Methodology
Every case in the casefile answers three questions: who deployed the agent, what it does, and how strong the evidence is that it worked. We only publish what a source says, and we show you the source.
Evidence grades
| Grade | Definition | Cases |
|---|---|---|
| Grade A | Named organization, quantified outcome, independent source (news, paper, filing, or customer quote outside vendor marketing). | 11 |
| Grade B | Named organization with a quantified outcome, from the vendor or the organization's own announcement. | 126 |
| Grade C | Named organization, no specific outcome (deployment announced only, or only vague qualitative claims). | 10 |
| Grade D | Anonymous or community report, or a demo/concept/product launch with no named deployment. Shown, but labeled and excluded from "proven" filters. | 5 |
Rules we apply when grading:
- Quantified means specific. “Stronger performance” or “a game-changer” is not an outcome. A number, a before/after, or a concrete measurable change is.
- Independent means a news outlet, peer-reviewed paper, regulatory filing, or customer quote outside the vendor's own marketing. Peer-reviewed papers count as independent even when vendor-authored, but we disclose funding and conflicts on the case page.
- Self-reported metrics are flagged. Most outcomes are reported by the organization or its vendor. We say so on every outcome.
- Anonymous = D. If the customer is not named (“a leading specialty pharmacy”), the case is grade D regardless of numbers.
- Product launches without a named deployment are D and carry the status
launch. - Conflicting numbers are disclosed. When sources disagree, we use the more rigorous source and note the discrepancy.
Deployment status
production: In regular use with real customers, patients, or employees.pilot: Limited rollout, proof of concept, early-access program, or retrospective evaluation.launch: Vendor product made generally available; no named customer deployment reported in the source.announced: Organization has said it is adopting or building the agent; no live usage reported.
How a case gets in
- Collect. Vendor customer stories, press releases, trade press (healthcare and MedTech titles included), research papers, and public developer sources are monitored on a schedule using RSS feeds and official APIs. We respect robots.txt and do not scrape behind logins.
- Deduplicate. Multiple articles about the same deployment become one case with multiple sources.
- Extract. A model fills the case schema using only the source text. Every factual field must be supported by an exact quote of 40 words or fewer. Fetched pages are treated as untrusted data: the extraction step has no tools and follows no instructions found in content.
- Verify. An automated check re-fetches every source and confirms (a) each quote appears verbatim on the page and (b) every number in every reported outcome appears in the cited source. Any mismatch fails the case. Pages that block automated access are checked through a mirror (e.g. Europe PMC for journals) or a dated manual check, recorded in the public repository.
- Grade and gate. We assign the grade, then reject anything without a primary source, with fewer than two facts beyond the headline, too similar to an existing case, or making patient-outcome claims without a peer-reviewed or regulatory source.
- Human review. During launch, every case is reviewed before publication.
- Refresh. Links and quotes are re-verified weekly; the “last verified” date on each case reflects the latest pass.
As of Oct 17, 2026, 152 of 152 published cases passed the latest automated verification.
Hub pages
Industry and role hubs are indexed by search engines only once they contain at least 5 verified cases. Smaller hubs stay visible to readers but carry noindex.
Healthcare and MedTech
We cover operational and commercial workflows. We do not publish PHI, discuss off-label use, or make device or drug efficacy claims. Clinical decision-support cases describe what the tool does and what the organization reported; they are not medical advice.
AI-assistance disclosure
Case summaries are drafted with AI assistance and checked by the verification process above. Outcome figures and quotes are copied from sources, never generated. Each case lists the date it was last verified.
Independence and money
No organization can pay to be included, to change a grade, or to rank higher. Sponsored content is labeled at the placement and flagged in the public data. See our sponsorship policy.
Corrections
Found an error, a newer outcome, or a dead source? See corrections. We aim to fix verified errors within 24 hours and log every change.