# 17 Million Messages Say Enterprise AI Is Mostly Writing > Source: > Published: 2026-08-15 00:08:32+00:00 [AI](https://sourcefeed.dev/c/ai)Article # 17 Million Messages Say Enterprise AI Is Mostly Writing OpenAI's first telemetry-based study shows broad but shallow adoption, driven by juniors and rich firms. [Mariana Souza](https://sourcefeed.dev/u/mariana_souza) OpenAI just did something vendors almost never do: it opened its enterprise telemetry to economists and published what it found. ["How Organizations Use AI: Evidence from ChatGPT"](https://cdn.openai.com/pdf/how-organizations-use-chatgpt.pdf) — written by OpenAI chief economist Aaron Chatterji with academic co-authors including Berkeley's David Holtz and Wharton's Prasanna Tambe, and posted to [arXiv](https://arxiv.org/abs/2608.12236) this week — links [ChatGPT Enterprise](https://openai.com/chatgpt/enterprise/) account records to worker roles, message-level task classifications, and public-company financials through March 2026. The worker-level sample alone covers more than 1,500 organizations and 17 million messages. That makes it the best administrative-data look at enterprise AI we have. It's also, read carefully, a corrective to two years of survey-driven hype. The picture that emerges isn't transformation. It's a very large, very fast-growing writing and documentation machine that companies are still figuring out what to do with — the authors' own closing line is that firms "are still actively learning how to integrate AI into organizational workflows." ## What the telemetry actually says Growth is real and steep: aggregate output tokens rose roughly sevenfold between June 2025 and March 2026, and about fourfold even within the cohort of firms that had already adopted by June 2025. So this isn't just new logos — existing customers are using it harder. But adoption is lopsided in a familiar way. Among U.S. public companies, adopters had median revenue of $2.28 billion versus $210 million for non-adopters, and median headcount of 2,934 versus 424. The strongest financial predictor of adoption wasn't R&D intensity — it was SG&A stock per employee, the accounting bucket where white-collar overhead lives. That's a tell. The firms buying enterprise AI first are the ones drowning in knowledge-work coordination costs, not necessarily the ones building technical products. And what do people do with it once it's deployed? In the task-classified subsample (973 organizations, 8.7 million messages sorted into 60 categories), the volume concentrates in documentation, technical writing, and communication. More than half of active users perform a documentation or technical-writing task at least weekly; nearly half do some form of technical digital work. Business research, legal, and financial tasks reach plenty of users but generate comparatively few messages. ## The juniors are the power users The finding that deserves the most attention inverts the standard survey result. Six months after adoption, early-career workers and trainees send roughly eight to nine more messages per week than the average active user at their firm, while managers and executives send fewer. Surveys keep telling us leadership is bullish and the rank-and-file is hesitant; the message logs say the opposite — enthusiasm at the top, usage at the bottom. That should reframe the "AI is coming for junior jobs" debate. Whatever happens to junior hiring, the juniors who are employed have become the heaviest users, presumably because AI substitutes for the thing they lack — accumulated context and someone senior with time to answer questions. There's a real organizational risk buried in that: if the entry-level ranks learn their craft by asking a model instead of a mentor, firms are quietly rewiring how expertise propagates, and nobody is measuring what that does to quality five years out. The paper doesn't touch outcomes — by design — so this is exactly the kind of question its data can't answer yet. ## The blind spot that matters for developers Here's the part SourceFeed readers should sit with: engineering and technical practitioners make up just 11% of weekly active users in this data, behind managers and directors (24%) and individual contributors generally (15%). Read naively, that says developers are laggards. The more likely explanation is that the paper is pointing its telescope at the wrong sky. This dataset sees only ChatGPT Enterprise chat traffic. It cannot see API usage, and it can't see the tools where serious engineering adoption actually happened over the study window — IDE-integrated assistants and agents like [GitHub Copilot](https://github.com/features/copilot), Cursor, and Claude Code, which live in the editor and terminal, not in a chat tab. The authors flag this limitation themselves. By March 2026, a developer's AI consumption is mostly agentic and mostly invisible to a chat-message counter. "Messages per week" is already a legacy metric for technical work, roughly the way "minutes of dial-up" was a bad proxy for internet use in 2005. That blind spot cuts both ways for interpretation. The paper likely *understates* technical adoption economy-wide, while accurately showing that the chat interface itself has settled into a specific niche: prose, docs, synthesis, and comms — the tasks where a conversation window is genuinely the right UI. ## What to do with this If you run a platform or enablement team, this data is a demand signal. The organic, high-volume use case inside enterprises is documentation and technical writing — so before you build a bespoke agent for some exotic workflow, make the boring path excellent: wire your internal docs, runbooks, and wikis into whatever assistant your company licenses, because that's where the messages already are. If you're deciding where AI budget goes, note that the spend gap between chat seats and API/agent infrastructure is also a measurement gap; don't let a dashboard of chat metrics convince leadership that engineering "isn't adopting." My read: this paper is genuinely valuable — large-N administrative data beats vibes and surveys, and OpenAI deserves credit for the caveats it prints about its own product. But its headline growth numbers describe adoption of a chat interface, and the frontier has moved to agents that this data cannot observe. Enterprise AI in 2026 is broad, shallow, junior-heavy, and concentrated in rich firms — and the most economically interesting usage is the part that never shows up as a message. ## Sources & further reading - [How Organizations Use AI: Evidence from ChatGPT [pdf]](https://cdn.openai.com/pdf/how-organizations-use-chatgpt.pdf)— cdn.openai.com - [How Organizations Use AI: Evidence from ChatGPT](https://arxiv.org/abs/2608.12236)— arxiv.org - [How ChatGPT Enterprise users use AI at work](https://techinformed.com/how-chatgpt-enterprise-users-use-ai-at-work/)— techinformed.com - [How Organizations Use AI: Evidence from ChatGPT [pdf]](https://news.ycombinator.com/item?id=49290768)— news.ycombinator.com [Mariana Souza](https://sourcefeed.dev/u/mariana_souza)· Senior Editor Mariana covers the fast-moving world of machine learning and generative AI, with a particular focus on how these technologies are reshaping development workflows. When she isn't stress-testing the latest foundation models, she's usually at a local hackathon. ## Discussion 0 No comments yet Be the first to weigh in.