Executive Summary
Every investment professional knows the risk of being wrong. Fewer think systematically about the risk of not knowing that you are uncertain. The difference between a wrong view and an uninformed view is not always obvious at the time — but it is almost always obvious in retrospect.
Current AI research tools have a structural confidence problem. They are trained and evaluated on the quality of the answers they produce. There is no reward signal for saying ‘insufficient evidence.’ In a consumer context, this is a minor annoyance. In an investment context, it is a systematic risk.
This paper makes the case that explicit acknowledgment of insufficient evidence is not a limitation to be engineered away. It is one of the most valuable features a research tool can have.
1. The Confidence Problem in AI Research Tools
Language models are, by design, optimised for fluency. They generate text that sounds confident, coherent, and knowledgeable — because that is what the training signal rewards. When a model does not have adequate evidence for a claim, it does not typically flag the gap. It fills it.
This filling behaviour is not a bug that will be patched out in the next model release. It is a consequence of how language models are built. They are not reasoning about what they know and what they do not. They are predicting what a plausible answer looks like.
In investment research, this creates a specific and dangerous failure mode. An analyst asks a question. The AI tool produces a fluent, well-structured response. The analyst cannot easily tell whether that response is grounded in authoritative licensed sources or synthesised from general training data and plausible inference.
A research tool that does not know what it does not know will not tell you when to be uncertain. That is not a limitation you can work around. It is a structural flaw in the research process.
2. What ‘Insufficient Evidence’ Actually Means in Investment Research
In an evidence-bounded research system, ‘insufficient evidence’ is a specific, actionable output. It means one or more of the following:
Coverage Gap
The licensed evidence universe contains little or no material on the specific question being asked. This is a signal to the analyst to either expand the evidence universe, seek additional licensed sources, or treat the question as one where the research base is thin.
Consensus Fragility
Multiple licensed sources address the question, but they disagree significantly. The system cannot report a consensus because there is not one. This is a materially different situation from strong consensus with a minority dissent — and it requires a different analytical response.
Evidence Staleness
The licensed sources that address the question are dated in ways that may affect their relevance. A sector view from eight months ago may be directionally useful or may be materially outdated depending on what has happened since. An evidence-bounded system flags the vintage of its sources.
Scope Mismatch
The question the analyst is asking falls partially or fully outside the defined evidence universe. This is common in cross-asset or multi-geography queries where licensed coverage is uneven. An evidence-bounded system reports the scope of coverage it has and what falls outside it.
3. The Cost of Not Knowing What You Don’t Know
False Conviction
A PM forms a view on the basis of research that appeared well-supported but was, in fact, grounded in thin or unverified evidence. The position is sized on the basis of that conviction. When the view is challenged or the position goes against, the analytical foundation is not there to defend it.
Undetected Consensus Risk
A team runs the same research tools as their competitors and, because those tools synthesise in similar ways from similar data, arrives at similar conclusions. Evidence fragility — the degree to which a consensus view is supported by thin or monolithic sourcing — is invisible to tools that do not surface it.
Investment Committee Exposure
A research output is presented to an Investment Committee or a risk committee. A committee member asks where a specific claim comes from. The analyst cannot answer — because the AI tool that produced the claim did not preserve the provenance. The question then is not just about the specific claim. It is about the reliability of every claim in the output.
Regulatory and Audit Risk
As AI-generated research outputs become more common and regulatory scrutiny of AI in financial services increases, the absence of an evidence trail becomes a governance risk. Regulators and auditors are not just interested in whether a decision was right. They are interested in how the decision was made.
4. Surfacing the Gap: What a Sufficiency-Aware System Does Differently
- It reports coverage — how many licensed sources address the query, from which providers, over what timeframe — before synthesising across them
- It distinguishes between supported conclusions and inferences — clearly marking where a claim is directly evidenced in a licensed source
- It surfaces disagreement — rather than averaging across different analyst views, it reports the distribution of views and flags material divergence
- It refuses to conclude when evidence is insufficient — rather than generating a plausible response, it reports that the evidence does not support a conclusion
- It preserves provenance — every claim carries a traceable link to the source passage
The research tool that says ‘insufficient evidence’ is not failing. It is doing the most important thing a research tool can do: preserving the integrity of the analyst’s judgement.
5. Why Buy-Side Professionals Should Demand This
The buy-side is in a buyer’s market for AI research tools. Every major terminal, every research aggregator, and every technology vendor is now offering some form of AI-assisted research capability. Very few are differentiating on evidence quality and sufficiency signalling.
This is a market mispricing. The feature that most investment professionals need most — an honest signal about the quality and completeness of the evidence behind an answer — is the feature that most tools are least equipped to provide.
Conclusion
The research blind spot — the confident answer generated in the absence of adequate evidence — is not a minor inconvenience. It is a systematic risk in investment processes that have not yet established the evidence governance layer they need.
Contours is built around explicit sufficiency thresholds. When the licensed evidence does not support a conclusion, Contours says so. When coverage is thin, fragmented, or stale, Contours surfaces that before synthesising across it. This is not a limitation. It is the feature that makes every other output trustworthy.
Contours is a product of KiteEdge Ltd. To learn how Contours handles evidence sufficiency and coverage gaps, contact us
