usecaseinai

Pattern library · 4 documented systems · updated 17 Sept 2026

AI use-cases, documented from source.

Real-world AI systems across industries and everyday life, each one sourced and broken down — the decision, the approach, and the result. No hype, no demos.

20% less inventory: how AB InBev makes 85% of its demand plans touchless

In production

01 · INGEST

Sales history, POS, distributor sell-through, promotions, price, weather, calendar events

02 · ENRICH

Correct censored demand (stock-outs logged as zero), build features: lags, seasonality, promo flags

03 · FORECAST

Ensemble (gradient boosting + deep learning) → a probabilistic forecast at SKU-location-week

04 · GATE

Auto-commit only the forecasts above the confidence threshold; these run touchless into the plan

05 · ROUTE

Everything below the line (new SKUs, promos, anomalies) goes to a human demand planner

Domain
Supply chain / CPG
Gating eval
Backtest WMAPE + Forecast Value Added; auto-commit only above a confidence threshold
Primary failure
Censored demand and silent drift — looks fine, compiles into the plan, quietly stale
Status
PRODUCTION

AI-driven demand forecasting behind a confidence gate: an ensemble produces a probabilistic forecast per SKU-location-week, and only the predictions above a confidence threshold commit touchless — everything else routes to a human planner.

Source: usecaseinai analysis. Operational figures are o9 Solutions' AB InBev case study (vendor-reported, marked in-text). Independent ranges are from McKinsey operations research. (Swap in the Substack long-form URL once live.)
Open teardown →

Pattern index

All entries →
IDPatternStatus
20% less inventory: how AB InBev makes 85% of its demand plans touchlessAI-driven demand forecasting behind a confidence gate: an ensemble produces a probabilistic forecast per SKU-location-week, and only the predictions above a confidence threshold commit touchless — everything else routes to a human planner.Source: usecaseinai analysis. Operational figures are o9 Solutions' AB InBev case study (vendor-reported, marked in-text). Independent ranges are from McKinsey operations research. (Swap in the Substack long-form URL once live.)PRODUCTION
The 80% problem: how banks modernize legacy COBOL with AIAI code comprehension, explanation, and review are in production at bank scale; autonomous COBOL-to-Java migration of core systems is still pilot-stage, gated by parallel-run validation against the original system.Source: usecaseinai — narrative long-form (TODO: add source URL once the essay is live)PRODUCTION
Gaming vs. saving lives: the $2.5M drug discovery problemEmbedding search over 200M protein sequences, GPU-accelerated structure prediction, and diffusion-based docking replace a wet-lab search that would otherwise take roughly two centuries.Source: usecaseinai — "Gaming vs Saving Lives: A $2.5 Million Drug Problem"PRODUCTION
Reclaim 2 hours: how Superhuman classifies 500M+ emails a dayMulti-task email classification (importance, category, spam) behind one shared encoder, with a calibration layer that only automates the predictions it's confident about.Source: usecaseinai — "Reclaim 2 Hours: Email Prioritization"PRODUCTION

Recent teardowns

  • 17 Sept 2026
    20% less inventory: how AB InBev makes 85% of its demand plans touchless

    AB InBev's demand forecasting isn't a smarter model — it's a confidence line. It automates the forecasts the system trusts and routes only the exceptions to a human planner.

    demand-forecasting, supply-chain, cpg

  • 12 Sept 2026
    The 80% problem: how banks modernize legacy COBOL with AI

    Citi's gen-AI developer tooling automated 740,000 code reviews and frees roughly 100,000 engineering hours a week. But the value isn't that AI writes code — it's where AI plugs into the modernization pipeline, and where it silently ships wrong-but-compiling code into systems that move money.

    legacy-modernization, financial-services, code-generation, agentic

  • 16 Dec 2025
    Gaming vs. saving lives: the $2.5M drug discovery problem

    Innophore needed one enzyme out of 200 million known protein sequences. A three-model AI pipeline on H100 GPUs found it in two weeks for under $100,000 — a search that would have taken a wet lab roughly 200 years.

    drug-discovery, life-sciences

  • 11 Dec 2025
    Reclaim 2 hours: how Superhuman classifies 500M+ emails a day

    Superhuman's email triage runs on dozens of fine-tuned classifiers, not one generic model — and a calibration pattern that only automates the predictions it's actually sure about.

    classification, productivity

One teardown a fortnight

New entries and revisions to existing ones, with a short note on what changed and why. No digests, no roundups.