Customer Service Automation
AT&T needed to scale AI-powered customer-care agents without letting latency, operating cost, stale documentation, or model drift degrade the quality of service delivered across a massive and constantly changing support environment.
Reported impact
AT&T improved AI-agent accuracy by up to 40% on key metrics, lowered latency and operational cost with lightweight fine-tuned models, and used the platform to cut call-center analytics cost by 84% while establishing a reusable fine-tuning platform for additional enterprise use cases.
AT&T built a feedback-driven AI platform around NVIDIA AI Enterprise, NVIDIA NeMo microservices, and NVIDIA NIM to continuously improve the Ask AT&T agent stack. The company uses NeMo Curator to cleanse and filter training data from domain experts and historical logs, NeMo Customizer to fine-tune candidate base models such as Mistral, Mixtral, and Llama, NeMo Evaluator to score answer quality with Rouge and BERT F1, and NeMo Retriever to connect agents to fast-changing internal repositories. Fine-tuned models are then deployed as NIM microservices on secure infrastructure, where a data flywheel feeds production feedback back into curation and retraining. This architecture lets AT&T trade heavyweight generic models for lightweight optimized ones while keeping agents current as thousands of documents change each week.
Application logs and internal documents are curated with NeMo Curator, candidate models are fine-tuned and evaluated with NeMo microservices, current enterprise knowledge is retrieved at inference time, and NIM microservices serve the optimized agents while production feedback loops back into the next curation cycle.
AI agent responses showed up to 40% improvement in accuracy in responses with post-training.
AT&T leveraged NVIDIA NeMo microservices to create a self-sustaining cycle of data curation, model fine-tuning, and real-time evaluation.
AT&T used NeMo Retriever to set up pipelines with quick access to the company's repository.
The deployment of Ask AT&T with NVIDIA NIM and NeMo enabled AT&T to realize an 84% decrease in call center analytics cost.
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
NVIDIA Sample Aug 29, 2025 T-Mobile T-Mobile needed more accurate, lower-latency call transcripts so its Expert Assist software could recommend answers in real time without making customers repeat themselves in noisy service interactions. 10 percent higher accuracy NVIDIA A100 Tensor Core GPU · NVIDIA NeMo · NVIDIA Riva NVIDIA Sample Aug 25, 2025 Singtel Singtel needed to make its 5G enterprise offering compelling for AI-heavy customers, but real-time video analytics, AR/VR, and chatbot workloads required low-latency edge compute, shared GPU capacity, and… NVIDIA A40 GPUs · NVIDIA AI Enterprise · NVIDIA T4 GPUs NVIDIA Sample Archived · Sep 25, 2024 Amdocs Amdocs needed to help telecom providers automate complex customer and network workflows with domain-specific AI while lowering inference cost and latency enough for near-real-time production use. Reduced query latency by approximately 80 percent Llama Nemotron · NVIDIA AI Enterprise · NVIDIA DGX Cloud 