Walmart · Microsoft
22% — global ecommerce growth
Independently reportedRetail & e-commerce · Search & retrieval · LLM · In production · 2024
Search & retrieval
K&L Wines · 923
Vendor-reportedRetail & e-commerceIn productionMid-marketLLM
Vendor-reported. The customer is named and the numbers are quoted from the source page, but the account comes from the vendor. No independent confirmation.
Reported by ninetwothree.co (vendor self-report). Checked against the source page on 2026-09-06. 6 of 6 figures below appear on that page word for word.
The problem
Team spent hours daily hunting through over 500,000 wine records to find the right SKU due to chaotic naming variations.
What was deployed
- Vendor
- 923
- Technique
- LLM
- Build or buy
- Bought off the shelf
- Deployment
- Custom build
- Data used
- 500,000+ wine records
- Scale
- 500,000+ wine records; 7 employees affected; family-owned retailer since 1976
What changed
Each row is quoted from the source. Figures we could not find on the page in those words are marked — they are kept, not deleted, so you can judge them.
14 hoursdaily busywork eliminateddaily
“eliminated 14 hours of daily busywork”
Quoted word for word from the source
99%accuracy rate
“delivered 99% accuracy”
Quoted word for word from the source
273000Annual SavingsAnnual
“$273K+ Annual Savings”
Quoted word for word from the source
60xSpeed Increase
“60x Speed Increase”
Quoted word for word from the source
1 minuteTime to match 10 winesfrom 1 hour
“1 hour → 1 minute (10 wines)”
Quoted word for word from the source
2 FTEEquivalent Capacity
“2 FTE Equivalent Capacity”
Quoted word for word from the source
ROI as stated: Solution paid for itself in first month. Everything after is pure profit.
“This saves a few hours per day for seven people. It brings two new employees to the company without having to onboard anyone.”
Difficulties and limits
- Off-the-shelf AI solutions were unreliable
- LLMs cannot handle 500K-record context windows, hallucinate, cost too much at scale, and fail in production
- Traditional engineering scripts did not work at scale with imperfect production data
- Risk of unnecessary technical debt: building a machine learning solution beyond the necessary scope. K&L needed a tool that worked, not another tool to manage.
Sources
- ninetwothree.co — vendor self-report, 2025-11-12
Similar deployments
Osprey · NOVOS
72.6% — share of AI answers
Vendor-reportedRetail & e-commerce · Search & retrieval · LLM · In production
Safe · Victoria Olsina
40.88% — Enterprise keyword visibility
Vendor-reportedOther · Search & retrieval · LLM · In production
Novartis · Market Logic Software
640% — User growth rate
Vendor-reportedPharma & biotech · Search & retrieval · LLM · Scaled · 2020
SoWork · iGrow
100% — AI Visibility Score
Vendor-reportedTechnology & software · Search & retrieval · NLP · In production · 2026
Philips · Market Logic
89% — Proficiency in answering business questions
Vendor-reportedHealthcare · Search & retrieval · LLM · Pilot