Classifier — McKinsey & Company
Role & scope: In an AI enablement engagement for a Banking client, we slashed delivery time of a highly manual classification and data mapping operation. By architecting a custom Generative AI Classifier, we achieved high classification rate, enabling teams to expand AI adoption.
Revolutionizing banking intelligence through GenAI taxonomy mapping—architecting a custom classifier that turned weeks of manual data mapping into days while freeing analysts for strategic client work.
The Opportunity
Analysts supporting Banking clients faced a critical taxonomy mapping bottleneck, consuming 30% of their time and introducing two-week latency per engagement. Traditional ML models peaked at ~40% accuracy; simple GenAI prompting achieved only ~20%. Classification required expert judgment to decode acronyms, nuanced definitions, and rapidly changing semantics—capping capacity at 3–4 concurrent engagements.
The Solution
I engineered a structured, multi-model GenAI Classifier workflow with candidate generation, reranking, confidence scoring, human-in-the-loop review, and quality dashboarding. Core differentiators included two-stage verification to prevent hallucinations, a quality assessment framework tracking match rates against every prompt change, and a controller-in-the-loop model shifting analysts from manual labor to strategic oversight.
The Impact
90% accuracy on existing datasets · 60% on new and cold-start lists · Throughput expanded from 3–4 to 10+ concurrent client scenarios · Research timelines reduced from 12+ months to 3–4 months · Reusable research frameworks and standardized protocols for firm-wide AI adoption
- 65% — cycle time reduction—from two weeks to three days per engagement
- 90% — accuracy match rate on existing taxonomy datasets
- 50% — reduction potential in manual analyst effort