) and find overlap plus differences. The shared patterns include:
- investment guidance
- payment management
- banking services
- taxation
- credit improvement

US vs India topic emphasis

Differences show up in what people lean on AI for:
- US conversations more often involve insurance, education finance, and public assistance
- India conversations more heavily feature investment research, stock analysis, and account security

If you’re thinking in product terms, this implies AI financial assistants may succeed differently across markets—not just due to language, but due to what decisions people are trying to solve.

What this means for products, users, and regulators

One of the most policy-relevant takeaways is the “middle space” problem. The paper argues that because Level 2 (“Shape”) interactions dominate, consumer protection frameworks may not fully address the ways AI can influence financial judgment without executing decisions directly.

For regulators: don’t only regulate “autonomous action”

Existing approaches often focus on whether AI systems can carry out trades, transfers, or other irreversible actions. But this paper suggests the earlier, subtler risk is widespread: AI shaping decisions can still affect outcomes dramatically.

For financial institutions and product builders: the near-term value is decision-support

The authors’ findings align with a practical product direction: near-term consumer demand seems concentrated on:
- product optimization
- financial problem resolution
- personalized analysis

rather than fully autonomous agents.

So the winning approach may be to build AI that:
- supports analysis and comparison transparently
- helps users check assumptions
- makes it easier to understand why a recommendation is being made

For researchers: the framework provides a baseline as AI becomes more “agentic”

Because the study measures intent and decision authority, it creates a baseline for future work. If agents become more capable, you should be able to observe shifts from:
- Level 1 → Level 2 → Level 3
over time—using the same behavioral lens.

This is exactly why the paper’s contribution matters beyond this snapshot: it gives a measuring stick for a transition that’s already underway.

Key Take Takeaways

• Finance is a major AI use case already: about 50.0% of US users and 44.3% of India users engaged in finance conversations during the study window.
• People mostly use AI to inform and shape decisions, not execute them:
  • Level 1 (“Inform”) dominates (63.5% US, 72.1% India)
  • Level 2 (“Shape”) is also large (58.6% US, 49.7% India)
  • Level 3 (“Act”) is extremely rare (0.3% US, 0.1% India)
• “Decision delegation” isn’t happening at scale yet—but AI influence is. That means consumer outcomes can change even without autonomous actions.
• What AI is “used for” differs by domain and country:
  • US leans toward Retail Banking & Credit, especially monitoring and deposits
  • India leans more toward Investments
  • US has a bigger share of Insurance conversations
• Practical implication for product teams: build systems that act as decision-support tools (comparison, optimization, problem resolution), and focus on clarity and user control rather than full automation.
• Practical implication for policymakers: regulate not only autonomous execution, but also the high-volume shaping behavior that can silently steer choices.
• Future research baseline: the paper’s intent+authority framework gives a consistent way to track how AI roles shift toward greater autonomy over time (and the original paper lays out that measurement approach).

If you want, I can also turn this into a “what should you do as a user?” guide—like a checklist for when to trust AI advice, when to double-check with primary sources, and what questions to ask before you act.