AI Papers Atlas Wave 1

Method

Using AI
to keep up with AI.

This site exists to keep scholars of management, organizations, and strategy current with AI research that bears on their fields—by putting AI itself to work on the reading problem, under human command and independent QA.

Motivation

The research most relevant to our disciplines—how AI-native firms organize, how agent organizations coordinate and fail, what AI does to productivity, tasks, and managerial work—is produced at preprint speed, scattered across arXiv, SSRN, NBER, journals, standards bodies, and lab reports. Traditional scholarly infrastructure was not built for this cadence: review articles lag by years, and each lab indexes only its own output.

As of our 2026-08-20 survey of the public landscape, we found no maintained, structured paper database at this intersection. The one repository ever aimed at it froze in 2021; the active curated lists nearby are either engineering-scoped or contain no organizational research at all. Our answer is to use AI to keep up with AI: agentic campaigns do the sweeping, screening, and evidence coding; independent engines cross-check the results; humans rule the judgment calls. The output is not a link list—it is a database with provenance.

How a work enters the atlas

  1. Discovery. Scoped research campaigns sweep the sources against a frozen inclusion boundary, logging every query.
  2. Screening. Every lead gets an explicit decision—deep-code, context, merge, or exclude—with lineage and its ruled stratum recorded.
  3. Evidence coding. Admitted works are read at a recorded location; each extracted claim carries an evidence class, an exact locator, a scope boundary, and a confidence grade. Claims stay at the level of what the inspected source says.
  4. Independent QA. A separate session audits the bundle; defects force a repair round before acceptance.
  5. Landing. Accepted bundles are hash-manifested (SHA-256) and committed append-only. This site is a projection of that ledger—the counts on these pages reproduce from it.

Multi-source provenance

The current corpus contains 365 AI × management-and-organizations works, 74 separately labelled management-canon works, and 567 work claims, plus a curated roster of 15 researchers and 68 researcher claims. A separate author index contains 526 OpenAlex Author ID groups derived only from rights-safe admitted-work metadata.

  • apw-a1-author-roster: campaign 2026-08-20__apw-a1-author-roster, commit ea9690ee35f8b15257753ce423081682f9d073ed, ledger SHA-256 157d69d2c8cb….
  • apw-b1-canon-identification: campaign 2026-08-20__apw-b1-canon-identification, commit cea7db879a4485e6ed0b69053dbac3cc0e65e35f, ledger SHA-256 4582bd0950a7….
  • apw-c1-obot-journals: campaign 2026-08-20__apw-c1-obot-journals, commit ea9690ee35f8b15257753ce423081682f9d073ed, ledger SHA-256 69e19c10ab3b….
  • apw-e1-ssrn-backfill: campaign 2026-08-20__apw-e1-ssrn-backfill, commit cea7db879a4485e6ed0b69053dbac3cc0e65e35f, ledger SHA-256 05ea51693c1b….
  • apw-r0-seed: campaign 2026-08-18__ai-native-organization, commit 799340c, ledger SHA-256 ce8aff0549a7….
  • apw-work-enrichments: campaign wave1-1-openalex-work-enrichments, commit a154a94a156aa48415ec63b9bdeacbeba74371d0, ledger SHA-256 a21b84240a91….

The original gold sentinel remains independently recomputable at 34 works and 70 claims. Campaign overlaps augment provenance and claims on one canonical work; unresolved identities and metadata conflicts remain held out.

The management-canon boundary

A classic management work enters only when the B1 corpus evidence passes the ruled invocation test. The executed, release-audited intake identified 110 candidates carrying 220 corpus-evidenced invocation statements. A qualifying admitted work is labelled management-canon and never counted as an AI paper. B1 publishes the source-level invocation evidence but names an invoking public work only when that endpoint is separately admitted and evidenced; it does not invent endpoints. Source-argument claims remain a separate field and B1 adds none. Unresolved identities stay held. The rights boundary remains metadata only—no full text, regardless of age or licence.

Current data machinery

  • Wave 1 seed: 804 identifiers from the research estate are the distinct normalized identifiers in the accepted research estate. Screened curated personal references contribute 82 additional identifiers, producing 886 identifiers queued for wave 1. That non-public seed authority remains unchanged and is not the publication base.
  • C1 journals: 179 source works carrying 269 claims yielded 179 admitted canonical works carrying 269 claims, 0 held identities carrying 0 claims, and 0 exclusions carrying 0 claims in this cut.
  • Identity augmentation: 30 exact-DOI C1 links to accepted queue identities are recorded independently of publication admission; 30 are active and 0 are held. A1 declares 11 claim-link occurrences across 8 C1 identities, with only admitted targets activating public links.
  • B1/E1 increments: the completed, content-addressed management-canon and SSRN cache revisions admit only DOI-matched, unique W-IDs that agree with accepted-queue identity evidence wherever an exact-DOI overlap exists. Unresolved and conflicting rows remain held. B1 candidates without a requestable DOI are outside that request universe and remain held unless a separately ruled identity path is supplied.
  • The citation graph: every admitted paper is a node; citation edges appear only when a source supplies the relationship, both endpoints are admitted, and that source work is covered by the partial OpenAlex enrichment pass.
  • The author index: authorships are grouped only by OpenAlex Author ID. Display-name strings never become identity joins; null-A-ID names remain unmerged on their work records, and the derived index remains distinct from the curated researcher roster.
  • Exports: CSV, CSL JSON, a researcher-roster JSON export, and a fixity/count manifest mirror the current public projection.
  • Themes: a small controlled vocabulary, assigned by recorded multi-vote panels, never silently.

Read the full coverage, frozen-lane, rights, and custody disclosures.

Claim discipline

Population words—“all”, “complete”, “comprehensive”—are banned as coverage claims. Waves are additive; corrections happen through an append-only adjudication overlay, never by silently rewriting landed records. Where metadata is unresolved, the record says so.

Who

Built by a human-directed AI research apparatus: research and build seats run by AI engines under human command, with every landing gated. The same machinery that keeps this database current is, itself, an instance of the phenomenon the corpus studies.