· JSON

2026-10-27-sources

{
  "status": "html-v11: F02–F04 in main; F01 and F05 in backup; F06–F07 retained as research reserves",
  "sources": [
    {
      "id": "F01",
      "publisher": "Holcim",
      "title": "Strong start to the year",
      "published": "2026-04-24",
      "url": "https://www.holcim.com/media/media-releases/q1-2026-results",
      "location": "Artificial Intelligence unlocking growth and value creation",
      "status": "company target",
      "facts": "CHF 200 million recurring EBIT benefits targeted by 2028; around CHF 20 million annual growth investment; 38 initiatives.",
      "limits": "Benefits include cost avoidance and savings. This is a future target, not realised savings or a demonstrated 10x ROI. Investment figure is not a complete lifecycle cost account.",
      "corroboration": "https://www.holcim.com/innovation/artificial-intelligence",
      "accessed": "2026-10-27",
      "presentation_usage": "backup A7"
    },
    {
      "id": "F02",
      "publisher": "Klarna",
      "title": "First Quarter 2025 Results",
      "published": "2025-05",
      "url": "https://mb.cision.com/Main/22566/4152195/3455723.pdf",
      "location": "page 7; footnote 4",
      "status": "company reported historical KPI",
      "facts": "Customer service and operations cost per completed transaction: USD 0.32 Q1 2023; 0.25 Q1 2024; 0.19 Q1 2025.",
      "limits": "Adjusted expenses exclude share-based payments, depreciation/amortisation, restructuring and other. Denominator is completed payment transactions, not chats. Before/after trend does not isolate AI causality.",
      "accessed": "2026-10-27",
      "presentation_usage": "slide 18"
    },
    {
      "id": "F03",
      "publisher": "Klarna / SEC",
      "title": "Klarna Group plc 2025 annual report",
      "published": "2026",
      "url": "https://www.sec.gov/Archives/edgar/data/2003292/000200329226000007/klar-20251231.htm",
      "location": "paragraph beginning Our AI assistant handled 80%",
      "status": "company reported estimated savings",
      "facts": "Approximately USD 59 million customer-support AI savings in 2025; assistant handled 80% of customer-service chats.",
      "limits": "Management claim in a regulatory filing; no independent causal audit or complete implementation/AI running-cost reconciliation established here. Human support remains available.",
      "accessed": "2026-10-27",
      "presentation_usage": "slide 18"
    },
    {
      "id": "F04",
      "publisher": "AWS and Care Access coauthors",
      "title": "How Care Access achieved 86% data processing cost reductions and 66% faster data processing with Amazon Bedrock prompt caching",
      "published": "2025-11-20",
      "url": "https://aws.amazon.com/blogs/machine-learning/how-care-access-achieved-86-data-processing-cost-reductions-and-66-faster-data-processing-with-amazon-bedrock-prompt-caching/",
      "location": "Benefits and impacts",
      "status": "customer and vendor reported production outcome",
      "facts": "86% lower Amazon Bedrock costs; 66% less processing time per record after prompt caching.",
      "limits": "Bedrock service spend, not total organisational cost. No absolute spend, matched sample size, uncertainty, or formal clinical-quality evaluation disclosed. Do not invent dollars or assert unchanged clinical accuracy.",
      "accessed": "2026-10-27",
      "presentation_usage": "slide 13"
    },
    {
      "id": "F05",
      "publisher": "NVIDIA",
      "title": "Pinterest Brings Conversational AI to Visual Discovery With NVIDIA Blackwell and Dynamo",
      "published": null,
      "url": "https://www.nvidia.com/en-us/case-studies/pinterest/",
      "location": "Building on Blackwell to Accelerate AI Inference and Ad Performance",
      "status": "vendor published customer case",
      "facts": "Post-trained open models on Pinterest data have inference transaction costs below 8% of comparable closed proprietary models.",
      "limits": "No named baseline model, audited absolute bill, full training/engineering TCO or matched quality test disclosed. Not evidence that local AI universally saves 92%.",
      "corroboration": "https://newsroom.pinterest.com/news/newsroom-pinterest-x-nvidia/",
      "accessed": "2026-10-27",
      "presentation_usage": "backup A6"
    },
    {
      "id": "F06",
      "publisher": "AWS",
      "title": "Amazon Q Developer just reached a $260 million dollar milestone",
      "published": "2024-08-01",
      "url": "https://aws.amazon.com/blogs/devops/amazon-q-developer-just-reached-a-260-million-dollar-milestone/",
      "location": "business impact methodology",
      "status": "company estimate",
      "facts": "USD 260 million annual cost savings estimated from infrastructure hosts removed after AI-assisted Java 17 upgrades.",
      "limits": "Distinct from the separately estimated 4500 developer-years. Not all labour savings, not lower model-inference spending, and not a full net ROI calculation.",
      "accessed": "2026-10-27",
      "presentation_usage": "research reserve; not in presentation"
    },
    {
      "id": "F07",
      "publisher": "Holcim",
      "title": "Holcim launches world's first cement plant digital twin",
      "published": null,
      "url": "https://www.holcim.com/media/company-news/first-cement-plant-digital-twin",
      "location": "Plants of Tomorrow programme paragraph",
      "status": "company reported programme outcome",
      "facts": "More than CHF 70 million annual savings across 1600+ applications and 140 sites.",
      "limits": "Broad Industry 4.0 programme. Do not attribute the entire amount to generative AI or combine with F01 as a comparable time series.",
      "accessed": "2026-10-27",
      "presentation_usage": "research reserve; not in presentation"
    }
  ]
}