What I Want From AI Decision Support Not Magic

When people hear “AI” near investing, they often reach for magic: perfect foresight, emotionless genius, automatic wealth. I want something less cinematic and more useful—decision support that helps humans process messy information, test assumptions, and stay inside risk envelopes. Magic marketing sets teams up for either blind trust or bitter disappointment.

Complementary labour, not replacement theatre

Manual analysis still matters for thesis formation, ethics, and institutional context. Machines excel at scanning large feature sets, monitoring streams, and enforcing rules without fatigue. The premise of serious AI quant work is complementary labour. Replacement theatre makes for better demos and worse governance.

In 2023 I organised research around AI-driven quantitative analysis and automated trading ideas for that reason. The bottleneck was not only idea scarcity. It was the cost of processing multi-source information without drowning in noise—while keeping humans responsible for capital.

Capabilities I actually budget for

I want better data hygiene, clearer regime diagnostics, wider scenario coverage, and sizing recommendations that respect hard constraints. I want alerts when correlations shift or when a strategy drifts outside its tested envelope. I do not want a black box that cannot explain why it likes a trade in language a risk committee can challenge.

Emotion reduction is not emotion deletion

Systems can reduce impulsive overrides by forcing checklists and cool-down rules. They cannot delete organisational politics. If seniors treat model output as optional when inconvenient, juniors will learn the real culture quickly.

Risk envelopes define intelligence

Dynamic position sizing and automated execution deserve respect only inside explicit, testable limits. Outside limits, speed is not intelligence. It is amplified error. Design the envelopes first. Then automate within them.

Interface ethics

How AI outputs are displayed shapes trust. Avoid fake precision. Show uncertainty bands. Show when data is stale. Show when a recommendation violates a risk envelope and is therefore blocked. Interfaces that only celebrate green buy signals are marketing, not decision support.

Human override should be logged with reasons. If overrides cluster, either the model is mis-specified or the culture is rejecting discipline. Both diagnoses are useful. Neither should be invisible.

Roadmaps with humility

Research programmes need milestones that include validation and compliance, not only feature demos. A projected commercial window is a plan, not a promise to markets or to students. Keep external communication aligned with internal evidence. Hype debt comes due.

Train users—especially in educational settings—to ask what the system cannot do. The most valuable AI literacy may be boundary literacy.

I want fewer magic claims and more boring reliability. That preference is not anti-AI. It is pro-capital and pro-trust.

Commercial timelines and technical roadmaps should stay humble. Plans are not proof. Validation and compliance are part of the product, not afterthoughts.

The ask

Evaluation metrics that match the job

Score decision-support systems on calibration, constraint adherence, operator time saved on drudgery, and incident rates—not on mythical hit ratios alone. A system that slightly improves discipline while reducing operational misses may beat a flashy signal with unstable live behaviour.

Publish internal scorecards. Make magic claims unwelcome because the scorecard language is better.

Training humans alongside models

Budget time to train operators on what the system optimises and where it fails. Untrained operators recreate magic thinking even if engineers were careful. Decision support is a socio-technical system. Neglect either half and the whole degrades.

In academy settings, demos should include a forced failure: bad data, regime break, blocked envelope. Students should leave impressed by discipline, not by omniscience cosplay.

Magic is a story. Support is a practice. Fund the practice.

Training humans alongside models

Budget time to train operators on what the system optimises and where it fails. Untrained operators recreate magic thinking even if engineers were careful. Decision support is a socio-technical system. Neglect either half and the whole degrades.

In academy settings, demos should include a forced failure: bad data, regime break, blocked envelope. Students should leave impressed by discipline, not by omniscience cosplay.

Magic is a story. Support is a practice. Fund the practice.

Give me decision support that makes discipline easier and self-deception harder. Keep the magic for fiction. Markets already supply enough surprises.