The Enterprise AI Adoption Gradient

Sources#

• How Organizations Use AI: Evidence from ChatGPT

Summary#

How Organizations Use AI: Evidence from ChatGPT (Chatterji, Holtz, Rakholia, Tambe & Weeratunga, arXiv 2608.12236, 2026-08-12, 69pp) is the first study in this corpus to read firm-level AI adoption off a vendor's own administrative account records rather than off occupational exposure scores, payment rails, or executive surveys — and then to link those records to Compustat financial statements. Its subject is not what AI can do but which firms bought it, how deeply they use it, and who inside them is using it.

Four documented facts, in the paper's own order:

1. Usage is growing on both margins at once. Aggregate output tokens across ChatGPT Enterprise customers grew roughly sevenfold between June 2025 and March 2026, and roughly fourfold within the fixed cohort of firms that had already adopted by June 2025 — so "about half of the growth in token consumption over this period occurred within already-adopting firms." Every adoption cohort accelerated simultaneously in early 2026, which the authors read as something happening to the product or the market rather than the normal post-adoption ramp.
2. Among U.S. public companies, adopters are the big, valuable, intangible-heavy end of the distribution — and the association strengthens with pre-existing organizational capital.
3. Inside adopting firms, use spans every job function and seniority level, but intensity varies systematically, with a negative seniority gradient.
4. Use spans a wide task range rather than concentrating in one workflow — the paper's evidence for generative AI as a general purpose technology for knowledge work.

The synthesis the authors put on it, and the sentence this page exists to carry: "adoption is only the beginning of deployment."

Evidence note. empirical, and correctly so — these are measured message, token and active-user counts, not self-report, which is the paper's stated motivation ("self-reported usage is typically less detailed and may suffer from imperfect recall or reporting biases"). Three constraints bind every number below. (1) Conflict of interest is total, not partial: three authors are at OpenAI, and the two academics — Holtz (Columbia) and Tambe (Wharton) — "contributed to this work in their capacity as paid contractors for OpenAI." There is no independent author on the paper and no external replication. (2) The sample is buyers, not a population. Everything here describes organizations that purchased a centrally administered ChatGPT Enterprise workspace; usage through personal accounts, the API, other subscription tiers, and every other vendor is invisible by construction. (3) Nothing is causal. The paper says so itself of its central regressions: the estimates "describe conditional associations… and should not be interpreted causally." No outcome is measured — not output, not productivity, not quality, not organizational change.

What is actually being counted#

The unit matters, because this page's numbers are not interchangeable with the survey-based adoption figures the wiki carries elsewhere. Four distinct samples, all descending from one organization-week panel of workspaces adopted between 2024-01-01 and 2026-03-31:

Sample	Size	What it supports
Aggregate usage panel	organization-weeks; zero-usage active weeks retained	the growth decomposition (Fact 1)
Worker-characteristics sample	1,764 organizations, 17,446,551 messages	composition and intensity by role (Fact 3)
Task-classification subsample	973 organizations, 8,696,657 classified messages	task mix (Fact 4)
Compustat-linked public-company panel	521 ticker-years / 417 tickers (509 / 410 with positive message volume), against 11,784 non-adopter tickers	who adopts (Fact 2)

Usage is measured three ways and they are not the same quantity: messages sent, weekly active users (WAU), and generated output tokens — with per-employee variants built by assigning each usage week to a Compustat fiscal year and dividing by emp. Role analyses are pinned to a common clock: week 26 (six months) after each organization's adoption date.

The control group is contaminated by construction, and this is the most important methodological fact on the page. To avoid disclosing commercially sensitive information about its enterprise business, OpenAI "draw[s] a random sample from the resulting set of ChatGPT Enterprise accounts with public-company identifiers." Non-adopters are then defined as public firms with no ticker-bridge match. Two consequences follow that the paper does not spell out: (a) genuine adopters that were not drawn into the sample sit in the 11,784-firm "non-adopter" group, so every adopter-versus-non-adopter contrast below is attenuated toward zero and should be read as a lower bound; and (b) no adoption rate can be computed from these counts — the 417: 11,784 ratio is a sampling artifact, not a share of US public companies using ChatGPT Enterprise. The paper separately notes a second attenuation in the same direction: firms using other vendors' enterprise LLM products are also classified as non-adopters.

Fact 2: who adopts — scale first, intangibles second#

Unadjusted medians, 2024 Compustat characteristics (paper's §4.2 prose; see the parse warning below before reusing these):

Median, FY2024	Adopters	Non-adopters
Revenue	$2,275.1M	$209.6M
Total assets	$4,394.2M	$667.6M
Employment	2,934 workers	424 workers
Net PP&E	$271.2M	$43.7M
Market value	$4,997.2M	$316.4M
R&D expense	$113.1M	$9.9M

An order of magnitude on every line. The authors' own summary: "early ChatGPT Enterprise adopters are not representative of the average public firm."

Conditional on the rest, the gradient is scale and revenue productivity, not physical capital. A one-log-point increase in lagged revenue per employee is associated with +0.4 to +0.9 percentage points of adoption probability across specifications; lagged log employment is positive and significant in all of them; and PP&E per employee is generally negatively associated with adoption once scale is controlled. The pattern survives dropping the information sector, so it is not a tech-industry artifact.

The concentration is at the very top of the size distribution. A one-log-point increase in lagged revenue is worth +1.1pp; being in the top revenue quartile is +6.9pp and the top 5% is +9.8pp. Measured within industry-year cells (the Autor et al. 2020 superstar-firm construction) the gradient is steeper still: +7.2pp for the top quartile and +11.3pp for the top 5%. The authors flag the obvious limit — the entire panel is US public companies, already large relative to the business population, so this says nothing about how adoption varies among startups, mid-market, or small private firms.

The complement stocks, ranked#

The paper's most novel move, and the reason Organizational Complements to AI needed this source: it tests whether pre-existing intangible capital predicts adoption, using stocks measured in fiscal year 2021 — three to four years before the FY2024–25 adoption window — so the measure cannot be an artifact of adopting. SG&A expense is cumulated at 20% annual depreciation, R&D at 15%, and capitalized software is taken directly from Compustat capsft; each is normalized per employee and entered as log(1 + stock/emp).

FY2021 complement stock per employee	Full sample	Excluding tech / high-R&D sectors
SG&A stock	0.020*** (0.005), N=5,943	0.010 (0.007), N=3,247 — not significant
Capitalized software	0.008** (0.003), N=1,076	0.006 (0.006), N=477 — not significant
R&D stock	0.004*** (0.001), N=7,117	0.003*** (0.001), N=4,029

(Table 4, reconciled cell-by-cell against pdftotext -layout -f 37 — the docling parse of this table is welded and shifted and must not be read directly.)

Read it as a ranking rather than as three separate results. SG&A stock — the crudest available proxy for accumulated organizational capital — is the strongest and most robust predictor, five times the R&D coefficient. Capitalized software, the literal "prior software investment" measure, is significant in the full sample only. And the ordering is fragile where it matters most: outside technology and high-R&D sectors, only R&D survives significance, with SG&A and software both losing it on roughly halved samples. So the headline is real but its generality outside tech is unestablished.

The inversion: larger firms adopt more and use less#

The paper's sharpest internal tension, and the finding most likely to be misread. The extensive and intensive margins run in opposite directions on firm size. Among adopters (Table 2, reconciled against pdftotext -layout -f 35), lagged log employment enters:

Usage-intensity outcome	Coefficient on L. log employment
Messages per active week per employee	−0.266*** (0.019)
Weekly active users per employee	−0.032*** (0.003)
Output tokens per employee	−0.667*** (0.048)
Messages per weekly active user	−0.002 (0.008) — indistinguishable from zero

The three per-employee measures fall sharply with firm size; messages per active user does not move at all. The decomposition is therefore clean: a big adopter's lower per-employee usage is entirely a breadth deficit — a smaller fraction of its workforce is active — and not a lower intensity among the people who are active. A large firm buys the workspace and a smaller share of its people ever shows up; those who do behave like everyone else. That is the complements thesis stated as an arithmetic identity, and it is the mechanism behind "adoption is only the beginning of deployment."

Revenue per employee runs the other way on the intensive margin but weakly: the point estimates on messages-per-active-week-per-employee and WAU-per-employee are positive and consistent with broader diffusion across employees, but neither is statistically significant, and the paper declines to conclude from them. Descriptively (Figure 4), high-intensity adopters — split at the median of weekly output tokens per employee — sit to the right of both low-intensity adopters and non-adopters on revenue per employee and market value per employee, and the rightward shift survives adjusting for industry and size class, most visibly for market value.

Fact 3: broad inside the firm, and steeply graded by seniority#

Six months after adoption, in the firm-level average (each organization weighted equally):

• By function: engineering and technical practitioners ≈11% of weekly active users; executives, founders and partners ≈9%; finance/accounting and marketing/communications ≈5% each; sales and account management ≈4%. Use is not confined to technical workers.
• By seniority: managers and directors ≈24%; ICs and professionals ≈15%; senior ICs and principals ≈14%; executives ≈10%; early-career workers and trainees ≈7%.

The caveat the prose never states, which Figure 5 makes unavoidable: the single largest category in both panels is "Other / unknown." Reading the chart against its gridlines, unclassifiable job titles are ≈37.5% of weekly active users in the job-title panel (≈43% population-weighted) and ≈30% in the seniority panel (≈34% population-weighted). Every percentage above is a share of a denominator in which the biggest block is workers the classifier could not place. The paper's §4.3 limitations are honest about the adjacent problem — these are shares of observed active users, not role-specific adoption rates, because the denominator of all employees by role is unobserved — but the size of the unclassified bucket is visible only in the figure.

Intensity inverts composition. The groups with the largest share of active users are not the heaviest users. Analysts and marketing/communications workers send more weekly messages than the average active user in their own firm, despite being small shares of the user base; executives, founders and partners send fewer. On seniority the gradient is monotone and negative: the paper states that early-career workers and trainees send "roughly eight to nine more weekly messages than the average active user within the same firm," while managers, directors and executives send fewer. (Figure 6 Panel B's plotted point estimates read closer to +6.5 population-centered and +7.4 with firm fixed effects; see the parse note. The direction and the ranking are not in doubt — the exact magnitude is.)

This lands directly against the early-career labor-market literature: the workers whose entry-level prospects are the subject of the canaries-in-the-coal-mine finding are, conditional on being active, the most intensive users of the tool. The paper flags the juxtaposition and adds the right hedge — "message volume should be interpreted as a measure of usage intensity rather than as a complete measure of economic importance."

Fact 4: breadth of reach, concentration of volume#

Messages are assigned to one of 60 task categories by a classifier available from 2025-10-30. The two measures — what share of active users touch a task at least once a week, and what share of messages a task accounts for — separate cleanly, which is the finding.

Task (Figure 7)	Share of WAU using it ≥once/week	Share of weekly messages
Documentation and technical writing	56.3%	18.3%
Technical digital and electronic	49.6%	12.7%
Interpersonal message	41.1%	11.9%
Topic overview	38.9%	4.4%
Other facts and figures	28.7%	—
Nontechnical professional or academic	27.0%	3.3%
Business and market research	21.4%	3.0%
Sales and marketing	19.4%	5.0%
Plan, design and create	18.7%	2.9%
Legal and regulatory	17.2%	2.7%
Data analysis	15.5%	—
Financial and tax	14.9%	2.5%
Debug and fix	—	3.3%
Translation	—	2.6%
All other task classifications	—	27.3%

Three readings the table supports. Reach and depth are separate margins: topic overviews reach 38.9% of active users and account for 4.4% of messages; legal, financial and research tasks are widespread and comparatively message-light. No single application dominates — the largest category is 18.3% of messages. And the long tail is the largest single block: pooled "all other task classifications" is 27.3% of messages, which the authors correctly treat as evidence in itself that enterprise use spreads across activities that do not fit the top categories.

Task mix varies by industry, role and seniority, but mostly on the extensive margin: financial and tax tasks are far more prevalent in finance and insurance, sales and marketing more prevalent in arts/entertainment, information and retail than in manufacturing — while the message-weighted distributions look similar across industries, concentrated in the same core of documentation, technical work and communication. By role the same shape recurs: engineers over-index on technical work and debugging, finance staff on financial and tax, sales and marketing staff on sales and marketing, and the role-specific tasks "do not overpower the small set of core tasks that are performed by all roles." By seniority, executives are relatively more represented in topic overviews, facts and figures, legal and financial tasks — consumption of synthesized information — while early-career workers and ICs concentrate in production-oriented categories.

This is conversational, not agentic, usage. The token series pools ChatGPT and Codex, but Appendix Figure A1 shows the growth is "overwhelmingly generated by ChatGPT and related non-agentic AI tools" during the sample window. Read it as the ChatGPT-side baseline against which Conversation-to-Delegation Shift measures the Codex-side move.

Where this sits among the vault's adoption instruments#

The wiki now carries five instrument families pointed at "how much AI are firms using," and they disagree by 4x on levels (Telemetry vs. Survey Measurement, AI and Market Power). This one is a sixth and structurally different: not a rate at all, but a within-buyer depth measure.

• Against Firm AI-Spend Intensity and Headcount Growth's Ramp × Revelio panel, this observes the same object — a firm's revealed paid adoption — through the vendor's ledger rather than the customer's card. Ramp sees any AI vendor spend across 21,559 firms and can compute an adoption rate; OpenAI sees one product across a random sample of its own public-company customers and, as established above, cannot. Both find adoption gated on something other than firm-size alone (Ramp's PEPM intensity tercile; this paper's FY2021 complement stocks), which is the most substantive agreement between them.
• Against AI and Market Power's Portuguese firm-level evidence, there is an apparent contradiction worth keeping visible. OECD finds the most GenAI-exposed firms are small (median 6 employees), young, and not market leaders; this paper finds the ChatGPT Enterprise adopters are large public incumbents. The two are reconcilable and the reconciliation is informative: exposure is an occupational-composition proxy that ignores purchasing, while enterprise adoption is a procurement event that only sizeable organizations run — and the US-public-company frame here structurally excludes the small firms where OECD's exposure peaks. The intensive margin above is where they actually touch, and there they agree: per-employee usage falls with employment, so within the adopter population the smaller firms are the intensive ones.
• Against Anthropic Economic Index, this is the closest sibling measurement in the corpus and the differences are the point. Both are first-party classifier-mediated telemetry from a frontier lab; the AEI's unit is the conversation, sampled across consumer, Cowork, Claude Code and API surfaces and mapped onto O*NET tasks, while this paper's unit is the organization-week inside a paid administrative workspace, mapped onto NAICS industries, job-title classes and a proprietary 60-category task taxonomy. The AEI can say what kind of work the economy does with AI; it cannot link a conversation to a firm's balance sheet. This paper can — and pays for it by seeing only enterprise buyers and only one product. The two samples are close to disjoint, so their numbers should never be pooled or averaged.
• Against Task Saturation: Broad but Shallow AI Diffusion's broad-but-shallow reading of diffusion, Figure 7 is the enterprise-interior version of the same shape: wide reach on the extensive margin, concentrated volume on the intensive one.

What it cannot say#

Collected in one place because nearly every claim above needs at least one of them:

• No causality. Adoption regressions are conditional associations on lagged characteristics, stated as such.
• No outcomes. Output, productivity, quality, time saved, organizational change: none observed. The paper's own closing call is for future work connecting enterprise telemetry to output measures.
• No population. ChatGPT Enterprise buyers only; other vendors, the API, Business tier and personal accounts are invisible; the public-company arm is US-only and drawn as a random sample.
• No role denominators. Composition shares describe observed active users, not the fraction of each role that adopts.
• Classifier-mediated throughout. Job titles are normalized and classified automatically; tasks are assigned by a classifier "evaluated against an internal benchmark of human- and model-labelled ChatGPT conversations" whose accuracy is not published — the same unquantified error bar Usage-Telemetry Classifier Validation tracks across this genre.
• Job titles are a one-time snapshot and coverage is incomplete, with unclassifiable users retained in denominators.

Connections#

• Organizational Complements to AI — the thesis this source supplies its first pre-dated, firm-level measurement for: complement stocks laid down in FY2021 predict FY2024–25 adoption, with SG&A (organizational capital) five times the R&D coefficient. The conclusion restates the page's electrification argument almost verbatim — "General purpose technologies rarely generate immediate, economy-wide gains; their impact unfolds through a slower process of co-invention"
• Conversation-to-Delegation Shift — the external organizational curve that page's frontier-preview question asks about, measured in tokens rather than survey phase: 7x aggregate and 4x within-cohort growth over nine months, accelerating across all cohorts at once. Also its baseline — this is the ChatGPT (non-agentic) side of the same vendor's token series
• Task Crossover — the population that page's method explicitly could not reach: crossover is measured on Business-account users' individual messages and is not generalizable to Enterprise. This is Enterprise, with role-resolved task composition — role-specific specialization sitting on top of a shared core — but no O*NET mapping and therefore no crossover ratio
• Firm AI-Spend Intensity and Headcount Growth — the payment-rail instrument aimed at the same object from the buyer's side; the complement to its adoption-rate ledger, and the source that cannot join it
• AI and Market Power — the apparent size contradiction (small exposed firms vs large adopting ones) and its reconciliation; and the mechanism behind that page's divergence worry, stated here as "diffusion may initially reinforce existing firm heterogeneity"
• Pilot-to-Production Gap — "adoption is only the beginning of deployment" is that page's thesis with a token series behind it; the breadth deficit at large firms is the gap measured as a coefficient rather than asserted from consulting practice
• Telemetry vs. Survey Measurement — the paper's own framing argument for its instrument, and a new entry in the aperture ledger: administrative vendor records see depth inside buyers with a precision no survey reaches, and cannot see non-adoption, other vendors, or shadow use at all
• Task Saturation: Broad but Shallow AI Diffusion — the same broad-reach/narrow-volume split, measured inside paying enterprises instead of across occupations
• Usage-Telemetry Classifier Validation — the unpublished-accuracy problem applies to both classifier layers here, the job-title one and the 60-category task one
• OpenAI — the vendor whose administrative records these are, and the employer or paymaster of all five authors
• Anthropic Economic Index — the rival program measuring the same phenomenon on a near-disjoint sample with a different unit of analysis
• Codex — present in the token series and deliberately backgrounded; the agentic share is small enough in-window that the authors defer it to the companion Codex paper
• Erik Brynjolfsson — the intangible-complements literature (Brynjolfsson, Rock & Syverson 2021) the complement-stock design operationalizes, and the early-career employment work the seniority gradient is set against
• AI Adoption in Scientific Work — the population this admin-record study excludes by construction, and that study excludes in reverse: Google's ATLAS telemetry drops enterprise contracts and paid API, while ChatGPT Enterprise records see only paying organizations — the two instruments partition the adopting population rather than overlap on it

Open Questions#

• The FY2021 complement stocks predict adoption; nothing here shows they predict value. Does a high SG&A or capitalized-software stock at adoption forecast deeper deployment (WAU per employee, task breadth) or better firm outcomes two years later — or does it only forecast who signs the contract? The same Compustat linkage run forward on the 2024 cohorts would settle it.
• Larger adopters show lower per-employee usage entirely through breadth (WAU/emp −0.032***) with per-active-user intensity flat (−0.002, n.s.). Is that a rollout-speed artifact that closes as seat deployment catches up with headcount, or a durable ceiling on how much of a large workforce a centrally administered workspace ever reaches? Falsifiable by re-measuring the same cohorts at week 52 and week 104 against week 26.
• "Other / unknown" is the largest single category of weekly active users (≈37.5% of the firm-level job-title panel). Are unclassifiable-title users a random slice of the workforce, or systematically different — contractors, frontline staff, non-English titles — in a way that biases the functional and seniority composition estimates? The classifier validation in Appendix C reports top titles per class but never characterizes the residual.

Sources#

• How Organizations Use AI: Evidence from ChatGPT — Aaron Chatterji, David Holtz, Neel Rakholia, Prasanna Tambe & Gawesha Weeratunga, How Organizations Use AI: Evidence from ChatGPT, arXiv 2608.12236 (2026-08-12, 69pp, empirical). §3.1–3.3 (the four samples, the Compustat linkage, the random-sample draw and the non-adopter definition — footnotes 9–12); §4.1 (growth decomposition); §4.2–4.2.4 (adopter characteristics, Tables 1–4); §4.3–4.3.2 (composition and intensity by role and seniority, Figures 5–6); §4.4–4.4.3 (task composition, Figures 7–10); §5 Conclusion (co-invention, "adoption is only the beginning of deployment," the scope limitations). Conflict of interest: no independent author — Chatterji, Rakholia and Weeratunga are at OpenAI, and Holtz and Tambe "contributed to this work in their capacity as paid contractors for OpenAI." Parse and reconciliation notes: the raw is docling-derived and carries table-collapse / table-split-row / table-weld warnings with no repair blocks — every table in it is unreconciled. Tables 2 and 4, the only tables quoted on this page, were reconciled cell-by-cell against pdftotext -layout (pages 35 and 37) and the values above are the PDF's; the docling renderings of both are shifted (Table 2 welds column 2's Year-FE Yes into the industry-FE cell and orphans an Obs. value onto its own row; Table 4 slides the Complement measure and Observations rows sideways). Table A5's task taxonomy is collapsed and shifted — the health-and-wellness second-level values appear under "how to and procedural guidance" — and is cited nowhere. Two source-internal prose↔figure disagreements, both flagged rather than resolved: (1) Figure 2's plotted medians differ from the §4.2 prose, systematically and only for non-adopters — the chart labels read $167M revenue / 354 employees / $549M assets / $251M market value / $34M PP&E / $8.4M R&D against the prose's $209.6M / 424 / $667.6M / $316.4M / $43.7M / $9.9M, a ratio of ~1.18–1.29 on every line, with adopter values agreeing to within a few percent. The prose figures are the ones quoted (verified present in the PDF, so not a parse artifact); the two are not reconcilable from the document and one of them is stale. (2) §4.3.2 states early-career workers send "roughly eight to nine more weekly messages" than the firm average, while Figure 6 Panel B's point estimates read ≈ +6.5 (population-centered) and ≈ +7.4 (firm FE) by gridline arithmetic. The prose number is quoted with the chart read given alongside. All figure-derived values on this page (Figures 5, 6, 7) are chart reads from the two-pass image review and are labelled as such where they carry weight.