How we build the AI Use Case Library
Every number is quoted from its source and says who made the claim, and every page is checked against its sources before it is published. This page explains how, including where AI does the work.
What a use case page is
A use case is a job AI does inside an organization, described independently of any vendor: "AI agent for card dispute intake", not a product. Each job has one canonical page. Deployments by different organizations are recorded as evidence on that page, instead of as separate thin pages.
How we find deployments
In each research run, an agent searches public sources: organization newsrooms and annual reports, earnings calls, regulator publications and government AI inventories, vendor case study libraries, and trade press. Trade press is used to discover a deployment; we then cite the primary source.
The research agent proposes records. Before a page is published, a fact check agent fetches every cited source again and checks every number, claim and regulation on the page, and a second, independent agent tries to refute the page. Only pages that pass both are published. The instructions the agents follow are the rules described on this page.
Evidence grades
- A
- Independent oversight: a regulator, supervisor or audit office reporting on the deployment, or an audited figure.
- B
- The deploying organization itself: press release, annual report or 10-K narrative, earnings call, executive talk, or a public body's own entry in an AI register.
- C
- A vendor case study that names the customer.
- D
- An anonymous vendor claim, an analyst estimate or an anonymized Blits.ai deployment.
Every number carries its quote and its claimant
A metric enters the library only with the exact sentence from the source that states it, the source link, and who made the claim: the organization itself, a vendor, a regulator, or an independent party. A vendor's case study number is labelled as a vendor claim even when it names the customer.
Our source check fetches the cited pages and confirms that each quoted sentence is on the page, before publication and whenever a record is checked again. Forecasts, targets and market sizing are never recorded as deployment results; analyst estimates appear separately and are labelled as estimates. "No outcome disclosed" is a valid and visible answer.
Verification levels
- AI extracted: Extracted by the research agent; the quote has not yet been checked against the source.
- Quote checked: The quote and number were found on the live or archived source by our source check, and the record was reviewed before publication.
- Organization confirmed: The deploying organization confirmed the record.
- Blits.ai record: An anonymized deployment on the Blits.ai platform, from Blits.ai's own records.
How benchmarks are calculated
For each KPI we take one data point per organization, from public records only, so one organization with several records cannot outweigh the rest. Within a record the first value reported for the KPI counts. When one organization has several records, we keep the value with the strongest evidence (a firm value before an "up to", then the better grade, then the more independent claimant), never the highest value. Anonymized records never enter a benchmark, because a reader cannot check them.
"Up to" values are ceilings, not typical results: we list them next to the benchmark but leave them out of the median. A median is shown from 3 pooled values up, with the reported range, the number of values and the mix of claimants. Below that we show the strongest single reported value, with its organization and who made the claim.
Absolute volumes (conversations handled, hours saved, money saved) depend on the size of the organization, so they are listed but not pooled into a median.
Indicative value
Each use case can carry a worked example for a named reference organization, such as "a retail bank with 1 million active customers". It shows a low and a high scenario, every input with its basis (a benchmark on the page, a cited source, or a stated editorial assumption) and the formula, so anyone can rebuild the number with their own inputs. It is an order of magnitude, not a forecast.
Neutrality and Blits.ai
The library covers every vendor and every organization, including Blits.ai competitors and their customers. Blits.ai content appears only in a separate, clearly marked block on each page that describes how the use case is built on the Blits.ai platform. Blits.ai deployments are anonymized unless the customer has given written approval, are graded D, and never count toward benchmarks.
When a page is published and indexed
A page is published after the fact check and the independent review have both passed. It is offered to search engines only when it has at least two public deployment records, or one of grade A or B. Hub pages for an industry, function or AI pattern are indexed once they hold at least three indexable use cases, KPI pages once they carry three reported values, and organization pages once an organization has three public records. On an industry hub, benchmarks count only organizations in that industry.
How AI is used to make this library
AI agents find sources, extract quotes, draft sections and fact check them, and an independent agent reviews every page before publication. Len Debets, Blits.ai's CTO, is the named author and editor, sets these rules and is accountable for what is published and for every correction. Every page shows when it was last verified and keeps a changelog.
Corrections and submissions
Spotted an error, an outdated number, or a deployment we missed? Tell us through the contact form with the page and the source. Organizations can confirm or correct records about themselves the same way. We fix verified errors and note them in the changelog.