06-reference/research

agent shaped job taxonomy grounding

2026-10-21·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
agentify-3000automation-exposuretaxonomyonetcredibility

The decomposition is borrowable from APQC, the exposure score is house-original, and the unit of analysis is the part no published framework can validate

The question

Verbatim from the Notion Research Backlog (page_id 3f9f7d4936d181a9b62acda3272e90bb): "Is 'agent-shaped job' (agentify-3000's core scoring taxonomy — find/judge/do/deliver/admin, scored 1-3 per job per archetype) grounded in any published task-automation-exposure framework (O*NET task-level automation exposure, Autor/Frey-Osborne-style labor literature), or is it a house-original taxonomy with no outside validation?"

The question misstates the taxonomy, and the correction matters. The live scheme in data/scripts/phase2/scheme.md is find / sell / deliver / support / admin, scored 0-3 (not 1-3), per business (2,566 of them), with archetype-level figures being means of business scores. There is no judge axis and no do axis. Everything below scores the real five. The corrected statement is also the first finding: if the queue note drifted off the rubric in two weeks, a published post will too.

What we already know (from the vault)

What the web says

O*NET publishes the task taxonomy and no exposure score

Fetched onetcenter.org/content.html directly. Work-activity content sits under Occupational Requirements → Work Activities ("41 variables / 74,702 ratings / 2,085 detailed work activities / 24,087 DWA-task linkages … This hierarchy includes generalized, intermediate, and detailed work activities"). Tasks is a separate domain — Occupation-Specific Information — with 18,838 statements. The four top-level groupings beneath Work Activities are, verbatim: Information Input ("Where and how are the information and data gained…"), Mental Processes ("What processing, planning, problem-solving, decision-making, and innovating activities are performed…"), Work Output ("What physical activities are performed, what equipment and vehicles are operated/controlled…"), Interacting With Others ("What interactions with other persons or supervisory activities occur…").

Two things follow. First, O*NET itself publishes no automation- or AI-exposure score per task — confirmed against the page text; the one near-miss is Work Context descriptor 4.C.3.b.2 "Degree of Automation — How automated is the job?", a survey-collected incumbent rating of the job as it is today, not a forward-looking per-task estimate. O*NET is the substrate everyone else scores. Second, O*NET's four groupings are cognitive-stage categories, not business-function categories — and they are an uncanny match for the axis set the queue note hallucinated (find→Information Input, judge→Mental Processes, do→Work Output, deliver→Interacting With Others). Those four have outside authority. The five we actually ship do not, on this axis.

Every exposure framework scores tasks and rolls up to occupations

Nobody scores a business

A deliberate negative search returned nothing that assigns an automation-exposure score to a firm or business entity: arXiv metadata queries for firm-level AI/automation exposure returned zero entries, and Anthropic states its index "does not map conversations to specific businesses or firms" (firm behaviour is inferred from first-party API traffic as a proxy, where "three-quarters of interactions classified as automation"). Every verified instrument indexes task → occupation (SML, Frey-Osborne, ALM, Anthropic) or task/project → deliverable (GDPval, RLI). Caveat: this negative rests on arXiv metadata search plus one fetched page; consultancy and vendor indices were not swept, so it is real but not conclusive.

Convergences and contradictions

Mapping verdict, axis by axis

Axis Closest named published construct Verdict
find APQC PCF 3.0 Market and Sell Products and Services, level-2 groups 3.1 Understand markets, customers and capabilities · 3.2 Develop marketing strategy · 3.3 Develop and manage marketing plans Named construct (process framework). No labor-literature analogue.
sell APQC PCF 3.4 Develop sales strategy · 3.5 Develop and manage sales plans Named construct. Our find/sell split is a sub-split inside one PCF category — defensible, and PCF itself draws the same seam at level 2.
deliver APQC PCF 4.0 Manage Supply Chain for Physical Products + 5.0 Deliver Services Defensible recombination: two PCF categories collapsed into one axis. Justified for small businesses that do one or the other, not both.
support APQC PCF 6.0 Manage Customer Service Named construct, effectively 1:1. The strongest single mapping we have.
admin APQC PCF categories 7.0-13.0, i.e. APQC's entire explicitly-labelled "Management and support services" block (human capital, IT, financial resources, assets, risk/compliance, external relationships, business capabilities) Defensible recombination, and the weakest axis: it collapses seven PCF categories into one score. Its 2.0-2.3 means across every archetype are near-certainly an artifact of that aggregation, not a finding.
(missing) APQC PCF 1.0 Develop Vision and Strategy and 2.0 Develop and Manage Products and Services No axis exists. And the rubric knows it: scheme.md's own agent_share instruction says "a SaaS's are build, support, find" — naming a sixth job the taxonomy does not contain. This is an internal inconsistency, not a philosophical gap.
moat field Frey & Osborne's three engineering bottlenecks: Perception and Manipulation · Creative Intelligence · Social Intelligence Near-exact match, and the most citable thing in the whole scheme. Our 0 anchor — "physical hands, licensed signature, in-person trust, taste the market pays for" — is Frey-Osborne's three bottlenecks plus licensure/legal standing, which F&O does not include. That fourth item is a genuine extension, not a restatement.
0-3 scale Eloundou et al.'s E0/E1/E2(+E3) ordinal exposure rubric Granularity defensible; anchors are not operational. Eloundou pins each tier to a measurable threshold (≥50% time reduction at equivalent quality, where "equivalent quality means a third party … would not notice or care about LLM assistance"). Ours are share-of-act descriptions with no threshold, so two raters can disagree without either being wrong.
unit of analysis (business, not occupation) none No outside analogue, qualified. Verified absence across the academic preprint surface; consultancy indices unswept.

Overall verdict: partially grounded. The decomposition is a defensible recombination of a named, current, primary-source-verified process framework (APQC PCF v7.4) — not a house invention, and we should stop treating it as one. The exposure scoring is house-original: no published framework scores business functions, the anchors have no operational threshold, and the 30-business check is author-agreement rather than independent inter-rater reliability. The moat field and the 0 anchor are the parts with the cleanest outside authority, and they are currently the parts we never cite.

Synthesis for RDCO

Can the scoring engine cite outside authority today? Partly, and not where it matters. Right now a data-literate reader of the Fifth Job series gets five axes, a 0-3 scale, and a 2,566-row n, with no named framework behind any of it and no reliability statistic. That reader's first move is the one Eloundou et al. pre-empted on themselves and we have not: who else scored these, and did they agree? Our answer is "the author checked 30 and liked them." Eloundou reports 80.8% raw agreement on direct exposure and a kappa of 0.223 on the identical problem shape — model-labelled exposure scoring against human labels, at scale, by the people who built the model. If the published state of the art admits weak chance-corrected agreement in print, a founder-run study claiming 30/30-within-one-point without a kappa is not reading as more rigorous; it is reading as unaudited. That is the credibility exposure, and it lands squarely on the "agents in production" positioning the phData sales motion is supposed to borrow from ([[2026-10-08-verification-ratio-senior-signoff-throughput]]).

The cheap fix is three citations and one number, and it is almost entirely rewriting, not rescoring. (1) Anchor the five axes to APQC PCF v7.4 explicitly — one sentence per post: "the five jobs are APQC's Process Classification Framework categories 3-6 plus its management-and-support block, collapsed for businesses under 20 people." That converts the taxonomy from invented to derived from a 1992-vintage open standard built with 80+ organizations, which is a materially different sentence to defend. The primary-source read is already in the vault ([[2026-08-17-apqc-pcf-function-motion-mapping]]), so this costs nothing. (2) Restate the 0-3 anchors as Eloundou-style time-reduction bands with a quality clause, and cite arXiv:2303.10130 for the threshold convention. (3) Cite Brynjolfsson/Mitchell/Rock for the finding, not the method — their "few if any occupations have all tasks that are SML … toward the redesign of jobs and reengineering of business processes" is our thesis one level up, and it is the single strongest external validation available to the series. And: run a held-out sample of 50-100 businesses past an independent rater and publish a weighted kappa. One number, generated once, that nobody else in this genre publishes.

The unit of analysis is the asset, and we have been apologizing for the wrong thing. Every verified instrument — ALM, Frey-Osborne, SML, Eloundou, Anthropic Economic Index, GDPval, RLI — indexes tasks and rolls up to occupations, projects, or deliverables. None scores a business. That is not a gap in our rigor; it is the whitespace. The economics literature answers "which jobs are exposed," which is a policy question. agentify-3000 answers "which businesses can be run by one person and an agent fleet," which is an operator question, and it is the one a phData prospect and a Sanity Check reader both actually have. The same pattern RDCO already banked in [[industry-process-reference-models]] applies: borrow the decomposition, own the layer above it. Here the layer above is the firm-level aggregation (agent_share, hours-weighted, deliberately not a mean) and the moat phrase. Those are the original contributions and they should be the headline, with the five axes positioned as borrowed scaffolding rather than as the insight.

Two defects to fix before anything else publishes. First, the missing axis: scheme.md names "build" in its own agent_share instruction while the taxonomy has no build/product-development job, which maps to PCF 2.0 and matters enormously for the saas and consumer_app archetypes that currently score highest. Any post claiming a SaaS is 58% agent-carryable is claiming it about a business whose largest hour sink was never scored. Second, admin collapses seven PCF categories, which almost certainly manufactures its own result — "admin is the universally agent-shaped job, top in 11 of 12 archetypes" is exactly what you would expect from averaging seven heterogeneous functions, and it is currently finding #1 in [[archetypes]]. Neither defect requires rerunning 2,566 businesses to state; both require stating before the series leans on them.

Why this is in the vault

agentify-3000's five-axis score is the engine behind the Fifth Job Ship 30 series ([[agentify-3000-series]]) and the opportunity ranking in opportunities.md, and the series is one of the few surfaces actively building the "agents in production" credibility the founder named as the phData sales lever. This brief is the grounding check that decides whether each post can cite APQC + Eloundou + Brynjolfsson or has to say "a taxonomy I made up," and it flags two scoring defects (no build axis, admin over-aggregation) that the series' two headline findings currently rest on.

Open follow-ups

Related

Sources

Vault

Web (fetched and read; quoted text is from these documents)

Could not verify (flagged, not guessed)

Research budget note: QMD 2 queries + direct grep/read of 8 vault docs; web 3 WebSearch + 4 WebFetch across two verification subagents (one fetch over the template's 3-source cap, spent to verify the SML scale rather than guess it). Paywalled sources were skipped and flagged, never inferred.