Stafford Brief — 2026-09-26

Saturday. Delta layer on the 12:00 UTC scan (2 staged, 9 skipped) — read that brief first; this one does not repeat it.

Three corrections and one find. The scan's headline today was that ADAPTOVATE's numbers force Brandon to narrow his incentive argument. They don't — he already published the narrow version, twice, in the exact words. What the numbers actually break is a smaller and more useful thing: the self-check question at the end of Paper 1. Separately, the scan is quoting the wrong half of its best new study, and it asserted that a certain kind of evidence doesn't exist while this repo has held an example of it since 9 September. The find is a paper the base didn't have, from OpenAI's own chief economist, saying that buying AI is not the same as getting anything from it.

📖 TODAY'S READ

Nobody has checked whether the organizations that adopt AI fastest actually do better — and the company with the data to check it just published a paper that pointedly doesn't.

"How Organizations Use AI: Evidence from ChatGPT" — Chatterji, Holtz, Rakholia, Tambe & Weeratunga; arXiv:2608.12236, 12 August 2026. Four of the five authors work at OpenAI. Not peer reviewed.

They took ChatGPT Enterprise's own account records — who logs in, how much they use it, what they use it for — and joined them to the published financial statements of the public companies that are customers. Over 1,500 organizations, over 17 million messages, running through March 2026. That combination is new here. Everything else in this base that measures which firms adopt AI is built either from job adverts or from surveys where people describe themselves.

The finding in plain terms: the firms buying enterprise AI are the big rich ones, by a wide margin. The typical adopter has about eleven times the revenue of the typical non-adopter ($2.3bn against $210m), seven times the staff (2,934 against 424), and eleven times the R&D spend ($113m against $10m). They also carry more of the softer stuff — accumulated spending on software, sales infrastructure and research — which the authors read as "prior investment in intangible and organizational capabilities."

And here is the line Brandon should take: the authors conclude that "adoption is only the beginning of deployment" and that rapid adoption "should not be equated with immediate productivity transformation." That is the Technology Illusion argument, written by the chief economist of the company that sells the technology, on the company's own usage logs. It is worth more to Brandon than any consultancy survey making the same point, because the people saying it are the ones who lose by saying it.

What it does not do, stated plainly. There is no outcome measure anywhere in the paper — no productivity, no margin, no revenue effect. The authors say their estimates are associations and "should not be interpreted causally," and they list connecting this telemetry to outcomes as future work. So the single best-placed group in the world to answer "do the fast adopters do better?" has the data linked up already and has not run it.

So what it changes: the story thread the scan opened this morning — does prior reorganization predict better outcomes, or only faster adoption? — stops being a hope and becomes a watch on a named group with the data in hand. File: knowledge-base/chatterji-holtz-rakholia-tambe-how-organizations-use-ai-chatgpt-enterprise-2026-08.md.

🔭 THINKER TO WATCH

Aaron "Ronnie" Chatterji — OpenAI's first Chief Economist, on leave from Duke's Fuqua School of Business, and previously Chief Economist at the U.S. Department of Commerce, where he worked on the CHIPS Act.

Why now: he is the most commercially conflicted person who could possibly make Brandon's argument, and he is making it. He controls the richest private dataset on enterprise AI use in existence, and what he has published from it is a description of who buys plus an explicit refusal to claim that buying produces value.

How Brandon differs, and it is not a small gap. Chatterji explains the adoption-to-value lag as co-invention — a natural, temporary delay while firms "discover use cases, invest in complements, and reorganize production." That is the standard economics story about general purpose technologies, and it implies waiting is the right response. Brandon's position is that the lag is not a phase, it is a diagnosis: the organizations that never close it are failing at five specific, nameable things, and time alone does not fix any of them. Chatterji says not yet. Brandon says not without the design work, and most of them will not do the design work. Those predictions diverge, which means one of them will be wrong in public — and that is exactly the kind of disagreement worth staking a paper on.

🏢 CASE IN THE WILD

Nothing today, and padding this would be the wrong call. The scan had no named organizational case either and said so. My own work today was verification and one instrument — neither produces a company story. An honest blank beats a market statistic dressed up as a case.

⚙️ FOUR FORCES CONCLUSION

Capability — and today's evidence points the opposite way from how the scan read it this morning.

The scan concluded that Capability should be understood as an accumulated stock: what the last change left behind sets the price of the next one. Today's paper appears at first to agree — adopters carry more accumulated investment in software, research and organizational infrastructure.

But look at what that accumulated investment actually is here, and the picture changes. It is measured as spending per employee on software, R&D and sales-and-admin. The adopters are eleven times bigger and eleven times richer on research. The simplest explanation for who adopts is not accumulated organizational learning. It is money. A firm with $113m of R&D can afford to try things a firm with $10m cannot, and no organizational capability is required for that.

So the honest Capability conclusion today is a fork, not a finding. Either prior organizational adaptation lowers the cost of the next change (the scan's reading, from the job-adverts study), or the firms that adopt everything early are simply the ones that can pay for everything early (this paper's pattern). The two studies are compatible with both stories and neither can separate them — the job-adverts study has a technique for isolating cause but a weak measure; this one has a strong measure and no technique for isolating cause. Putting them in a draft as mutually reinforcing would be the error. Put them in as two halves of a question neither answers, and a referee has nowhere to go.

💭 OPEN QUESTION

If the organizations that adopt AI fastest are mostly just the richest ones, what is left of "readiness" as a distinct idea — and is Brandon's framework describing a property organizations can build, or one they can only buy?

This is uncomfortable in the right direction and it is not the question the scan asked this morning. The scan worried that readiness might buy absorption speed and absorption speed might buy nothing. The sharper worry is upstream of that: if the measurable correlates of "ready" turn out to be revenue, headcount and R&D budget, then readiness is largely a restatement of size, and a framework built on it tells a mid-market CEO that the answer is to be a large-cap. Brandon's whole practice rests on readiness being something a $200m company can design its way into without a $2bn balance sheet. Nothing in this base currently demonstrates that, and today's paper is the first instrument that makes the money explanation concrete rather than theoretical. The one instrument here that cuts the other way is the 2012 management-practices study, where the effect came from how firms were run rather than how big they were — which is why the next section matters.

🥊 CHALLENGE

The scan is quoting the wrong half of its best new study, and the half it quotes cannot mean what it is being asked to mean.

The Schubert job-adverts paper reports two numbers: an effect of 0.4 when comparing different firms, and a larger effect of 0.7 when comparing different job types inside the same firm. The scan's case-file entry calls the bigger, within-company number "the structurally important result" and treats it as the strongest new support for the claim that organizational readiness drives technology uptake. I verified the wording against the paper's own abstract today: "0.4 percentage points across firms and 0.7 percentage points across occupations within firms."

Here is the problem. When you compare job types inside one company, everything that is true of the company as a whole — its culture, its leadership, its operating model, its readiness — is identical on both sides of the comparison and therefore cancels out. You cannot detect a company-level property by looking only at differences within a single company. The scan has taken as its best evidence for organizational readiness the one number from which organizational readiness has been mathematically removed.

The mundane rival explanation, which that number cannot rule out: the jobs that went remote in 2021 and the jobs that use generative AI in 2024 are the same jobs — desk work, screens, documents. No organizational learning needed. That overlap sits at the job-type level, so a company-level research design does not touch it.

What would settle it is cheap: read the paper's method section for whether its statistical technique varies by job type at all, and whether it controls for how digital each occupation is. Nobody has done that — the scan quoted the abstract and noted it had not read the error margins. Until someone does, the number Brandon can cite for an organizational claim is the smaller one, 0.4. Logged in full at knowledge/thesis-challenges.md.

🧱 ASSET SUGGESTION

Paper 1's incentive self-check has been broken by AI — not by being wrong, but by being passed.

The question at the end of that section is: "If the transformation succeeds but a leader's metrics do not improve, will that leader still treat the work as a priority?" It worked when few executives had personal skin in a given transformation. ADAPTOVATE's survey says 83% of executives now report that their own job security depends on scaling AI — and 19% of their companies have a plan. Five in six pass the check. A diagnostic question that nearly everyone passes has stopped diagnosing anything.

The replacement is already in Brandon's own prose, three paragraphs above it: "It is whether the system makes it rational for them to prioritize it when tradeoffs appear." The self-check asks about one leader's metrics; the prose asks about the system's tradeoff. Fix the check to match the paragraph — something closer to when this transformation and a leader's quarterly number collide in the same week, which one do the calendar and the budget actually protect?

Form: a one-sentence revision to Paper 1, not an essay. Full spec, including the one thing to retrieve first, at knowledge/asset-suggestions.md asset 15.

And the framing correction that goes with it, because it matters for how Brandon reads this morning's brief. The scan routed ADAPTOVATE to him as "Paper 2's Incentive Fragmentation argument needs narrowing before it is published, not after." Paper 2 was published on 14 May and it already says, at line 48: "Commitment is not endorsement. It is the designed priority that holds when tradeoffs appear." Paper 1 says: "The issue is not whether leaders verbally support the change. It is whether the system makes it rational for them to prioritize it when tradeoffs appear." The loose version of the argument was never his — it had drifted into the evidence file. So ADAPTOVATE is not a correction to Brandon. It is the first hard number behind a distinction he drew four months ago from a single story about a CISO, and he should use it that way: I argued this from one case; here is n=300.

🎓 SPARRING NOTE

Not issued, and the cap is the reason. Three notes ran in the week beginning Monday 21 September — 09-22, 09-23 and 09-24 — against a stated cap of two, a breach the 09-25 run recorded against itself. Six notes now stand unanswered. The standing recommendation holds: pick 09-17 and answer it; a prompt backlog is not a development plan.

SOURCES & THREADS

Threads: none due, and I checked independently rather than taking the scan's word. The nearest live date is 2026-09-30. One file defect found and worth fixing before it bites: story-threads.md line 178 carries an orphaned Next check: 2026-09-16 line stranded inside the log of the verification-cost thread, left behind by an earlier date change. The thread's real date is on line 139 (2026-10-11) and its log runs to 09-25, so nothing was missed — but any run that scans the file for due dates, as I did, reads that thread as ten days overdue. It is the only instance in the file. Fix in Routing.

Exploratory query — null on target, and one find the base didn't hold. Target: the gap the scan named this morning as its highest-value one — anything that measures organizational outcomes against a pre-existing organizational condition, with AI as the technology. I composed it against a sharper specification than the scan could, because this repo holds the design template (see below). The specific cell is genuinely empty. Three of the top results were things this base had already correctly disposed of — including Zeng/Wang/Sun's "Domain AI Readiness," assessed and retired on 2026-08-17, resurfacing for the third time on a third wording exactly as the candidates ledger predicted. That is the coverage confirmation the null needed. The find is today's read. Logged at knowledge/query-log.md.

Candidates: none added. Chatterji and Tambe are field-map proposals, not new outlets; arXiv econ.GN is long since standing.

The thing worth Brandon's attention in this section, though, is a system problem rather than a search result. The scan opened today by pre-registering an excellent rule for itself — check what the base already holds before searching the web — after catching itself asserting a false absence yesterday. In the same brief it asserted another one: that no instrument puts an outcome against a prior-reorganization measure. This repo has held the canonical example since 9 September — Bloom, Sadun & Van Reenen's 2012 study in the American Economic Review, where firms with strong people-management practices got roughly twice the productivity out of the same computer spending and firms without got essentially none. The scan's sentence survives only on two words: BSVR scored management at the end of its window rather than beforehand, and its technology is PCs, not AI.

This is an architecture problem, not carelessness, and it is the second one this month. The scan cannot read stafford-research. Its new rule is unsatisfiable for the half of the base Stafford owns, because the 2026-08-30 split left Stafford's exploratory finds in a repo the scan never opens. Starting today, every Stafford brief carries a BASE-HOLDS block in Routing so the scan can satisfy its own rule from a file it does read. The durable fix is a PROPOSED-UPDATE below.

Routing

🔒 BASE-HOLDS — instruments in `stafford-research` bearing on the scan's currently-open gaps (new standing block, 2026-09-26)

For the next scan run: these exist, they are in the other repo, and you cannot see them.

Gap: "no instrument puts an outcome on the left-hand side of a prior-reorganization measure" (asserted in today's scan brief and in the new thread's criterion (a)). → Bloom, Sadun & Van Reenen, "Americans Do I.T. Better," AER 102(1):167–201. knowledge-base/bloom-sadun-van-reenen-americans-do-it-better-2012-02.md. Blind management-practice scores joined to statutory accounts; management × IT interaction 0.1451 (0.0331); the level effect of management is null, so the result lives entirely in the interaction. Two disqualifiers keep the gap open and both should be stated whenever it is cited: management was scored once in summer 2006, at the end of both panels, so it is contemporaneous rather than prior; and the technology is PCs per worker, not AI. It is the design template, not the finding. The gap is correctly described as this design, run with a predetermined organizational measure and AI as the technology.
Gap: the ESCALATED Résumé-Panel Consensus pattern's shared-genre complaint. → Chatterji et al. (today, chatterji-...-chatgpt-enterprise-2026-08.md) is the first firm-level AI adoption instrument in either repo that is not postings or résumé inference — vendor usage telemetry joined to public financial statements — and the first whose window runs into 2026.
Gap: Claim 3's Incentive Fragmentation evidence under AI. → Characterized, not merely open, since 2026-09-11: the multitask principal-agent literature is mature and theoretical, and nobody has instrumented it for AI. See knowledge/query-log.md, 2026-09-11.

PROPOSED-UPDATE — `knowledge/story-threads.md` (research-evidence), line 178

Delete the orphaned line beginning - **Next check:** brought forward to **2026-09-16** on the strength of the Faros result…, which sits inside the Log block of "Does anyone design for verification cost…" and is superseded by the authoritative header line 139 (2026-10-11). Retain its criteria text by appending it to the line-139 entry if it is not already carried there. Reason: any run that greps the file for due dates reads this thread as ten days overdue. Only instance in the file.

PROPOSED-UPDATE — `knowledge/story-threads.md` (research-evidence), thread "The ladder's second half", Log

Append: 2026-09-26 (Stafford): criterion (a) now has a named party holding the data. Chatterji, Holtz, Rakholia, Tambe & Weeratunga, "How Organizations Use AI: Evidence from ChatGPT" (arXiv:2608.12236, 12 Aug 2026; four of five authors OpenAI) have already linked ChatGPT Enterprise usage telemetry to public-company financial statements — 1,500+ organizations, 17M+ messages, through March 2026 — and report only descriptive facts, explicitly non-causal, naming the outcome linkage as "future work." The second half of the ladder is therefore not blocked on data assembly; it is blocked on someone choosing to run it, and the best-placed group has the panel built. Check OpenAI's economic-research output as the primary channel for criterion (a), ahead of general search. Note also that this instrument partly satisfies criterion (c) — a non-postings replication of the antecedent — though associationally rather than causally.

PROPOSED-UPDATE — `knowledge/emerging-patterns.md` (research-evidence), Résumé-Panel Consensus (`ESCALATED`)

Append: 2026-09-26 (Stafford): the shared-genre complaint has its first genuine exception at the firm level. Chatterji et al. (arXiv:2608.12236) measure enterprise AI adoption from vendor usage telemetry joined to statutory financial statements — no postings, no résumés, no self-report — over 1,500 organizations through March 2026, and reach a conclusion compatible with Schubert's antecedent (pre-existing organizational and intangible capital predicts who adopts). Two cautions against reading this as independent confirmation. (1) It is explicitly non-causal and cannot separate accumulated organizational capability from sheer firm size and spending power; adopters are ~11x median revenue and ~11x median R&D. (2) It is the second vendor-telemetry instrument here after Yu/Chen/Hu/Suri/Counts (M365 Copilot) — two vendors measuring their own products with their own logs share a structural interest even without sharing data, so treat them as one genre with two members. The pattern's staleness complaint is now weak on two counts (Schubert, Chatterji); the shared-instrument complaint stands, narrowed to postings specifically.

PROPOSED-UPDATE — `knowledge/paper-2-case-file.md` (research-evidence), today's Claim 5 entry

Two amendments to the entry filed this morning. (1) Demote the within-firm coefficient. The entry calls Schubert's 0.7pp within-firm estimate "the structurally important result" for Claim 5. A within-firm estimate differences out firm-level properties, including organizational readiness — so it cannot evidence a firm-level readiness claim, and the occupational confound (remote-capable work and genAI-applicable work are the same digital work) is unaddressed by a firm-level instrument. The citable coefficient for Claim 5 is 0.4pp across firms until someone reads the specification for occupation-level instrument variation and occupational-digital-intensity controls. Full statement in stafford-research/knowledge/thesis-challenges.md (2026-09-26). (2) Add Chatterji et al. as the cross-genre corroboration and the rival reading — see BASE-HOLDS above; it supports the antecedent from a different data genre while making "money, not organizational learning" a live alternative the Schubert entry does not currently concede.

PROPOSED-UPDATE — `knowledge/field-map.md` (research-evidence)

New entry. Aaron "Ronnie" Chatterji — Chief Economist, OpenAI (the company's first, appointed Oct 2024); on leave from Duke Fuqua (Mark Burgess & Lisa Benson-Burgess Distinguished Professor of Business and Public Policy); previously Chief Economist, U.S. Department of Commerce (CHIPS Act). Why tracked: controls the richest private dataset on enterprise AI use and publishes the adoption-is-not-value conclusion from inside the seller. Brandon's differentiation: Chatterji frames the adoption-to-value lag as co-invention — a temporary, self-resolving phase of general purpose technology diffusion; Brandon frames it as a diagnosis — five nameable failures that time does not fix. Divergent predictions, publicly testable. Also add Prasanna Tambe (Wharton; OpenAI paid contractor) as a secondary. Source: stafford-research/knowledge-base/chatterji-holtz-rakholia-tambe-how-organizations-use-ai-chatgpt-enterprise-2026-08.md.

PROPOSED-UPDATE — division of labor (needs Brandon, or the scan's next run)

The scan's new "check the base before you search" rule cannot be satisfied as written. It pre-registered that rule this morning and broke it in the same brief, because the evidence it needed sits in stafford-research, which it does not read. Cheapest durable fix: the scan's run reads stafford-research/knowledge/query-log.md at bootstrap — one file, already a compact ledger of every Stafford exploratory target and its yield, which is exactly the index the rule requires. Failing that, Stafford commits a pointer digest into a path the scan does read. Until either is in place, the BASE-HOLDS block above is the manual workaround and it depends on Stafford correctly guessing which gaps the scan will assert.

STAGE-CANDIDATE — `chatterji-...-chatgpt-enterprise-2026-08.md`

python3 build/add-research-json.py --json '{
  "title": "How Organizations Use AI: Evidence from ChatGPT",
  "publisher": "arXiv (OpenAI / Columbia Business School / Wharton)",
  "url": "https://arxiv.org/abs/2608.12236",
  "published_date": "2026-08-12",
  "credibility": "moderate",
  "instrument": "ChatGPT Enterprise internal account records linked to message-level usage telemetry, administratively-inferred job titles and seniority, LLM task classification, and public-company financial statements. Descriptive decomposition plus linear-probability regressions of adoption on LAGGED firm characteristics and employee-normalized stocks of SG&A, R&D and capitalized software. Authors state estimates are conditional associations and must not be interpreted causally; no outcome variable in the paper. Window: usage through March 2026; token series June 2025 - March 2026; adoption cohorts from January 2024. Vendor research: 4 of 5 authors OpenAI; Holtz and Tambe paid OpenAI contractors. Not peer reviewed.",
  "breakpoints": [
    {"name": "technology_illusion", "rationale": "Accumulated intangible and organizational capital (SG&A, R&D and capitalized software stocks per employee) predicts which firms deploy enterprise AI while the paper measures no value variable of any kind, and the authors conclude that adoption should not be equated with immediate productivity transformation."},
    {"name": "momentum_mirage", "rationale": "Aggregate output tokens grew roughly sevenfold between June 2025 and March 2026 - the strongest available scaling metric - in the same paper whose authors decline to infer transformation from it and report that firms are still actively learning how to integrate AI into organizational workflows."}
  ],
  "forces": [
    {"name": "Capability", "rationale": "Adoption is predicted by pre-existing employee-normalized stocks of SG&A, R&D and capitalized software, the authors explicit proxies for accumulated investment in organizational capabilities, measured on financial statements rather than inferred from hiring text."}
  ],
  "key_findings": [
    "Median adopter revenue $2,275.1M vs $209.6M for non-adopters; median employment 2,934 vs 424; median market value $4,997.2M vs $316.4M; median R&D $113.1M vs $9.9M",
    "A one-log-point increase in lagged revenue per employee is associated with roughly 0.4 to 0.9 percentage points higher adoption probability - explicitly non-causal",
    "Aggregate ChatGPT Enterprise output tokens grew roughly sevenfold June 2025 to March 2026; roughly fourfold within the fixed cohort already adopting by June 2025, so about half of growth was intensification inside existing customers",
    "Weekly active users after six months by seniority: managers and directors ~24%, ICs and professionals 15%, senior ICs and principals 14%, executives ~10%, early-career and trainees 7% - while early-career workers send many more messages each, so extensive and intensive margins point opposite ways",
    "Authors conclude adoption is only the beginning of deployment and should not be equated with immediate productivity transformation; outcome linkage named as future work"
  ],
  "sample_size": 1500
}'

Note for whoever stages this: sample_size is organizations at the six-month worker-level horizon (paper says "over 1,500"); the message count is 17M+. Confidence filed moderate — vendor research on the vendor's own product, not peer reviewed.

For Brandon

Strategy. The one decision on the table is the Claim 5 split recommended on 09-24 — 5a as a stated premise, 5b as the empirical claim about captured gain. Today is the fifth wording pressure in twenty-one days, and running this morning's "absorption versus benefit" worry against the split shows it does not even arise: 5b is about captured gain, Schubert measures uptake, the two never touch. The question the scan escalated to you is a symptom of the unsplit claim. Implementing the split retires it, and retires the next one too.

Strategy. The Open Question above is the one I cannot answer from the base and would like your read on: if the measurable correlates of organizational readiness keep turning out to be size and budget, is readiness something a mid-market company can design its way into? Your entire practice assumes yes. The 2012 management-practices study is currently the only instrument here that supports it, and it is fourteen years old and about PCs.

Personal. Two process items, both cheap. (1) Six sparring notes stand unanswered; pick 09-17 and the rest should be retired rather than carried. (2) The scan-cannot-read-Stafford gap above has now produced a false absence in a brief warning against false absences — worth one decision from you on whether the scan's bootstrap adds one file read.

Ship: see first line of the run's final message.