Enterprise ChatGPT Adoption: What Actually Happens Inside Firms

Enterprise ChatGPT Adoption: What Actually Happens Inside Firms

Introduction: What this new research reveals about ChatGPT in the workplace

If you’ve ever wondered “Who in the company is really using ChatGPT—and for what?” you’re not alone. A lot of discussion about enterprise AI is built on guesses, surveys, or small pilots. But new research from Aaron Chatterji et al. on arXiv (2608.12236) takes a much more direct look: it studies how organizations use a real, centrally managed product—ChatGPT Enterprise.

The authors link administrative ChatGPT Enterprise account records to workplace context like worker roles, task categories, and public-company financial data (through March 2026). The scale is huge: the worker-level sample analyzed at a six-month horizon includes 1,500+ organizations and over 17 million messages. Importantly, the message content itself is classified automatically—no manual review of customer messages—so the analysis stays privacy-preserving while still getting at what people actually do.

This means the paper isn’t just asking whether firms adopt AI. It asks how adoption grows, who uses it most intensely, where it shows up in the organizational hierarchy, and what kinds of knowledge work it’s being used for—writing, communication, technical tasks, research, analysis, legal/regulatory work, finance, and more.

Why this matters (and what you can do with it right now)

This research is especially relevant now because enterprise AI has moved from “can we access it?” to “what will we do with it?” Many companies already have accounts, champions, and vague rollout plans. But the hard part isn’t buying access—it’s learning where AI fits into workflows, who should use it, and how to avoid bottlenecks when only a small slice of the workforce actually turns usage into output.

A concrete scenario: imagine you’re an HR or operations leader trying to write a realistic AI adoption plan for a mid-sized public company. You might be tempted to target “knowledge workers” broadly, assuming junior roles will naturally pick it up and leaders will later operationalize it. This paper suggests that’s not just a vibe—it’s measurable. It finds early-career workers send dramatically more messages per active user (about 8 to 9 more weekly messages than the within-firm average active user). So if you launch training and onboarding only for executives or only for “the usual suspects,” you may be steering effort away from where usage intensity is already happening.

Compared to earlier AI research that focused on model capability or broad occupational exposure, this work is closer to the “deployment reality check.” Prior surveys help, but they rely on recall and self-report. Here, the authors use telemetry-style administrative usage data plus job titles and task classification. In other words, it builds on the research that AI is a general-purpose tool by showing what that looks like inside organizations: adoption is broad but uneven, and the economic story depends on how firms convert “access” into “co-invention” with internal workflows.

How the study measured enterprise AI use without getting lost in noise

A key reason this paper is credible is its measurement strategy. Instead of relying on self-reported usage, it uses organization-week panels from ChatGPT Enterprise—a centrally administered workspace. That means the “who” and “how much” are grounded in actual product activity.

What data the authors actually used

The study builds four related datasets:
- Aggregate enterprise usage sample: an organization-week panel for organizations whose ChatGPT Enterprise adoption dates fall between Jan 1, 2024 and Mar 31, 2026. They observe messages sent, active users, and output tokens over time.
- Worker characteristics + industry sample: a subset where employee job titles and firm industry (NAICS) are available with enough coverage. This sample includes 1,764 organizations and 17,446,551 messages.
- Task classification subsample: a further subset for message-level task categorization, including 973 organizations and 8,696,657 classified messages. The task classifier becomes available beginning Oct 30, 2025, and analyses focus on the 26-week (six-month) horizon after adoption.
- Public company financial sample: ChatGPT Enterprise adopters mapped to U.S. public-company tickers via an account-to-ticker bridge, linked to Compustat data. The linked panel includes 417 public-company tickers and 521 ticker-years with usage matching requirements.

How adoption and usage are framed

The paper tracks:
- Adoption on the “extensive margin”: whether a firm starts using.
- Usage intensity on the “intensive margin”: how much they use after adopting.
- Within-firm heterogeneity: how different worker groups use AI at the six-month horizon.

This is a big deal because many AI conversations jump straight from “someone can do it” to “it changes performance.” The authors are careful to separate capability, adoption, deployment, and realized usage.

Four facts about enterprise AI adoption and use that show “learning in progress”

The paper’s headline contribution is four stylized facts about enterprise adoption and usage patterns. Here they are in plain language.

Fact 1: Enterprise usage grew fast—both from new adopters and deeper usage

Across the period, total output tokens from ChatGPT Enterprise customers increased dramatically. The authors report that output tokens grew roughly sevenfold between June 2025 and March 2026.

Crucially, this isn’t only because more firms got onboarded. For firms that adopted by June 2025, tokens increased about fourfold over the same window. In other words, existing adopters kept deepening their usage after onboarding—and usage accelerated in early 2026 across cohorts.

Fact 2: Early U.S. public-company adopters are bigger and more “intangible-invested”

Compared with non-adopters, early adopters in U.S. public companies look like a specific type of organization:
- Median revenue: $2,275.1M for adopters vs $209.6M for non-adopters
- Median total assets: $4,394.2M vs $667.6M
- Median R&D expense: $113.1M vs $9.9M
- Median employment: 2,934 vs 424

The regression evidence supports this pattern as an association, not a causal claim: larger and more productive firms appear more likely to adopt, and adoption is tied to business attributes beyond just size.

Fact 3: Use shows up across functions and seniority—but intensity varies a lot

Using job title classes and seniority levels, the paper shows two complementary patterns:
- Breadth of participation (who is active)
- Intensity of use (how much active users message)

Active usage spans multiple job functions and ranks, but message volume is uneven. One of the clearest patterns is a negative seniority gradient: early-career workers and trainees send far more messages per active user than senior employees.

Fact 4: The work is broad—especially writing, communication, and synthesis

At the task level, ChatGPT Enterprise use isn’t dominated by one single workflow. Instead, it covers many categories of knowledge work tasks, including:
- Writing / technical writing
- Technical digital work
- Communication
- Information synthesis (“topic overviews,” “facts and figures”)
- Research / planning
- Data analysis
- Legal and regulatory work
- Finance and tax

This matters because generative AI behaves like a general-purpose tool inside organizations. The paper’s results align with that broader idea, while still showing where adoption effort concentrates in practice.

What predicts who adopts? Scale and “complements” look like the real gating factors

The adoption story here isn’t “AI adoption is random.” The authors test firm characteristics linked to adoption probability and usage intensity.

Which firm attributes correlate with adoption (public-company evidence)

The paper reports several linked patterns:
- Adoption probability rises with log revenue per employee (conditional on controls), and also rises strongly with log employment.
- There’s a more nuanced relationship with physical capital intensity (PP&E/emp), which is often not positively associated once scale and other controls are included.
- Beyond scale, adoption correlates with pre-existing investments in organizational/“intangible” capabilities:
- SG&A stock per employee is the strongest and most robust complement associated with adoption.
- R&D stock per employee is also positively associated (robust especially depending on sector exclusions).
- Capitalized software shows a positive association in the full sample.

A comparison of “who adopts” vs “how intensely they use” (important distinction)

A subtle but important insight: adopters are different, but conditional usage intensity doesn’t mirror the same pattern in a simple way.

Here’s how the paper’s findings contrast the extensive margin (adoption) with the intensive margin (per-employee usage):

This is exactly the kind of “not all adoption equals realized value” nuance you want.

Who uses it inside the firm? The workforce isn’t uniform—and early-career workers are the power users

The paper’s internal workforce analysis is one of the most actionable parts for readers trying to plan adoption.

Breadth: which roles are active after six months

Six months after adoption, the distribution of weekly active users spans multiple job title classes. At the average firm, for example:
- Engineering/technical practitioners: ~11%
- Executives/founders/partners: ~9%
- Finance/accounting: ~5%
- Marketing/communications: ~5%
- Sales/account management: ~4%

Across seniority levels, active usage is also spread:
- Managers/directors: ~24%
- Individual contributors/professionals: ~15%
- Senior IC/principals: ~14%
- Executives: ~10%
- Early-career/trainees: ~7%

Two things can be true at once: broad participation, and unequal participation intensity.

Intensity: who messages the most once they’ve become active

Conditional on being active, intensity varies systematically:
- Early-career workers/trainees send about 8 to 9 more weekly messages than the within-firm average active user.
- Executives/founders/partners send fewer messages than other active users within the same firm.

The paper also finds intensity differs by job function: analysts and marketing/communications workers send more messages per active user than the average active user, while executives tend to send fewer.

Practical implication: plan for “user energy,” not just “seat count”

If you’re building training programs or internal playbooks, this suggests you should measure adoption success in terms of:
- Who becomes an active user
- How intensely they use the tool
- What tasks they apply it to

“Everyone has access” isn’t enough—intensity and task fit decide whether AI becomes workflow infrastructure or remains a curiosity.

What tasks get done with enterprise ChatGPT? The adoption map looks like a knowledge-work operating system

This is where the research shifts from organizational sociology to work design.

Two ways to measure “task use”

The authors use a message/task taxonomy with two complementary measures:
1. Task prevalence: share of weekly active users who did a task at least once
2. Message share: share of total classified messages assigned to each task category

These measure different things:
- Prevalence is “how many people try it”
- Message share is “how much effort goes into it”

Overall: writing and information work dominate, but many domains show up

Overall task structure includes (with emphasis varies between prevalence and message share):
- Documentation / technical writing
- Technical digital work
- Drafting messages and communication
- Topic overviews, facts and figures
- Research and planning
- Data analysis
- Legal/regulatory work
- Finance/tax tasks
- plus a long tail (“other task classifications”)

The authors emphasize that task use is broad exposure among active users, not one single workflow. So even if your org thinks of “AI” as one thing (like “draft emails”), it’s probably being used as a multi-tool across functions.

Industry differences: broader similarity, sharper differences in “first tries”

Task patterns vary by industry in ways consistent with what job content looks like. For example:
- Financial and tax tasks show up more in finance/insurance contexts
- Sales/marketing tasks appear more in arts/entertainment/retail than in manufacturing

But the industry story is weaker when tasks are weighted by message volume. That suggests that industry differences show up more on the prevalence margin than on the intensity margin—a subtle but important planning clue.

Role and seniority differences: overlap with specialization

Task mix differs by job title class in role-consistent ways:
- engineering roles: more technical digital work / debugging
- finance roles: more financial/tax tasks
- sales/marketing roles: more sales/marketing tasks

Across seniority, early-career users and ICs are represented more in production-like categories, while executives are relatively more in categories like:
- topic overviews
- facts and figures
- legal and regulatory work
- financial/tax-related tasks

One interpretation: AI may be used differently at different hierarchy levels—not just “how much,” but what kind of work gets delegated vs synthesized vs reviewed.

What this means for leaders: adopt faster, but deploy smarter

The conclusion the authors land on is basically: adoption is only the beginning. General-purpose technologies generate value only when firms learn how to integrate them into workflows—through experimentation and organizational change.

If you’re a leader deciding what to do next, this paper supports a more disciplined approach:
- Treat enterprise AI deployment as a co-invention process, not an implementation checklist.
- Expect heterogeneous adoption speeds across firms and heterogeneous intensity across worker groups.
- Use telemetry insights (or close equivalents) to track where AI is actually landing in your workflow—not just where you told it to land.

The paper also flags limitations: it studies ChatGPT Enterprise specifically (not all AI tools), worker role coverage is incomplete for some users, and tasks measured from message content don’t automatically prove downstream productivity effects. Still, even with those constraints, the patterns are strong enough to guide strategy.

Key Takeaways

• Enterprise ChatGPT usage grew rapidly (about 7x output tokens from June 2025 to March 2026), driven by both new adoption and deeper use within existing adopters.
• Early public-company adopters are not average firms: they’re larger and more heavily invested in intangible/organizational complements (especially SG&A, plus R&D and software).
• Inside adopting firms, usage is broad but uneven: many roles participate, but early-career workers/trainees are especially intense (about 8–9 more weekly messages per active user than the within-firm average).
• The work is multi-domain: most enterprise usage clusters around writing/technical writing, technical digital work, communication, and synthesis—but also includes research, planning, data analysis, legal/regulatory, and finance/tax tasks.
• Practical planning implication: don’t judge AI rollout by access alone. Track who becomes active, how intensely they use, and which task categories dominate—because those determine whether AI becomes workflow infrastructure.
• Strategic implication for the future: firms are still learning how to integrate a general-purpose AI tool into organizational hierarchies. The economic payoff likely depends on whether individual usage turns into complementary organizational capabilities over time.

If you want, tell me what kind of organization you’re thinking about (industry + rough size + which departments care most). I can translate these findings into a rollout and measurement plan tailored to your situation.