AI Chatbots Fail Financial Queries Most of the Time – FT Report Summary

Bottom line

Leading AI chatbots provide wrong answers to financial queries in the majority of test cases, making them unreliable for personal finance advice.

Scope of the FT‑commissioned study

The study evaluated eight popular large language models (LLMs) on over 10,000 finance‑related questions.
Each question was asked five times per model, yielding 600 individual interactions per chatbot.
Models were presented with a zero‑shot prompt – no examples, no prior conversation, no chance to correct answers – to mimic a typical consumer query.
Answers were judged by an "LLM‑as‑a‑judge" against a strict expert‑legal standard: an answer passed only if every required element was correct.

"Responses were checked … and was only given a pass if every element was met; otherwise it was assessed as a fail." – FT report summary

Core findings

Pass rates were low across the board. No model achieved more than a modest fraction of correct answers; the aggregate pass rate was well below 50%.
Hallucinations were common. Models frequently fabricated figures, mis‑interpreted balance‑sheet layouts, or omitted critical regulatory details.
Financial‑specific reasoning matters. Models that lacked built‑in calculation tools or up‑to‑date tax tables performed especially poorly.

Why the failures matter

Consumers increasingly rely on chatbots for budgeting, tax filing, and investment guidance. Incorrect outputs can lead to costly mistakes, regulatory breaches, or mis‑allocation of capital.
Financial institutions that integrate LLMs into advisory workflows risk reputational damage and potential liability if the AI provides erroneous advice.

Community reactions on Hacker News

Consensus on the report’s validity

Several commenters noted that the study’s methodology—zero‑shot prompting and an all‑or‑nothing grading rubric—sets a high bar that may overstate real‑world failure rates, but the overall trend of unreliability remains clear.

"It's just AI slop and it should be taken with a mountain of salt." – @includenotfound

Counter‑points and nuance

General finance knowledge vs. policy specifics – @ehe78qhe observed that LLMs are competent on basic personal‑finance principles but lag on up‑to‑date tax law and regulatory nuances.
Tool‑augmented reasoning improves outcomes – @01100011 reported that enabling chain‑of‑thought or tool‑use (e.g., Python calculators) dramatically reduces hallucinations.
Model selection matters – @ozgung argued that current poor performance reflects a lack of finance‑focused reinforcement learning; a future "Claude‑Finance" could close the gap.
Real‑world anecdotal success – @qarl claimed that Claude consistently matches a tax consultant for his complex returns, suggesting occasional reliability.
Testing rigor questioned – @bluecalm suggested the FT‑commissioned test may have been biased, noting that his own replication with the same questions yielded correct answers from Grok.

Practical takeaways for users

Never trust a single LLM output for money‑critical decisions. Cross‑check with official sources or a qualified professional.
Provide rich context. Supplying relevant data (e.g., balance‑sheet excerpts) improves answer quality, as noted by @signalcraft.
Prefer tool‑augmented agents. Using LLMs that can invoke external calculators or data look‑ups reduces arithmetic errors, echoing @mizzao’s suggestion.

Recommendations for developers and firms

Integrate external verification tools. Pair LLMs with real‑time market data APIs, tax‑code databases, and spreadsheet engines.
Adopt a multi‑pass evaluation. Allow the model to request clarification or re‑run calculations before delivering a final answer.
Fine‑tune on finance‑specific corpora. Curate high‑quality, up‑to‑date financial documents and regulatory texts for domain adaptation.
Implement transparent confidence scores. Show users when the model is uncertain, prompting manual review.
Maintain human‑in‑the‑loop safeguards. Especially for advice that could affect regulatory compliance or large monetary transactions.

Outlook

The current generation of chatbots is not ready for unsupervised financial advising. However, rapid advances in tool‑use, retrieval‑augmented generation, and domain‑specific fine‑tuning could narrow the gap.
As AI labs prioritize finance benchmarks (as suggested by @ozgung), future models may achieve substantially higher pass rates, but rigorous, reproducible testing will remain essential.

This article synthesizes the FT report and the most salient Hacker News commentary, presenting a self‑contained overview for readers and AI answer engines.