OpenAI’s “frontier gap” study: what the enterprise AI data shows
OpenAI has released a detailed empirical study examining how its enterprise customers actually put generative AI to work, and the findings describe a widening operational split that the report labels the “frontier gap”. Rather than measuring adoption by headcount or licence numbers, OpenAI analysed output token consumption across its business customer base, effectively tracking how much genuine AI-driven work organisations are producing. The top 10 per cent of adopting organisations, described as frontier firms, generate 8.3 times as many output tokens per active user as typical enterprise customers.
This matters well beyond the technology sector. For consulting, engineering and professional services firms, including those operating in environmental assessment, planning advice and regulatory reporting, the research provides one of the clearest data points yet on where AI value is actually accumulating. It is not accumulating in firms that treat AI as a faster typing tool. It is accumulating in firms that have restructured workflows so models can act with a degree of autonomy across multi-step tasks.
For managing partners, technical directors and in-house counsel evaluating AI investment decisions, the practical question this research answers is blunt. Is your organisation using AI to answer questions, or is it using AI to complete work? OpenAI’s data suggests those are now two fundamentally different cost and productivity categories, and the gap between them is growing rather than narrowing.
How frontier firms use agentic AI workflows differently
The headline metric in the report is the 8.3x output token differential between frontier firms and average enterprise users. Output tokens are a reasonable proxy for the volume of substantive work a model is producing, as opposed to the volume of prompts a human is typing. A single conversational exchange, where a user asks a question and receives a short answer, generates comparatively few output tokens. A multi-step agentic workflow, where a model retrieves data, cross-checks it against internal systems, drafts a document, validates that draft against rules, and iterates, generates substantially more.
OpenAI identifies three structural factors driving the gap. First, tool and skill binding: frontier firms have connected models directly to enterprise databases and internal applications through custom plugins and application programming interface calls, rather than relying on manual file uploads into a chat window. Second, workforce dynamics: the report finds that early-career employees are adopting agentic execution tools fastest, using connected models to complete cross-functional tasks that previously required extensive manual coordination between teams. Third, a shift in training investment: frontier organisations are moving internal training spend away from prompt engineering and towards structured agent infrastructure, governance controls and tool-calling interfaces that allow safe delegation of full task lifecycles.
The report frames this as a transition from AI as conversational assistant to AI as autonomous execution engine. In practical terms, average organisations are still using models for short, single-turn interactions, drafting an email, summarising a document, answering a discrete question. Frontier organisations have built repeatable, multi-agent workflows where a model can run a task from start to finish with human review only at defined checkpoints, rather than at every step.
OpenAI’s own recommendation to enterprise leaders is explicit: shift focus from training staff on prompt syntax towards building the infrastructure and control planes that let AI systems safely execute full task lifecycles. This is a change in emphasis from the “prompt engineering” training that dominated corporate AI upskilling through 2023 and 2024, and it reflects a maturing view of where the productivity gains actually sit.
What the frontier gap means for Australian environmental consultancies
There is no Australian-specific data in this OpenAI report, and the research does not address environmental regulation, contaminated land frameworks or planning law directly. Its relevance to Australian professional services firms, including environmental consultancies, engineering practices and legal advisers working in the contaminated land and planning space, is structural rather than regulatory. The finding that value is shifting from conversational AI use to agentic, tool-connected execution has direct implications for how Australian firms should be thinking about internal technology investment and workforce training.
Australian environmental consulting firms produce a high volume of repeatable, rules-based documentation, site assessment reports, waste classification determinations, regulatory cross-referencing against frameworks such as the National Environment Protection (Assessment of Site Contamination) Measure 1999 as varied in 2013, state EPA guidelines, and the PFAS National Environmental Management Plan. This is precisely the category of work OpenAI’s research suggests benefits most from agentic execution rather than conversational assistance, because it involves multi-step data retrieval, cross-checking against structured criteria, and iterative validation, not one-off question answering.
For Australian firms benchmarking their own AI maturity against this research, the practical question is not whether staff have access to a chat interface. It is whether the organisation has connected models to its own data systems, document templates and quality assurance checkpoints in a way that allows repeatable, auditable, multi-step work to run with reduced manual overhead. Firms that have not made that connection are, on this data, operating at roughly one-eighth the output intensity of frontier adopters, even if licence spend per employee looks similar on paper.
Practical steps for professional services firms adopting agentic AI
For firm leaders and technical directors, the immediate implication is that AI investment should be assessed against output, not access. Buying licences and running prompt-writing workshops places an organisation in the average cohort of this research, not the frontier one. The differentiating investments are integration work, connecting models to document management systems, project databases and report templates, and governance work, defining which task lifecycles can be delegated, where human review checkpoints sit, and how outputs are audited.
Quality assurance and professional liability considerations remain central. Agentic workflows do not remove the need for qualified professional sign-off; they change where that sign-off occurs. Firms adopting multi-step AI execution should set review checkpoints deliberately, rather than defaulting to reviewing every intermediate step, which erodes the productivity gain, or reviewing nothing, which creates unacceptable risk in regulated work such as site contamination assessment and statutory reporting.
Training budgets warrant a fresh look. On OpenAI’s data, spend directed at prompt technique is delivering diminishing returns compared with spend on infrastructure, tool connections and staff capability in supervising delegated work. For firms in regulated sectors, that supervision skill, knowing when an automated output needs escalation to a qualified professional, is likely to become a core competency rather than a niche one.
Finally, the 8.3x differential suggests timing matters. The report ind
References and related sources
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: 14 Aug 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.
Environmental management plans Contaminated land advice Environmental management Talk to iEnvi