OpenAI’s internal data reveals a near-complete shift from ChatGPT to agentic workflows
On 25 June 2026, OpenAI published a landmark internal analysis titled “How agents are transforming work,” revealing a near-complete migration away from conversational chatbot interfaces within their own organisation. The data is striking: in August 2025, the average OpenAI employee was directing less than 10 per cent of their computational token output through Codex, the company’s agentic platform. By June 2026, that figure had reached 99.8 per cent of total weekly output tokens across the entire organisation. In practical terms, the creators of ChatGPT have largely stopped using ChatGPT for their own work.
This is not a story about engineers adopting new developer tooling. The adoption data covers finance, recruiting, and legal departments, with active user rates of 91 per cent, 89 per cent, and 88 per cent respectively across those non-technical business units. The engineering cohort sits at 99 per cent active adoption. What OpenAI is describing is an organisation-wide operational transformation, not a niche technical experiment. OpenAI President Greg Brockman publicly characterised the shift by stating that agents are being adopted rapidly and are accelerating everyone’s work.
For environmental consulting practices, infrastructure project teams, legal counsel advising on planning and contamination matters, and local government organisations managing large asset and compliance portfolios, this development is worth close attention. It signals that the productivity ceiling of prompt-and-response AI is already being left behind by the organisations that built it, and that the professional services sector needs to think seriously about what agentic workflows mean for how technical work is scoped, delegated, reviewed, and assured.
Key details from OpenAI’s published analysis
The core statistical finding from OpenAI’s published report is the shift in token distribution. Token consumption is the standard unit for measuring computational workload in large language model systems: every word, sentence, and instruction processed by the AI consumes tokens. When OpenAI reports that 99.8 per cent of their weekly output tokens are now generated through Codex rather than standard ChatGPT interfaces, they are describing where almost all meaningful AI-assisted work is actually taking place within the organisation. The residual 0.2 per cent attributed to conventional chatbot interaction represents a near-complete functional abandonment of the prompt-and-response model for internal work purposes.
The growth figures for non-developer adoption are particularly significant. According to the published analysis, individual non-developer usage of Codex grew 137 times over the period studied, while organisational non-developer usage across non-technical teams grew 189 times over the same interval. OpenAI attributes this acceleration to two specific platform changes: connecting Codex to more capable underlying models, and introducing business-process features including macOS computer-use capabilities and role-specific plug-ins. These additions allowed the platform to move beyond a coding assistant function and into a generalised workflow-automation role accessible to staff without technical backgrounds.
The nature of the tasks being delegated to agents is also shifting toward greater complexity. OpenAI’s data indicates that nearly a quarter of all Codex requests now correspond to what the company classifies as long-horizon tasks, defined as work that would require a human more than one hour to complete manually. These are not simple queries or single-step information retrievals. They are multi-step processes where the agent autonomously orchestrates tool calls, interacts with software environments, and iterates toward a defined high-level goal without requiring continuous human prompting at each stage.
The economic implications for software and AI budget structures are material. OpenAI’s analysis notes that as agentic workloads increase, token consumption shifts from the incremental, human-paced pattern of conversational use to large, autonomous background bursts. This fundamentally changes how organisations should model AI-related costs. Budgeting based on active user seat-time or on anticipated human-initiated prompts will significantly underestimate actual consumption once agentic workflows are running in parallel across an organisation. Enterprise software procurement teams and finance functions will need to restructure cost models accordingly.
Australian professional services context and business implications
Australian professional services firms, including environmental consultancies, engineering practices, planning advisory groups, and legal teams handling project approvals and environmental compliance, operate in a market where technical labour is the primary input cost and regulatory documentation is the primary output. The OpenAI findings are directly relevant to these business models. Agentic AI platforms capable of autonomously drafting, cross-referencing, and iterating on complex technical documents represent a structural shift in how that labour is deployed. The key insight from OpenAI’s data is that non-technical staff, including those in legal and finance roles comparable to staff in any professional services firm, have adopted agentic tools at rates comparable to software engineers once business-process features were made accessible.
Australian businesses adopting agentic AI will need to engage with existing professional and regulatory frameworks around quality assurance, professional responsibility, and document integrity. In environmental consulting specifically, technical reports submitted to state environment protection authorities, planning bodies, and accredited auditors carry statutory weight. The professional certifying a preliminary site investigation, detailed site investigation, or remediation action plan retains legal and professional responsibility for the conclusions reached, regardless of the tools used to produce the document. Agentic workflows that accelerate drafting, cross-referencing, or data synthesis do not alter that responsibility, but they do require firms to develop clear internal protocols around review, version control, and sign-off to ensure that AI-assisted outputs meet the same standards as manually produced work.
The pace of internal adoption documented by OpenAI also carries a competitive signal for Australian firms. When an organisation of that technical sophistication moves 99.8 per cent of its productive AI workload onto agentic platforms within roughly ten months, it suggests that the window for treating agentic AI as a future consideration rather than a present operational question is narrowing. Firms that establish governance frameworks and workflow integration now will be better positioned than those that wait for sector-wide standards to emerge before engaging.
References and related sources
- Primary source: openai.com
- htx.com
- gigazine.net
- substack.com
- thenextweb.com
How iEnvi can help
iEnvi integrates technology and data-driven approaches into environmental consulting. We monitor AI and technology developments that affect how environmental professionals deliver services to clients.
This is an iEnvi Machete news summary. Prepared by iEnvi to summarise the source article for environmental professionals tracking AI, data, and technology developments that affect consulting and project delivery.
Published: 29 Jun 2026
Need advice on this topic? Speak to an iEnvi expert at info@ienvi.com.au or 1300 043 684, or contact us online.
Need advice on this issue? iEnvi provides practical, senior-led environmental consulting across contaminated land, remediation, ecology and environmental risk.
PSI services DSI services Remediation services Talk to iEnvi