AI Data Foundation ROI Calculator

The business case compares the governed AI against your actual alternative: analysts answering questions manually. It nets out review labor, rework, residual errors above the human baseline, run costs, and the implementation itself — that is the number a CFO can fund. The avoided exposure compares the governed AI against shipping the raw 21% setup at full volume, which nobody would actually do. It is a real risk picture and a useful argument, but it is not realized savings. Never add the two together.

Because at low question volume, the implementation and run costs can outweigh the deflected analyst time, and the model says so rather than hiding it. That is a feature: a calculator that can say no is one whose yes means something. If the verdict is negative, the levers are volume (more questions routed through the AI), a leaner implementation, or revisiting the value inputs if your analysts spend more time per question than the default.

Anthropic published the results of testing its own internal analytics agent. With a raw, unstructured warehouse setup it answered roughly 21% of questions correctly. After layering in governed datasets, a semantic layer, and structured skills, accuracy passed 95%. Notably, giving the agent read access to thousands of past correct queries moved accuracy by less than one point — the bottleneck was governed context, not data access.

Because decision costs are skewed: most swayed decisions are cheap, a few are very expensive, and a single number invites you to anchor on the typical case while the rare severe one drives the real average. The model blends your three tier costs at a fixed 70% / 25% / 5% mix — the mix is a stated model assumption, the costs are yours to calibrate. With the defaults that works out to about $2,850 expected cost per swayed decision.

No. Only two points are published (21% raw, 95% governed); everything between them is a model. The estimate weights your six scores by how much each dimension mattered in the published findings, then maps them through an S-curve anchored to those two points. The S-shape and the weights are assumptions, stated openly. The only way to know your actual accuracy is to test your AI's answers against known-correct results — which is also why Evaluation practice is one of the six dimensions.

Deliberately excluded: trust erosion in dollars (no defensible way to price it — it lives in the severe tier narrative instead), IRR (NPV at your discount rate answers the same question more reliably), and broad AI productivity gains, which belong to our separate AI Productivity ROI Calculator — claiming them in both places would double-count. Also not modeled: accuracy differences across question types, question-volume growth, and compliance or audit risk. The exclusions mostly make the case smaller, which is the point.

In Anthropic's experience the biggest single lever was the semantic layer — shared, governed definitions that map a business question to the right metric. The note under the readiness chart computes this for your specific scores: it shows which one dimension, improved by three points, moves your estimated accuracy the most. Raw data access mattered least, which is the counterintuitive finding: wiring the LLM to more tables is the one move that barely helps.