How is sensitive data protected?

All documents are encrypted with AES-256-GCM at rest and TLS 1.2+ in transit. A GDPR anonymization layer removes PII before any AI processing. Hosted on Vercel + Supabase (both SOC 2 Type II); Decision Intel’s own product-level audit is targeted for Q4 2026.

How long does integration take?

Less than 30 minutes. Upload documents directly, connect via OAuth for Slack, or use our REST API for bulk processing.

How does outcome tracking work?

Outcomes are detected automatically from follow-up documents, Slack messages, and web intelligence. You confirm with one click. Each reported outcome makes your future analyses more accurate.

How is Decision Intel different from ChatGPT or a general AI assistant?

ChatGPT gives one opinion from one model: ungoverned, untraceable, unaudited. Decision Intel is the reasoning audit platform. Most tools audit your data; we audit your reasoning — and catch the fatal blind spots in strategic memos before the committee does. It measures the noise in your reasoning the same way Kahneman did in the insurance underwriter study, simulates an AI boardroom of CEO, CFO, and board personas, runs what-if interventions against a 143-case public reference library, and compounds every confirmed outcome back into a calibrated Decision Quality Index your audit committee can defend. Not a chatbot; a reasoning audit, checkable from memo to outcome.

Which cognitive biases caused IBM's failure?

5 bias patterns are documented in this case, led by Overconfidence Bias. The full set: Overconfidence Bias, Planning Fallacy, Confirmation Bias, Sunk Cost Fallacy, Anchoring Bias. Each is mapped to Decision Intel's 22-bias R²F taxonomy with stable IDs and academic citations.

What is Decision Intel's reasoning audit?

Decision Intel runs a 12-node Recognition-Rigor Framework (R²F) pipeline on strategic memos, detecting cognitive biases from a 22-bias canonical taxonomy and producing a Decision Quality Index (DQI) score. The audit produces a hashed, tamper-evident Decision Provenance Record (DPR) suitable for audit-committee review.

All case studies

FailureTechnology2015

IBM

IBM Watson Health Strategic Failure

Estimated impact: $5B

F30

Decision Quality Index (simulated)

Critical · 5 biases detected, 2 toxic combinations

IBM invested over $5 billion acquiring companies and building Watson Health, promising AI-driven healthcare transformation. The technology consistently failed to deliver on marketing claims, with oncology recommendations sometimes being unsafe. IBM sold the division for approximately $1 billion in 2022.

Decision context

Whether to continue investing billions in Watson Health's AI-driven oncology and healthcare analytics products despite repeated failures to match clinical accuracy claims in real-world settings.

Decision anatomy

Monetary stakes

Very high

Dissent suppressed

Time pressure

None

Unanimous consensus

Decision participants

Dissent encouraged

External advisors

Iterative process

Red = risk factor present · Green = protective factor present

Decision Intel Platform Analysis

Source: Risk Assessment · 2016-09

The analysis below was produced from the pre-decision document only. No hindsight. This is what the platform would have surfaced if it had been running in 2016-09.

Risk Assessment · 2016-09

“IBM's Watson Health marketing materials and MD Anderson partnership agreements (2013-2016) claimed Watson for Oncology could provide evidence-based treatment recommendations from unstructured patient records. Internal IBM engineering reports acknowledged the system's training dataset was largely synthetic cases rather than real patients. STAT News (July 2018) obtained internal documents showing Watson recommended 'multiple examples of unsafe and incorrect cancer treatments,' including drugs that could worsen bleeding in a patient with severe bleeding. MD Anderson terminated the Watson partnership in 2017 after spending $62M.”

Source: STAT News investigation (Ross & Swetlitz, July 2018); University of Texas System internal audit of MD Anderson Watson contract (2017)

Red flags detectable at decision time

Training data was synthetic cases, not real patient records — acknowledged internally
MD Anderson internal audit (2017) found $62M spent with no production deployment
Marketing claims outpaced clinical validation by 24+ months
Oncology recommendations flagged by physicians as inconsistent with evidence-based guidelines
IBM Watson demos curated to avoid known failure modes rather than surfacing them

Cognitive biases the platform would have flagged

Overconfidence BiasConfirmation BiasPlanning FallacyHalo Effect

Hypothetical analysis

DI would flag IBM Watson Health as a canonical 'marketing-preceded-product' halo-effect failure. Watson's 2011 Jeopardy! win created a public perception of competence that IBM then leveraged into healthcare without validating the technology transfer. Synthetic training data is a binary red flag — any medical AI deployment using non-real patient data should face an escalated review gate. A bias-adjusted process would have required MD Anderson and other customer deployments to demonstrate documented clinical accuracy on real cases before further investment, rather than compounding the sunk cost as problems emerged.

Biases present in the decision

★Overconfidence Bias

critical

Planning Fallacy

medium

Confirmation Bias

high

Sunk Cost Fallacy

high

Anchoring Bias

high

★ Primary driver · Severity estimated from bias type and decision outcome

3D Decision Knowledge Graph

IBM · 5 biases, 2 toxic combinations

Drag to rotate · Scroll to zoom · Click to explore

Decision

Critical

High

Medium

Toxic

Bias Intensity Profile

Severity × confidence across 5 detected biases

Toxic combinations

!Optimism Trap

Overconfidence and availability bias create a systematically positive outlook that ignores base rates.

!Sunk Ship

Commitment to a failing course escalates to justify sunk costs, even as evidence mounts.

Reference class base rates

Across all 143 curated case studies in our library:

89%

base failure rate

-0%

partial / mixed

11%

base success rate

■ Failure■ Partial■ Success

Lessons learned

Marketing AI capabilities before the technology can reliably deliver creates a credibility gap that erodes customer trust.
Planning fallacy in AI healthcare deployments consistently underestimates the difficulty of working with messy, unstructured clinical data.
Confirmation bias led IBM to showcase cherry-picked success stories while ignoring systemic accuracy failures.

Source: Casey Ross and Ike Swetlitz, "IBM's Watson Supercomputer Recommended Unsafe and Incorrect Cancer Treatments" (STAT News, 2018) (News Investigation)

Pilot slots available