A new research paper from OpenAI, Columbia Business School, and the Wharton School has put hard numbers on something most business leaders have been sensing: AI adoption is not splitting evenly across companies. The gap between the businesses using AI well and those using it minimally is now 8.3 times wider — and it tripled in under eight months.

The paper, posted to arXiv on August 12, analysed more than 17 million real ChatGPT Enterprise messages from 1,764 organisations, making it one of the largest empirical studies of workplace AI use to date. The findings are worth understanding carefully, because they point to where enterprise AI is actually going.

The Frontier Gap Is Real — and Growing Fast

In January 2026, frontier firms (the top 10% of organisations by monthly AI usage) generated 2.6 times more output per active user than typical firms. By June, that figure had reached 8.3 times.

This is not just about which companies have more employees using AI. It is about depth. Frontier firms have moved from using AI as a writing assistant or search tool to embedding it in workflows where it executes tasks directly. More agents. More integrations. More output per session.

For context, the aggregate volume of output tokens produced across all ChatGPT Enterprise customers grew roughly sevenfold between June 2025 and March 2026. That growth is not evenly spread. A small number of organisations are responsible for a disproportionate share of it.

Who Is Actually Using AI at Work

One of the more counterintuitive findings: junior employees are driving usage, not executives.

Early-career workers and trainees send roughly eight to nine more weekly messages than the average active user within their own firm. Six months after an organisation adopts the tool, that gap has shifted further — junior staff sent 13 more messages per week than their most senior colleagues.

Executives, founders, and partners send fewer messages on average. This contradicts what most organisations assumed when they licensed enterprise AI tools. The productivity unlock is happening at the bottom of the org chart, not the top.

There are a few ways to read this. One is that junior employees have more repetitive, codifiable tasks that AI handles well — drafting, summarising, researching, formatting. Another is that they are less constrained by habits around how work should look. Either way, if your rollout strategy focused on getting executives on board first, the data suggests the real ROI is coming from the team underneath.

The Wealth Gap Problem

The paper also flagged something uncomfortable. Organisations that have adopted ChatGPT Enterprise have a median income, market capitalisation, and R&D spend that are all roughly ten times higher than those of non-adopters.

AI is not democratising competitive advantage yet. It is concentrating it among companies that were already well-resourced enough to move fast. That is partly a cost issue, partly a talent issue, and partly a data readiness issue. Companies that do not have the internal capability to find and act on high-value use cases are not seeing returns from their licences.

This finding is worth sitting with. Simply buying access to frontier models does not close the gap. What closes it is the ability to identify where AI can replace significant amounts of work, build the workflows around it, and keep iterating.

What This Means for Business

The gap between frontier firms and typical firms is not primarily a technology gap. Both groups have access to the same models. The difference is in how they are using them.

Frontier firms have identified specific business processes where AI can do substantial work autonomously. They have built agents that connect to company context and tools. They have turned individual use cases into repeatable practices. And they are iterating faster because they have people who understand what good AI deployment looks like.

For leaders watching this data, the questions to ask are operational ones. Where in your business is significant human time going to tasks that are structured enough for an agent? What does your team’s data literacy look like? Are your people able to evaluate AI output, spot errors, and improve prompts over time?

The companies pulling ahead are not using a different product. They are using the same products differently — with more intention, more skill, and more capability to build on what works.

Enterprise DNA exists to help organisations develop exactly that capability — the combination of data skills, AI literacy, and operational know-how that turns AI access into AI advantage. The frontier firms in this study did not get there by accident.

---

Source: “From assistance to execution: How enterprises put AI to work” — OpenAI, Columbia Business School, Wharton School. Working paper posted to arXiv August 12, 2026.