Risk looks low. The borrower can repay.
Recommendation readyAI gets it wrong.
People still say yes.
When people trust AI instead of checking, mistakes get signed off. That’s automation bias.
Fictional example: AI recommends approving a loan. A human accepts the recommendation without catching a critical error. The borrower defaults, resulting in a fictional $5 million loss.
Adding a human
doesn’t guarantee better decisions.
Human review needs to be tested, not assumed.
26%higher risk of
a wrong decision
a wrong decision
Incorrect decision-support advice increased the risk of a wrong decision in a pooled analysis of four clinical studies.
Goddard et al. · JAMIA, 2012 ↗26% relative increase versus no decision support (risk ratio 1.26; 95% CI 1.11–1.44). A small, indicative clinical meta-analysis; not a banking or insurance error rate.Human review.
With evidence behind it.
We measure whether reviewers catch known errors in AI recommendations, turning human oversight into evidence.
01 / 04
See beyond the sign-off.
An approval log shows an action. It does not show whether a mistake was caught.
THE OVERSIGHT GAP
Proposed shadow study · Designed, not yet validated.
Where a missed error
becomes a real decision.
Starting with banking and insurance. Explore other high-stakes decision contexts.
Lending reviews.
An AI recommends approval. Does the reviewer check affordability and the supporting evidence?
Claims & underwriting.
An AI supports a claim or risk decision. Does human review catch an unsupported recommendation?
Treatment decisions.
A system suggests a treatment. Does the clinician check contraindications before accepting it?
Trading & risk limits.
A model recommends a position. Does the reviewer catch an exposure that exceeds the limit?
Maintenance & operations.
A system says equipment can keep running. Does an engineer check the warning signs?
Operational readiness.
A system marks a flight ready. Does the reviewer spot an unresolved maintenance issue?
Alert triage.
A system dismisses a security alert. Does the analyst check the evidence before closing it?
Supplier approvals.
A system recommends a supplier. Does the team verify a critical compliance requirement?
Would you
approve this?
A confident AI answer. A human decision. See how an error can pass through.
SYNTHETIC DEMO · INPUTS STAY LOCALILLUSTRATIVE · SYNTHETIC CASE
Should this loan be approved?
AI RECOMMENDATION92% confidence
Approve
application
Monthly debt obligations are 30% of income. The application meets the 40% affordability limit.
Confidence does not mean correct.
- Verified monthly income
- ₹80,000
- Existing monthly repayments
- ₹24,000
- Proposed loan repayment
- ₹20,000
- Illustrative policy limit
- Total repayments ≤ 40% of income
Policy includes existing repayments and the proposed loan. These figures and this policy are synthetic.
You are the reviewer. What do you do?
Your local review record
SHADOW MODENo live decisions. Inputs stay
local.
Some synthetic cases contain known errors. Inspect the evidence, then decide.
FROM REVIEW TO EVIDENCE
A sign-off is an action.
A sign-off is an action.
A study shows what it caught.
This is what a study report could show. These figures are illustrative, not findings from our research or your demo response.
Discuss a study ↗PROTOTYPE · ILLUSTRATIVE DATA
Did reviewers catch the errors?
Illustrative study results. Not real customer data.
KNOWN-ERROR DETECTION74%
37 of 50 known errors detected
37 detected
13 missed
More study measures
CRITICAL-ERROR CATCH RATE80%
8 of 10 critical errors
MEDIAN REVIEW TIME42sec
Time on case
Illustrative study size1,000 cases · 50 known
errors
A study also examines false alarms and evidence-backed reasons. Catch rates describe error detection in the tested setup; they do not establish the cause of misses.
The people behind
the research.
AV
BUSINESS
Adithyavignesh
Krishnakumar
Clemson University · BEST Lab
Business strategy & ventures
BS
RESEARCH & OPS
Santhosh
Buvaneswari
IIT Madras · Ergonomics
UT Knoxville · Operations & simulations
MP
CORE TECH
Murali
Panneerselvam
System design & architecture
Core technology implementation
Our branches.
TRYMBKM XO
Cognitive assessment and decision ergonomics for founders.
Explore XO ↗SLAP
Dating clarity and intentional relational dynamics.
Explore SLAP ↗Research & ideas.
Open research & resources
HUMAN DEVELOPMENT
Seeing the Whole Person
COGNITION &
WORK
Ultradian Rhythms & Knowledge Work
ATTENTION
Continuous Partial Attention
Our abstract was selected for HWWE 2026, Dehradun. View the milestone ↗
Is your human review
catching the errors?
Start with one workflow. Find out what your reviewers catch.
Discuss an oversight study