OpenAI’s enterprise data suggests that leading adopters are not merely using AI more—they are connecting it to longer, agentic units of work.
OpenAI’s August enterprise analysis describes a widening gap between typical organisations and “frontier firms”, defined in its dataset as the top 10% of AI usage each month. Those firms generated 8.3 times more output tokens per active user than typical firms. OpenAI also reported that Codex accounted for 64% of combined Codex and ChatGPT output tokens among enterprise customers in June.
These are provider-specific usage measures, not direct productivity or ROI. Output tokens are especially imperfect because longer agent tasks naturally produce more output. Even with those caveats, the pattern is important: advanced adopters appear to be shifting from isolated assistance to longer units of delegated work.
Adoption depth matters more than seat count
An organisation can buy thousands of licences and remain at a shallow level of adoption. Employees may draft emails or summarise documents while the underlying process remains unchanged.
Execution-level adoption looks different. AI is connected to:
- repositories and live business context;
- tools and APIs;
- coding and delivery pipelines;
- repeatable skills;
- multi-step workflows;
- approvals and policy;
- output verification;
- operational metrics.
The difference is not simply “better prompting”. It is whether the organisation redesigns work around the new capability.
Coding agents are an early indicator
Coding is one of the first domains where agents can receive a goal, inspect a large environment, modify artefacts, run tests and iterate. The output is machine-verifiable more often than in general knowledge work.
That makes software engineering a useful proving ground for enterprise agents. It exposes the same questions other functions will face:
- What context may the agent inspect?
- Which branches, environments or systems may it change?
- Which tests define acceptance?
- What requires review?
- How are secrets and external network access controlled?
- Who owns regressions?
- How is work resumed after interruption?
Organisations that learn to govern coding agents are building patterns transferable to finance, operations and customer service.
Skills convert tacit practice into reusable capability
OpenAI reports that plugins and skills are more common among frontier firms. The underlying idea is strategically significant.
A skill packages instructions, tools, resources, validations and sometimes business logic for a repeatable task. It turns “ask the model to do something” into an organisational capability that can be versioned and evaluated.
Examples include:
- prepare a board briefing from approved sources;
- assess a supplier onboarding package;
- create a release note from commits and incidents;
- investigate a service case;
- produce a jurisdiction-specific contract checklist.
The skill becomes a boundary between general model intelligence and company-specific practice.
The frontier gap is organisational
The leading constraint is often not access to a stronger model. It is the surrounding organisation.
Frontier adoption requires:
- owners who can redesign processes;
- domain experts who define acceptance;
- engineering that connects data and tools;
- security and legal teams involved early;
- evaluation infrastructure;
- incentives to change how work is allocated;
- a portfolio process that stops weak experiments.
Without those conditions, more usage can simply mean more ungoverned variation.
What leaders can do now
Choose one unit of work, not one department
Select a meaningful task that crosses enough systems to prove value but has a controllable consequence.
Measure completion
Track verified outcomes, cycle time, rework and review—not prompt volume.
Productise the successful pattern
Once a workflow works, package its tools, policy, evaluation and ownership as a reusable capability.
Build a shared control plane
Identity, traces, budgets, approvals and incident response should not be reinvented by each team.
Preserve model optionality
The organisational capability should survive a change in model or provider.
The Lexdatum view
The frontier gap is not a gap in enthusiasm. It is a gap in operating architecture.
Typical firms give people access to AI. Frontier firms begin to give AI access to controlled work. That transition is where both the value and the risk increase.
The next stage of transformation will be measured not by how many employees opened a chat window, but by how many important processes can be completed faster, with evidence, inside a defined authority envelope.
Sources
Research note: This analysis is based on public information available on 31 August 2026. Product capabilities, prices and regulatory interpretations can change. Vendor-reported metrics are identified as such. This is not legal, security or investment advice.