40 Questions Finance Should Ask Before Trusting AI
Your regulators require you to explain every material decision, yet your AI models cannot explain why they chose what they chose. Your fiduciary duty demands independent judgement, yet your team is trained on the same Bloomberg AI outputs as your competitors. These 40 questions help you keep human judgement in control when AI pressure is high.
These are suggestions. Use the ones that fit your situation.
1How many other institutions in our sector use the same Aladdin model configuration that we do for portfolio construction?
2What happens to our risk models if the underlying training data becomes unrepresentative, as it did during the 2008 crisis or COVID shock?
3Can we identify which specific Bloomberg AI recommendations our competitors acted on in the last quarter?
4If Palantir's data integration fails silently, how many decisions downstream would be made on corrupted inputs before we notice?
5What sector-wide positions would unwind simultaneously if all AI models recognised the same risk signal at the same time?
6Are we monitoring whether our credit scoring AI and our competitors' credit scoring AIs are rejecting the same loan applicants for the same reasons?
7If our Microsoft Copilot investment analysis tool goes offline, what percentage of our analysts cannot complete their weekly reports without it?
8How would our risk committee identify a failure mode that affects all instances of a widely adopted model simultaneously?
9Do we have an analyst or team member whose job is explicitly to produce recommendations that contradict the AI consensus?
10What manual processes would need to restart immediately if all our AI tools failed at once?
Regulatory Compliance and Explainability
11Can we write a paragraph explaining why ChatGPT recommended this specific investment to a regulator without using the phrase 'the model learned patterns'?
12Does our AI output include the confidence level required by our regulator, or only the recommendation?
13How do we prove that a Palantir flagged transaction was not rejected based on protected characteristics when the model itself cannot articulate why it flagged it?
14If a regulator demands we show our work for a portfolio decision, can we produce an audit trail that does not simply say 'the AI said so'?
15What is our documented process for overriding an AI recommendation, and how often do we actually do this compared to how often we should?
16Does our risk appetite statement explicitly address which decisions AI can make alone and which require human sign-off?
17How do we comply with MiFID II suitability rules when the AI recommendation came from a model trained on data outside the EU?
18Can our compliance team explain to auditors why we chose this particular AI tool over another, beyond cost or convenience?
19If we relied on an AI output that turned out to be wrong, does our insurance cover it, or does the use of AI void our professional indemnity?
20What internal governance approval was required before deploying this AI tool, and who remains accountable if it fails?
Risk Management and Model Failure
21Does our market risk framework have limits built in for AI recommendations that diverge sharply from human analyst consensus?
22What happens to our Value at Risk calculation if the AI model that feeds into it starts producing outlier predictions?
23Have we stress tested our portfolio against scenarios where all AI-driven sell signals trigger simultaneously?
24If Bloomberg AI recommends overweight in a sector we have already flagged as concentrated, how does that conflict get resolved?
25When was the last time we tested whether an AI tool still works correctly with data it was never trained on?
26Can we identify specific decisions we made based on ChatGPT analysis that we would reverse if we knew the model had been fine-tuned on publicly available market commentary?
27What is our procedure if a Palantir data pipeline produces silently corrupted outputs, and how quickly would we catch it?
28How do we detect model drift in our credit risk AI, and how frequently do we run that detection?
29Which of our key risk metrics rely on AI outputs, and what manual checks validate those outputs?
30If our insurance pricing AI was trained on historical claims from a period of abnormally low losses, how do we adjust for that bias now?
Analyst Judgement and Decision Quality
31How many of our portfolio decisions this quarter were made because Aladdin recommended them, versus because our analysts independently reached the same conclusion?
32Does our investment process include a mandatory step where an analyst must explain why they disagree with the AI recommendation before approving it?
33When was the last time an analyst on our desk spotted a risk that the AI models missed, and what was it?
34Are we still hiring analysts who think differently from the consensus, or have we optimised for candidates who work well with AI tools?
35If a junior analyst questions an AI recommendation, what is our process for escalating that concern above the algorithm?
36How do we ensure that our credit committee is not simply ratifying AI decisions rather than genuinely evaluating the loan application?
37What percentage of our analysts have been trained on how to identify when an AI output is plausible but wrong?
38Do we track whether our Copilot-assisted research reports reach different conclusions than pre-AI reports on the same topic?
39How many of our traders still maintain independent market views, or have they converged toward the consensus generated by shared AI tools?
40When we hire a new analyst, do we first teach them how to think independently, or how to use the AI tools?
How to use these questions
Assign one person on your risk team to explicitly contradict the AI consensus each month. Their job is not to be right, but to keep your organisation's ability to think independently alive.
Before deploying any AI tool firm-wide, test what happens when it fails. Can your settlement team process trades without Bloomberg AI? Can your underwriters approve policies without the AI scorecard? If the answer is no, you are not using AI correctly.
Document the last three times a human analyst at your firm caught a risk that the models missed. If you cannot remember any, ask yourself whether your best people have left because their judgement was no longer valued.
Request your AI vendor provide the training data dates and geographic composition of the model. If it was trained on pre-2020 data or data from only developed markets, treat its recommendations with scepticism during crises.
Run a quarterly test: take a decision your AI made, remove the AI explanation, and see if your team would have made that same decision using old methods. If the answer is always yes, the AI is probably just automating human consensus rather than adding insight.