Skip to main content
crumpled paper in a trash can

Economics & Finance

Why Leaders Should Keep the Ideas AI Rejects

Why Leaders Should Keep the Ideas AI Rejects

An agentic AI currency-trading desk keeps its rejected proposals and dissenting votes. The case shows how that habit can turn fleeting AI interactions into an auditable record that organisations can learn from.
Loading the Elevenlabs Text to Speech AudioNative Player...

When people talk about using AI as an “agent”, they tend to focus on the result – what the AI has produced. But in many high-stakes decisions, the things that the AI didn’t do are as important as the things that it did. 

An experimental AI currency-trading desk* makes the point. It deliberately keeps unsuccessful proposals next to the trades it executed. Financial markets are noisy and, in global macro trading, the sample of decisions is far too small for short-run performance to show whether an AI decision process is “working”. The desk is therefore designed to make its process observable rather than to optimise for returns. Its main feature is a record of the alternatives considered, arguments made and decisions rejected.  

This is useful in any domain where it’s hard to tell whether a good decision led to a bad outcome or a bad decision happened to work. Keeping the alternatives gives evidence that outcomes alone cannot provide. 

Preserve the alternatives 

Most organisations don’t do this – perhaps simply because management hadn’t thought to. Yet, organisations can’t judge the quality of a decision properly if they only record what was chosen. A conventional trade ledger shows the position a desk took. It says nothing about the other options on the table, why they lost out or whether important objections were ignored. Once those alternatives are gone, later reviews are exposed to hindsight bias, with people working out what they “would have known” from an outcome they can now see. 

The trading desk, which covers five currency pairs (EUR/USD, GBP/USD, AUD/USD, USD/JPY and USD/CAD), keeps the options on file before anyone knows how they turn out. Three AI models each nominate a trade independently. The models then act as a committee, debating the competing proposals and voting on which, if any, should go ahead. Every proposal must state a directional claim, a catalyst with an explicit timeframe and a measurable condition that would invalidate the trade. The system stores the proposals together with the debate that followed and every dissenting vote. 

Suppose one model wants to buy the Japanese yen because it expects Japanese interest rates to rise, and the committee picks a Canadian dollar trade instead. The yen proposal stays in the record with its catalyst, timing and invalidation condition. Months later, managers can check more than whether the chosen trade made money. They can ask whether the yen thesis was better conceived, whether its catalyst happened and how it would’ve done under the same risk rules. 

Make rejection scorable 

Keeping alternatives only helps if managers can assess them without rewriting history or falling for confirmation bias. An agentic system that holds rejected ideas to the same rules as accepted ones is well suited to this. A monthly review can replay the nominations that weren’t selected, using standardised assumptions for position size, stop losses and holding periods, and managers can then compare the committee’s decision with what the models first proposed. Over time, that comparison can help answer an unresolved question about multi-agent AI systems: Does debate improve decisions or can it sometimes suppress good ideas? 

Getting a straight answer takes discipline. Both the proposal and the rules for scoring it must be fixed before market outcomes are visible. If they aren’t, reviewers can, without realising it, improve an abandoned idea after the fact by moving its hypothetical entry point, holding period or exit rule. 

Managers also need to keep decision quality and outcome quality apart. Suppose a trade is rejected for breaching a risk limit and then rises sharply. The rejection wasn’t necessarily wrong, since the risk constraint may have done exactly the job it was designed for. Likewise, one rejected trade that later does well is no proof that the entire deliberation process is broken. The distinction matters wherever luck can dominate individual outcomes, as it can in financial markets, strategic investments, hiring and product launches. 

Some questions about the process can be answered straight away. Reviewers can check whether a thesis relied on verifiable information, whether it named a specific catalyst and an explicit timeframe, and whether another reviewer could tell objectively if the invalidation condition had been met. Other questions need time. Did rejected ideas systematically beat selected ones? Did the dissenting models see risks earlier? Did deliberation improve calibration? The organisation can't answer these questions yet, but it's now collecting the evidence to do so. 

Keep the record grounded 

A useful decision record must capture the facts and rules that shaped the decision, alongside what the AI said. The currency-trading platform therefore splits quantitative calculation from qualitative judgment. Conventional Python code – which gives the same answer every time for the same inputs – calculates market indicators, applies position-sizing rules, watches risk limits and logs workflow events. The language models interpret the evidence, form hypotheses and challenge each other’s arguments.  

Any number in an AI-generated research narrative must match a field the code has already calculated, and the model receives those figures as read-only inputs. It can explain why an inflation surprise or interest-rate differential might matter, but it can’t make up the number itself. This reduces a familiar source of AI error: the confident narrative built on bad arithmetic or invented statistics. It also gives human reviewers something concrete to check. 

The system also keeps disagreement, because consensus doesn’t prove the majority right. A minority view may later turn out to have spotted a risk that others missed. Opening a new trade requires a unanimous vote, while decisions on existing positions need only a majority. When a position is closed on a two-to-one vote, the dissenting model’s objection stays attached to that decision. Without a record made at the time, nobody could later check whether the dissenter was right. For the same reason, the system logs the market inputs, proposals, votes, governing rules and human approvals that applied when each decision was made.

A record of past mistakes should feed into governance. It shouldn’t give an AI system permission to rewrite its own objectives.

The purpose of keeping rejected ideas is to give people better evidence for improving the system. People set the trading mandate, authorise capital and approve changes to the system. The AI can spot patterns and suggest changes to prompts, models or risk parameters, but it can’t put those changes into effect on its own. The separation matters, because a record of past mistakes should feed into governance. It shouldn’t give an AI system permission to rewrite its own objectives. 

Implications for organisational learning 

AI agents produce large volumes of proposals, objections, discarded drafts and dissenting judgments. Most organisations keep only the final output, potentially throwing away the information they need to tell whether their AI systems are actually improving decision-making. 

Managers could instead treat rejected alternatives as structured evidence. Such a record would show which objections proved correct; which models produced useful ideas that the consensus routinely rejected; and which constraints protected the organisation from bad risks or blocked attractive opportunities. A pile of successful outputs can’t answer any of that. 

Under uncertainty, a good outcome doesn’t prove the organisation decided well, and a bad outcome doesn’t prove it decided badly. To learn, it must reconstruct what was knowable at the time, which alternatives were considered and where disagreement lay. Preserving the record turns fleeting AI interactions into institutional memory, giving leaders a stronger basis for deciding what their organisations should learn. 

*This article draws on Joshua Dawe’s INSEAD MBA Independent Study Project, supervised by Ben Charoenwong. The underlying trading logs have not been independently reviewed. Examples are illustrative.

Edited by:

Rachel Eva Lim

About the author(s)

Related Tags

Artificial Intelligence

About the series

AI: Disruption and Adaptation
Summary
Delve deeper into how artificial intelligence is disrupting and enhancing sectors – including business consulting, education and the media – and learn more about the associated regulatory and ethical issues.
View Comments
No comments yet.
Leave a Comment
Please log in or sign up to comment.