[{"data":1,"prerenderedAt":105},["ShallowReactive",2],{"uc-org-at-and-t":3},{"organization":4,"includeUnpublished":13,"evidence":14},{"slug":5,"name":6,"country":7,"region":8,"industry":9,"wikidata":10,"records":11,"useCases":11,"indexable":12},"at-and-t","AT&T","US","north-america","telecommunications","Q35476",3,true,false,[15,46,77],{"title":16,"useCases":17,"organization":19,"vendors":20,"summary":26,"stage":27,"year":28,"channels":29,"languages":30,"metrics":31,"outcomeDisclosed":12,"sources":32,"verification":38,"grade":41,"id":42,"useCaseTitles":43},"AT&T: AI incident management system that finds outage causes and alerts customers",[18],"network-outage-communication-agent",{"name":6,"anonymized":13,"country":7,"region":8,"industry":9,"wikidata":10},[21,23],{"name":6,"role":22},"in-house",{"name":24,"role":25},"MongoDB","platform","AT&T runs an in house end to end incident management system (EEIM) that its chief data and AI officer, Andy Markus, describes to Business Insider. The system combines AI and machine learning to identify the root cause of a disruption, suggest fixes that can be made remotely or by a technician in the field, and proactively alert customers when an outage occurs. AT&T launched the proactive customer notification system in the first quarter of 2021, and since the first quarter of 2025 AI agents on the platform interact with customers to collect information about outages. Markus says the system helped prevent 3.1 million unnecessary field dispatches and reduced customer downtime by more than 12 million hours over the last year. Those figures cover the whole incident system, not the customer notices, so they are not recorded as metrics for this use case.","scaled",2026,[],[],[],[33],{"url":34,"title":35,"publisher":36,"date":37},"https:\u002F\u002Fwww.businessinsider.com\u002Fatt-telecom-ai-network-outage-prevention-2026-7","AT&T built an AI system to prevent network outages. It reduced customer downtime by more than 12 million hours.","Business Insider","2026-07-16",{"level":39,"checkedAt":40},"source-verified","2026-10-13","C","att-eeim-incident-management-customer-notification",[44],{"slug":18,"title":45},"AI agent for network outage detection and customer communication",{"title":47,"useCases":48,"organization":50,"vendors":51,"summary":57,"stage":58,"year":59,"channels":60,"languages":63,"metrics":65,"outcomeDisclosed":12,"sources":66,"verification":71,"grade":41,"id":73,"useCaseTitles":74},"AT&T: fine tuned Llama model that turns plain English into SQL for business users",[49],"governed-text-to-sql-analytics",{"name":6,"anonymized":13,"country":7,"region":8,"industry":9,"wikidata":10},[52,54],{"name":53,"role":25},"H2O.ai",{"name":55,"role":56},"Meta (Llama)","model-provider","AT&T's Chief Data Office wanted non technical business users to query databases in plain English for customer targeting and campaign optimization, and running GPT-4 against every query cost too much. AT&T fine tuned an 8 billion parameter Llama SQL model with H2O LLM Studio, choosing fine tuning over retrieval to cut cost and latency, and taught it the company's domain terms through data curation and profiling, avoiding synthetic data for legal reasons. H2O.ai reports that the deployed model was cheaper and more accurate than the GPT-4 solution it replaced. No accuracy figure or usage volume is published.","production",2025,[61,62],"internal-tools","api",[64],"en",[],[67],{"url":68,"title":69,"publisher":53,"archivedUrl":70},"https:\u002F\u002Fh2o.ai\u002Fcase-studies\u002Fatt-text-to-sql\u002F","How AT&T became an AI company with H2O.ai","https:\u002F\u002Fweb.archive.org\u002Fweb\u002F20250206185454\u002Fhttps:\u002F\u002Fh2o.ai\u002Fcase-studies\u002Fatt-text-to-sql\u002F",{"level":39,"checkedAt":72},"2026-10-14","att-fine-tuned-text-to-sql",[75],{"slug":49,"title":76},"Governed text to SQL analytics assistant",{"title":78,"useCases":79,"organization":81,"vendors":82,"summary":85,"stage":58,"year":59,"channels":86,"languages":87,"metrics":88,"outcomeDisclosed":12,"sources":98,"verification":100,"grade":41,"id":101,"useCaseTitles":102},"AT&T: fine tuned small language models that classify customer call transcripts",[80],"customer-feedback-analysis",{"name":6,"anonymized":13,"country":7,"region":8,"industry":9,"wikidata":10},[83,84],{"name":53,"role":25},{"name":55,"role":56},"AT&T records, transcribes and summarizes its customer calls and wanted insight from every one of them, such as appointments where a technician never showed up, but running GPT-4 over that volume cost too much. According to H2O.ai's case study, AT&T distilled large models into three smaller fine tuned open source models that work as an ensemble across an 80 category classification: a Llama model handles 10 categories, the 1.8 billion parameter H2O Danube model 20, and a classifier 50. H2O.ai reports 91 percent accuracy, close to the previous and far more expensive solution, at about a third of its cost.",[61,62],[64],[89],{"kpi":90,"value":91,"unit":92,"qualifier":93,"period":94,"claimant":95,"quote":96,"sourceUrl":97},"accuracy",91,"percent","exact","call classification ensemble in production","vendor","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.","https:\u002F\u002Fh2o.ai\u002Fcase-studies\u002Fatt-call-center\u002F",[99],{"url":97,"title":69,"publisher":53},{"level":39,"checkedAt":72},"att-fine-tuned-small-models-call-analytics",[103],{"slug":80,"title":104},"AI for voice of the customer and feedback analysis",1792049868959]