AT&T Cuts AI Costs 80% by Shifting Routine Tasks to Open-Source Models

AT&T has significantly reduced its AI spending by migrating high-volume, routine tasks like call transcription and customer service to open-weight models, with its chief data and AI officer Andy Markus reporting costs down as much as 80 percent compared to earlier this year. Open models now account for 40 percent of AT&T's AI usage, up from 20 percent in May, with Markus projecting that figure could reach 60 percent in the coming months. The company is specifically using Meta's Llama and Google's Gemma for customized tasks, while deliberately avoiding Chinese open models such as DeepSeek and Moonshot despite their lower costs. Closed frontier models from OpenAI and Anthropic are being retained for more complex tasks like heavy coding and multimodal work where open alternatives fall short. The broader trend is echoed by companies like Airbnb and Deloitte, reflecting a growing corporate strategy of reserving expensive frontier AI for specialized needs while offloading simpler, high-volume workloads to cheaper open-source alternatives.