Welcome Back to XcessAI

A few years ago, measuring AI adoption seemed relatively straightforward.

Does the company use AI? Does it have ChatGPT licences? Has it rolled out Copilot? How many employees have access?

Those questions made sense when artificial intelligence was still new inside most organisations. But they are becoming increasingly unhelpful.

Two companies can give exactly the same AI tools to exactly the same number of employees and end up using them in completely different ways.

One asks AI to summarise emails. The other asks it to investigate a problem, analyse files, write code, interact with software and complete a multi-step workflow.

Technically, both companies have "adopted AI". Operationally, they are doing something very different.

New data from OpenAI makes the distinction more visible.

Among the EU enterprise customers in its latest analysis, OpenAI found that Codex and ChatGPT Work, tools designed to carry out work rather than simply answer questions, now account for more than half of output tokens.

Yet adoption intensity varies greatly.

The most intensive EU firms generate more than four times as many output tokens per active user as typical EU firms. And even those leading EU companies generate less than half as many tokens per user as frontier firms outside the EU. OpenAI

There is an important caveat: Tokens are not profits. They are not productivity. They are not proof that one company is better at AI than another. But they do reveal something increasingly important.

The AI divide may no longer be primarily about access. It may be about how deeply organisations use the intelligence they already have.

The original enterprise AI question was binary.

Access or no access.

Some companies moved quickly.

They purchased licences. Connected employees. Created policies. Started pilots.

Others waited.

That produced an obvious adoption gap. But access is becoming commoditised.

Powerful AI is increasingly available through software employees already use. Models themselves continue to improve. Prices fall. Interfaces get easier.

The technical barrier to putting capable AI in front of an employee is becoming relatively small.

The organisational barrier is not. Because giving someone ChatGPT is not the same as changing how they work.

Most people initially use new technology by inserting it into an existing workflow.

The process remains fundamentally human. AI sits beside it. That is adoption.

But deeper adoption looks different. The workflow itself changes.

OpenAI's data captures part of this transition through the rise of what it calls agentic AI.

Traditional ChatGPT interactions generally start with a question or request.

An agent can go further. It can inspect information. Use tools. Edit files. Write and execute code. Move between applications. Perform several steps before returning with the result.

The difference is subtle but important.

One interaction looks like:

"Help me do this."

The other increasingly looks like:

"Do this."

OpenAI says that among the EU enterprise customers in its sample, Codex and ChatGPT Work now account for more than half of output tokens. Among small and medium businesses, weekly active users of those tools have increased more than eightfold since May. OpenAI

The direction is therefore clear. Enterprise AI is moving from assistance towards execution.

OpenAI's broader enterprise data has been showing the same pattern. Earlier this year, it found that firms at the frontier of usage were increasingly distinguished not simply by sending more messages, but by asking AI to perform richer, longer and more complex work. OpenAI

That distinction matters. Because ten prompts asking for email rewrites and one agent performing a thirty-minute analytical workflow may both appear as AI usage.

But economically they are not remotely equivalent.

Imagine two consulting firms.

Both purchase enterprise AI for every employee.

At Firm A, consultants use it to improve writing, summarise research and brainstorm presentation titles.

Useful.

At Firm B, teams build repeatable workflows around it.

An agent searches internal research.

Pulls relevant client information.

Analyses spreadsheets.

Produces a first version of the analysis.

Checks it against internal standards.

Creates supporting materials.

A consultant reviews the result.

Both firms can truthfully say:

"We use AI."

But the statement tells us almost nothing.

The more interesting question is:

What fraction of the workflow has been redesigned around AI?

That is much harder to measure.

And it is likely to become much more important.

This is where OpenAI's data needs to be handled carefully.

The company measures usage intensity using output tokens per active user.

A token is essentially a small unit of generated text or machine-readable output.

If a model generates more output, it generates more tokens.

Agentic workflows naturally tend to produce many of them because an agent may perform numerous intermediate steps while completing a task.

So firms producing many more tokens per worker are probably asking AI to do more work, more complex work, or both.

But we should stop there.

A million tokens can produce tremendous economic value.

They can also produce tremendous amounts of useless text.

OpenAI itself describes token volume as an imperfect proxy rather than a direct measure of business value. OpenAI

That distinction is crucial.

The data does not tell us that companies generating four times more tokens are four times more productive.

It certainly doesn't tell us they earn four times the return.

What it does tell us is that the intensity with which organisations are asking AI to work differs dramatically.

And that alone is interesting.

The EU numbers make the divergence particularly visible.

OpenAI defines "frontier firms" as the top 10% of companies in its sample based on output tokens generated per active user.

In September, those leading EU firms generated more than four times as many tokens per user as typical EU firms.

But frontier firms outside the EU generated more than twice as many again as the EU leaders. OpenAI

There is also enormous variation within Europe.

Among the ChatGPT Enterprise accounts studied, the proportion of output generated by Codex and ChatGPT Work ranged from 22% in Slovenia to 85% in Lithuania. OpenAI

It would be a mistake to turn those numbers into a league table of national AI capability.

The sample contains only OpenAI Enterprise and Business customers. Different countries have different industry mixes, customer mixes and sample sizes.

It tells us nothing directly about companies using Google, Anthropic, Microsoft, open-source models or internal AI systems.

So this is not evidence that "Europe is falling behind in AI." It is something narrower.

Among companies already paying for broadly similar OpenAI capabilities, usage intensity differs enormously.

That is more interesting. Because access alone cannot explain the gap.

Why would two companies with access to similar technology use it so differently?

Probably because enterprise AI adoption increasingly depends on things that have relatively little to do with the model.

Imagine an analyst who discovers that AI can automate 70% of a weekly report.

That discovery creates individual productivity.

But unless the workflow changes, the analyst may simply save a few hours every Friday.

For the organisation to capture the full benefit, somebody needs to ask a more difficult question:

Why are we still producing the report in exactly the same way?

Perhaps the AI should generate it automatically.

Perhaps the underlying data should flow directly into an agent.

Perhaps managers should receive exceptions rather than a full report.

Perhaps the analyst should stop producing the report entirely and spend the time investigating what the numbers mean.

Now we are no longer implementing an AI tool.

We are redesigning work.

And organisations are usually much slower at redesigning themselves than technology companies are at improving models.

This creates a potential problem for management dashboards.

Suppose a CIO reports:

92% of employees now have access to our approved AI platform.

That sounds impressive.

Perhaps 70% used it last month. Also impressive.

Perhaps employees collectively sent one million prompts. Excellent.

But none of those numbers tells management whether the organisation has actually changed.

A company can achieve near-universal AI access while remaining remarkably shallow in its use. People might be generating meeting summaries, fixing grammar and asking occasional questions.

The technology is everywhere. The operating model is unchanged.

That creates what we might call an adoption illusion.

The company has successfully deployed AI without necessarily becoming an AI-enabled organisation.

This distinction becomes increasingly important as AI moves into agents.

With a traditional productivity tool, value often scales with the number of employees using it.

Give more people better spreadsheets.

Give more people better communication software.

Give more people better search.

Agentic AI introduces another dimension.

How much work is each user delegating?

One employee might use an agent once a week.

Another might have ten agents performing work continuously.

A third might have built a workflow used by an entire department.

The number of licensed users could remain identical while the amount of economically useful work performed by AI differs by orders of magnitude.

That means traditional software metrics may become less informative.

Seats matter. Active users matter.

But increasingly, depth matters too.

Perhaps we therefore need a better word than adoption.

Adoption describes whether technology has entered the organisation.

What matters next is whether the organisation has absorbed it.

Has AI become part of how decisions are made? Is it connected to company information? Can it use internal tools? Are repeatable workflows built around it? Can employees delegate meaningful pieces of work rather than isolated tasks? Has management redesigned processes to take advantage of capabilities that didn't exist two years ago?

This gives us a simple progression:

Access → Usage → Integration → Delegation

Most companies are already somewhere on that path.

The competitive difference may increasingly come from how far along it they have travelled.

There was a time when simply owning a computer differentiated a company. Then computers became universal and the advantage moved somewhere else.

Into software. Processes. Data. Skills.

Eventually, nobody asked whether a company "used computers."

The question became meaningless.

AI may be heading in the same direction.

Soon almost every professional organisation will have access to powerful models.

Almost every employee will be able to ask them questions. Almost every company will claim to use AI.

At that point, AI adoption itself stops being particularly interesting.

What matters is what happens after adoption.

OpenAI's token data cannot tell us whether companies doing more of this are generating superior returns. We will need much better evidence for that.

But it does show something important. Companies with access to similar intelligence are already using radically different amounts of it.

The first divide was between businesses that had AI and businesses that didn't.

The next may be harder to see. It will run directly through companies that all appear to have adopted it.

On one side, AI is another tool.

On the other, it is becoming part of how the company works.

Until next time,
Stay adaptive. Stay strategic.
And keep exploring the frontier of AI.

Fabio Lopes
XcessAI

💡Next week: I’m breaking down one of the most misunderstood AI shifts happening right now. Stay tuned. Subscribe above.

Read our previous episodes online!