Popular searches
Not sure where to start?Talk to our team about running sovereign AI in your environment.
↑ ↓ Navigate↵ OpenEsc Close
AT&T cut iPhone fraud 80%, saved millions a year on fleet operations, and slashed LLM costs up to 90% by building predictive and generative AI on H2O.ai.
AT&T's data science team was small in 2015, ran on open source, and took months to ship AI projects that returned value. The company needed to democratize AI across business units and deliver solutions far faster.
To scale AI across the organization, AT&T had to democratize it across business units and shorten delivery timelines dramatically.
AT&T moved from open source to AutoML to a fully integrated AI-as-a-Service platform. In 2021 it deployed an AIaaS layer that links on-premises infrastructure, data cloud providers, and H2O AI Cloud with AT&T's own proprietary services.
The team used H2O Driverless AI for AutoML and co-created the H2O AI Feature Store with H2O.ai to store and share features across data scientists and engineers, so proven work could be reused instead of rebuilt.
The platform reduced iPhone fraud by 80%, addressing a fraud problem worth roughly $1 billion a year in the U.S. market. Predictive maintenance that cuts truck breakdowns across the 7,000-vehicle fleet saves about $7 million annually, and route optimization that lowers fuel consumption saves about $10 million a year.
Project deployment is far faster, and the number of AI users across AT&T has grown exponentially.
“H2O.ai continues to innovate and define what's next in AI. With their partnership, we're moving faster, delivering more value, and co-creating new AI products.”Mark AustinVP Data Science, AT&T
AT&T handles 15 million customer calls a year, all recorded, transcribed, and summarized. Extracting insight from that volume with large models like GPT-4 was accurate but too expensive to scale.
Doing that with large language models such as GPT-4 delivered accuracy but carried computational costs too high to run at that scale.
AT&T distilled large models like GPT-4 into three smaller, fine-tuned open-source models to cut compute cost and improve scalability. For 20 of the 80 categories it used H2O Danube 1.8B, a 1.8 billion parameter small language model fine-tuned with the no-code H2O LLM Studio.
The production system is an ensemble: 10 categories handled by Llama, 20 by Danube, and 50 by a classifier, working together to pull actionable insight from every call.
The ensemble reached 91% accuracy while cutting cost by 90%, bringing total cost to 35% of the previous solution, with Danube accounting for just 10% of it.
Processing runs 3x faster with a 75% latency improvement and 5x greater scalability.
AT&T's CDO Office wanted non-technical business users to query databases in plain English for customer targeting and campaign optimization. The work demanded high precision, but running large language models against it was prohibitively expensive.
Precision was critical, since the insights would drive key decisions. Large language models like GPT-4 had the analytical power, but running them against every query was prohibitively expensive.
AT&T fine-tuned a smaller model, the Llama SQL Coder 8 billion parameter model, with the no-code H2O LLM Studio, choosing fine-tuning over a RAG approach to cut both cost and latency. The team used semantic similarity search, careful data curation, and data profiling to teach the model AT&T's domain terminology, avoiding synthetic data for legal reasons.
Using LLM Studio, the team ran a grid search for the right learning rate and epochs, tracked training logs for convergence, then deployed and benchmarked the result against the original GPT-4 solution.
Benchmarked against GPT-4, the fine-tuned model proved not only cheaper but more accurate. AT&T secured the #1 spot on the global BIRD-SQL text-to-SQL benchmark, used by organizations such as Google, IBM, Alibaba, and Stanford, and currently ranks #2.
Running the 8 billion parameter model is effectively free, versus two cents per call for the previous large-model solution, and business users get customized, domain-specific query generation without a data analyst in the loop.
“While a small language model may not necessarily outperform a large one, some tasks are simple enough that a small model can achieve the same quality with significantly lower costs.”Farbod TavakkoliData Scientist, AT&T
AT&T Inc. helps more than 100 million U.S. families, friends and neighbors, plus nearly 2.5 million businesses, connect to greater possibility. From the first phone call 150 years ago to its 5G wireless and multi-gig internet offerings today, AT&T innovates to improve lives.
More information about AT&T is available at about.att.com.