{
  "post": {
    "title": "A GraphRAG business case, done the evidence-based way",
    "url": "https://bardak.ai/posts/graphrag-business-case",
    "updated": "2026-10-29T18:55:41.411Z"
  },
  "citation": "A GraphRAG business case, done the evidence-based way, Ivan Bardak, bardak.ai, https://bardak.ai/posts/graphrag-business-case",
  "datasets": [
    {
      "key": "evidence-based-management-for-a-graphrag-decision",
      "title": "Evidence-based management for a GraphRAG decision",
      "type": "benchmarkTable",
      "source": "Based on the six steps of evidence-based management (Center for Evidence-Based Management)",
      "caption": "",
      "csv": "https://bardak.ai/data/graphrag-business-case/evidence-based-management-for-a-graphrag-decision.csv",
      "header": [
        "Step",
        "Question it answers",
        "What goes into the business case"
      ],
      "rows": [
        [
          "Ask",
          "What decision are we making, and for whom?",
          "The decision in one sentence, the baseline, the owner"
        ],
        [
          "Acquire",
          "What evidence exists?",
          "Studies, vendor claims, your own data, and what your experts and users say"
        ],
        [
          "Appraise",
          "How much should we trust each piece?",
          "Who produced it, how it was measured, how close it is to our situation"
        ],
        [
          "Aggregate",
          "What does it add up to?",
          "A range for the expected improvement and a cost model, not a single number"
        ],
        [
          "Apply",
          "What do we do now?",
          "A pilot with a go or no-go threshold agreed in advance"
        ],
        [
          "Assess",
          "Did it work?",
          "Metrics after rollout and a date on which to decide"
        ]
      ]
    },
    {
      "key": "appraising-the-graphrag-evidence",
      "title": "Appraising the GraphRAG evidence",
      "type": "benchmarkTable",
      "source": "Edge et al. 2024; Han et al. 2025; Xiang et al. 2025; Microsoft Research 2024",
      "caption": "The weights are a judgement, and the point is to make that judgement visible so others can challenge it.",
      "csv": "https://bardak.ai/data/graphrag-business-case/appraising-the-graphrag-evidence.csv",
      "header": [
        "Evidence",
        "Weight",
        "Why"
      ],
      "rows": [
        [
          "Edge et al. 2024: 72-83% win rate for global answers",
          "Moderate for whole-corpus questions, low otherwise",
          "Written by the method's authors; LLM judge; questions built to need a whole-corpus view"
        ],
        [
          "Han et al. 2025: RAG vs GraphRAG on public benchmarks",
          "Moderate to high",
          "Independent team; ground truth; same models for every method; public data, not enterprise data"
        ],
        [
          "GraphRAG-Bench 2025: accuracy and token costs",
          "Moderate",
          "Independent; ground truth; two corpora; some token counts inconsistent between text and table"
        ],
        [
          "Microsoft: LazyGraphRAG indexes at 0.1% of GraphRAG's cost",
          "Low until reproduced",
          "Vendor blog; an independent benchmark measured about 90% of GraphRAG's indexing tokens"
        ],
        [
          "Your labelled question sample and pilot results",
          "High for this decision",
          "Your questions, your documents, your users"
        ]
      ]
    },
    {
      "key": "prompt-token-cost-per-month-at-20-000-questions-illustrative",
      "title": "Prompt-token cost per month at 20,000 questions (illustrative)",
      "type": "chart",
      "source": "Calculation from Xiang et al. 2025 token measurements",
      "caption": "Token ratios measured in GraphRAG-Bench; volumes and price are assumptions. A full index rebuild would add about $292 for GraphRAG and $147 for HippoRAG2 on this corpus, priced at the input rate for simplicity; completion tokens usually cost more.",
      "csv": "https://bardak.ai/data/graphrag-business-case/prompt-token-cost-per-month-at-20-000-questions-illustrative.csv",
      "header": [
        "method",
        "Cost per month"
      ],
      "rows": [
        [
          "Vector RAG",
          "8.79"
        ],
        [
          "HippoRAG2",
          "10.08"
        ],
        [
          "GraphRAG local search",
          "387.07"
        ]
      ]
    },
    {
      "key": "monthly-value-by-accuracy-gain-and-multi-hop-share-illustrative",
      "title": "Monthly value by accuracy gain and multi-hop share (illustrative)",
      "type": "chart",
      "source": "Illustrative calculation; gain range informed by Han et al. 2025",
      "caption": "Value = 20,000 questions x multi-hop share x gain x $10. The share comes from your labelled sample; the gain from your pilot.",
      "csv": "https://bardak.ai/data/graphrag-business-case/monthly-value-by-accuracy-gain-and-multi-hop-share-illustrative.csv",
      "header": [
        "gain",
        "10% of questions multi-hop",
        "20% multi-hop",
        "30% multi-hop"
      ],
      "rows": [
        [
          "0",
          "0",
          "0",
          "0"
        ],
        [
          "3",
          "600",
          "1200",
          "1800"
        ],
        [
          "6",
          "1200",
          "2400",
          "3600"
        ],
        [
          "9",
          "1800",
          "3600",
          "5400"
        ],
        [
          "12",
          "2400",
          "4800",
          "7200"
        ],
        [
          "15",
          "3000",
          "6000",
          "9000"
        ]
      ]
    }
  ]
}