The AI Adoption Gap Widened in 2026: What to Measure Now

The gap between companies that use AI the most and those that use it the least isn't widening slowly anymore, it more than tripled in six months. By June 2026, the most intensive decile of enterprise users consumed 8.3 times more output tokens per user than the mid-tier, up from 2.6 times in January of the same year (OpenAI, 12 August 2026). This isn't an alarming finding. It simply puts a number on something many business owners already sensed but couldn't quantify.

What exactly does the OpenAI report say?

On 12 August 2026, OpenAI published two reports on how AI spreads through organisations and what the most advanced ones do differently. The metric is output tokens consumed per user, compared between the top 10% of usage and the mid-tier, among the company's enterprise customers.

In January 2026, the gap stood at 2.6x. By June 2026, it had reached 8.3x. The growth isn't linear, it's accelerating, which points to a compounding effect. The organisations pulling ahead are consolidating their lead rather than levelling off, while the mid-tier progresses at a noticeably slower pace.

Why does the gap vary so much by sector?

The report breaks the gap down by sector, and the two extremes are worth noting. In information and technology, it reaches 11.7x. In manufacturing, it drops to 5.3x. Same measure, same method, very different result depending on the nature of the work observed.

This difference likely comes down to the raw material of the work in each sector. Text, code and documents lend themselves more directly to intensive use of a generative assistant than a physical production line, where AI tends to sit at the edges (planning, predictive maintenance, quality control) rather than at the core of the task itself. The report doesn't state the cause, so this reading is a plausible interpretation, not an established fact.

What does the shift to agentic usage actually mean?

A second figure in the report changes the nature of the conversation. By June 2026, agentic usage (measured in Codex tokens) accounted for 64% of combined output tokens among enterprise customers, Codex and ChatGPT together. In other words, most of the activity is no longer asking a question and reading a reply, it's handing over a task for an agent to complete from start to finish.

That shift matters more than the raw token volume. A company that asks a lot of questions is still in an assistance mode. A company that delegates whole tasks has moved into a different register: it has defined a scope, put checks in place, and accepted that an agent produces a result without validation at every intermediate step. That shift, more than the headline figure, is what separates advanced organisations from the rest.

What this figure tells you, and what it doesn't

This is the most important point in this article, and the one a rushed news brief would tend to skip. This data measures token consumption at a single vendor, not value created inside the businesses consuming those tokens. It says one solid thing: the gap in practice is widening, and it's accelerating between advanced and average organisations. It says nothing about another, equally important thing: return on investment.

A company can generate a lot of output tokens and get little value back, if usage stays scattered, poorly scoped, or concentrated on low-stakes tasks. Conversely, a company that uses AI modestly but on two or three well-chosen use cases can get more out of it than one showing ten times the volume with no method behind it. The adoption gap is a practice indicator, not a performance score, and treating it as a signal of success would be the first misreading to avoid.

It's also worth stating what this figure doesn't cover: it's proprietary data from one vendor about its own enterprise customers, not an independent measure of the whole market, let alone a reflection of SMEs generally. It points to an underlying trend, not an exhaustive ranking.

How does an SME know where it stands?

You don't need sophisticated measurement tools to get a sense of where you stand. Three observations, available without any special instrumentation, already give an honest picture.

The share of staff with genuine weekly usage. An open account means nothing on its own. What matters is how many people actually open the tool at least once a week for a real work task, not out of occasional curiosity. A gap between the number of accounts created and the number of people using them regularly is often the first thing worth looking at.

The number of tasks delegated end to end. Asking an occasional question and delegating a whole task (drafting a first version of a report, sorting an inbox, preparing a recurring summary) sit at very different levels of maturity. Counting how many tasks, even modest ones, are currently handed over without systematic rework gives a concrete read on the shift towards agentic usage described above.

Whether at least one procedure has been written down and reused. Usage that relies entirely on what individual people remember erodes over time and with staff turnover. A single documented method, reused from one session to the next, is worth more than a dozen one-off uses that were never captured.

These three observations don't replace a structured assessment, but they're enough to give an honest read on where an organisation stands before going further. The point isn't to catch up with an abstract ranking, it's knowing precisely what you're measuring in-house, and why.

What should you actually do with this?

The gap OpenAI measured isn't a call to chase volume: generating more tokens is nobody's goal. It's a call to move away from scattered usage and identify two or three specific use cases, with a defined scope and a way to check the result. That's exactly what a funded assessment is for: scoping where the real gains sit in your organisation, before spending time or budget on poorly targeted projects.