A New ChatGPT Enterprise Study Analysed 1,500 Organisations and 17 Million Messages: What Real AI Adoption Looks Like

A new study using ChatGPT Enterprise data from more than 1,500 organisations and 17 million messages suggests real AI adoption is not defined by licences, pilots or executive enthusiasm. Usage spreads unevenly across roles, tasks and seniority. For Sydney and NSW organisations, the practical lesson is to measure where AI is repeatedly used, what work it changes, who depends on it and whether that activity produces a controlled business outcome.

The enterprise AI market has spent much of the past three years measuring progress through announcements: how many employees received access, which model was purchased, how many agents were deployed and how much money companies planned to invest.

A new working paper offers a more revealing view.

How Organizations Use AI: Evidence from ChatGPT, authored by researchers affiliated with OpenAI, Columbia Business School and the Wharton School, examines privacy-preserving ChatGPT Enterprise usage data through March 2026.

At its six-month organisational adoption horizon, the worker-level sample covers more than 1,500 organisations and more than 17 million messages. The researchers connect usage patterns with worker roles, task classifications and, for a subset of US public companies, financial information.

The findings complicate the conventional enterprise AI story. Adoption is growing quickly, but not uniformly. Employees at different levels use AI differently. Some functions use it far more intensively than others. Common tasks span writing, communication, technical work and information synthesis rather than converging on one dominant enterprise workflow.

Most importantly, the researchers caution against interpreting adoption as an immediate productivity transformation.

That distinction matters for Sydney businesses now moving from AI access into operational deployment.

The First Mistake Is Measuring Licences Instead Of Use

An organisation can purchase 500 enterprise AI seats and still have weak AI adoption.

Another business may have a much smaller user base but several teams using AI repeatedly inside valuable work. The second organisation may be further along operationally despite appearing smaller on a procurement dashboard.

The study illustrates why.

Aggregate output tokens from ChatGPT Enterprise customers increased roughly sevenfold between June 2025 and March 2026. This was not only the result of new organisations joining. Among firms that had already adopted by June 2025, output increased roughly fourfold over the same period.

In other words, adoption deepened after purchase.

That should change how management teams think about AI implementation. The first contract date is not the adoption event. It is the beginning of an organisational learning period.

Sydney businesses conducting an AI readiness assessment should therefore distinguish between at least four different questions:

• How many employees have access?
• How many employees actually use the system?
• How intensively do active employees use it?
• Which recurring business tasks account for that usage?

Those measures describe very different operating conditions.

Real Adoption Is Broad, But It Is Not Evenly Distributed

One of the more commercially important findings is that enterprise AI does not appear to remain trapped inside technology teams.

Six months after organisational adoption, active users were distributed across engineering, management, finance, marketing, sales, executive and other functions.

Yet participation and intensity were not the same thing.

At the average organisation in the researchers' sample, managers and directors represented approximately 24 per cent of observed weekly active users. Individual contributors and professionals represented around 15 per cent, senior individual contributors and principals around 14 per cent, executives around 10 per cent and early-career workers and trainees around 7 per cent.

But the smaller early-career group used the system particularly intensively.

Among active users, early-career workers and trainees sent roughly eight to nine more messages per week than the average user inside the same firm. Analysts and marketing and communications workers also displayed comparatively high usage intensity, while executives generally sent fewer messages.

This creates an important management problem.

The people who approve AI expenditure may not be the people discovering its most frequent uses.

A board or executive team can therefore misunderstand adoption if it looks only at senior leadership behaviour. The practical innovation may be occurring lower in the organisational structure, where employees repeatedly perform research, prepare documentation, compare information, draft communications and complete technical tasks.

The Junior Employee May Be The Enterprise AI Laboratory

The seniority finding deserves particular attention.

Early-career workers often spend more of their day producing first drafts, researching unfamiliar subjects, reorganising information, preparing documentation, responding to requests and translating instructions into completed work.

Those activities overlap strongly with current generative AI capabilities.

That does not establish that junior employees are generating greater economic value with AI. The study measures usage intensity, not productivity, profitability or output quality.

It does suggest, however, that management teams looking for useful AI workflows should observe where frequent use is already emerging rather than assuming transformation must begin with an executive-designed use case.

A Sydney property or renovation operator, for example, may discover frequent AI use among project administrators rather than senior project managers.

Those administrators may repeatedly use AI to:

• Structure incoming job information
• Summarise supplier or contractor correspondence
• Prepare project updates
• Compare quotation information
• Organise photographs and scope notes
• Identify missing information before scheduling
• Prepare handover records
• Draft routine client communication for review

None of these activities necessarily justifies a fully autonomous agent.

Together, however, they can reveal where a formal workflow automation program may remove significant administrative friction.

The Most Important Enterprise AI Use Cases Are Surprisingly Ordinary

Another striking finding is the breadth of tasks.

Enterprise AI usage was not concentrated in one breakthrough application. Documentation and technical writing, technical digital work and communication were among the most common categories across industries.

Other activities included business and market research, information retrieval, planning, data analysis, legal and regulatory work, financial tasks, sales and marketing work and topic overviews.

Task mix also reflected the underlying responsibilities of different functions. Engineering and technical workers showed more technical digital activity. Finance workers were more likely to use AI for financial and tax-related tasks. Sales and marketing roles were more likely to use it for their corresponding commercial work.

This suggests that enterprise AI may scale less like the rollout of one new department and more like a horizontal capability attached to existing jobs.

The accounting team does not become an AI team. The project team does not become an AI team. The marketing department does not become an AI team.

Instead, each function gradually inserts AI into a different part of its existing workload.

That Creates A Measurement Problem For Management

When adoption is dispersed across hundreds of small tasks, traditional transformation reporting can miss what is happening.

A company may not have one dramatic AI project capable of producing a board-level case study. It may instead have hundreds of employees each removing 10, 20 or 40 minutes of friction from different parts of the working day.

That activity still needs to be measured.

• Licensed users
• What it tells management: Potential organisational reach
• What it does not prove: That employees are actually using AI
• Weekly active users
• What it tells management: Breadth of active participation
• What it does not prove: That use is frequent or commercially important
• Messages per active user
• What it tells management: Intensity among employees already using AI
• What it does not prove: That more messages equal better productivity
• Task distribution
• What it tells management: Where AI is entering existing work
• What it does not prove: Whether the output is accurate or valuable
• Controlled business outcomes
• What it tells management: Whether AI improves time, quality, cost or delivery
• What it does not prove: That every surrounding process should be automated

This distinction is central to mature AI consulting and implementation planning.

Usage analytics should identify candidates for operational improvement. They should not be mistaken for the improvement itself.

A Sydney Project Team Shows Why The Difference Matters

Consider a property services business managing renovation projects across Sydney.

The visible physical job may involve carpet removal, timber removal, tile removal, concrete grinding, floor levelling, installation, painting or another trade package.

Around that physical work sits an information system.

The business may need to coordinate:

• Customer enquiry details
• Site photographs
• Measurements and substrate conditions
• Strata requirements
• Building access and lift bookings
• Noise restrictions
• Quotations and variations
• Deposits and payment milestones
• Supplier lead times
• Contractor availability
• Site sequencing
• Completion photographs
• Handover communication

An employee might use ChatGPT dozens of times across that process without the business having a formal "AI renovation project".

That organic behaviour can be useful evidence.

If staff repeatedly use AI to turn incoming emails into structured scope notes, the business may have discovered an intake automation opportunity.

If AI is repeatedly asked to find missing access information, the problem may actually be the project record.

If staff repeatedly use AI to reconcile customer, supplier and scheduling information, the organisation may need better integration between its CRM, project system and calendar.

If employees repeatedly rewrite poorly structured internal information before it can be sent externally, the underlying workflow may need redesign.

This is why business process automation should begin with observed work, not technology enthusiasm.

High Usage Can Reveal A Good Workflow Or A Bad One

Management should also resist assuming that heavier AI usage is automatically desirable.

Twenty prompts used to complete a task may indicate strong adoption. They may also indicate fragmented records, poor instructions or repeated correction.

A useful adoption review therefore needs to distinguish between productive repetition and compensating behaviour.

Consider three employees using AI heavily.

1. Employee A uses AI to prepare a consistent project summary from approved records, reducing a 30-minute administrative task to five minutes.
2. Employee B repeatedly asks AI to locate information because project records are scattered across email, messages and personal folders.
3. Employee C repeatedly regenerates client correspondence because the AI does not have access to approved service rules.

All three may appear as highly active users.

Only the first example clearly demonstrates an efficient operating pattern.

The second reveals an information architecture problem. The third suggests a knowledge and governance problem.

Real adoption analysis therefore requires operations knowledge, not only platform analytics.

The Study Does Not Prove An AI Productivity Boom

This is where the research needs careful interpretation.

The working paper explicitly notes that its task data does not measure downstream work products, productivity improvements or changes to organisational routines.

Its financial analysis also requires caution. Early ChatGPT Enterprise adoption among the US public-company sample was concentrated among larger, more valuable companies with greater investment in areas such as research and development and selling, general and administrative activity.

Those relationships are associations, not evidence that using ChatGPT caused stronger financial performance.

Early adopters may simply have more of the organisational capabilities required to experiment effectively with new technology: larger technology budgets, stronger internal systems, more knowledge workers, more management capacity and greater previous investment in digital processes.

The researchers' broader conclusion is more restrained and more useful: adoption is only the beginning of deployment.

Businesses are still learning where generative AI belongs.

For NSW Organisations, Governance Has To Follow The Actual Use

Decentralised adoption creates a second challenge.

Once employees across functions start finding their own AI use cases, governance cannot exist only in a procurement document.

The NSW AI Operational Policy, which applies to NSW Government agencies, illustrates the direction of mature governance. It emphasises accountability, use-case registration, risk assessment, governance, AI literacy and lifecycle oversight.

Private NSW businesses are not automatically subject to that government policy. The operating principle is nevertheless instructive: organisations need visibility into how AI is actually being used.

The NSW AI Assessment Framework similarly uses a risk-based approach to AI design, deployment, procurement and use.

Privacy creates another layer. The Office of the Australian Information Commissioner: Guidance on privacy and the use of commercially available AI products advises organisations using commercially available AI products to assess privacy risks and their obligations where personal information is involved.

For Sydney organisations handling customer records, property addresses, contracts, employee information, photographs, access instructions or financial information, AI adoption therefore has to be mapped against data handling and responsibility, not just productivity.

The Better AI Dashboard Measures Organisational Learning

The study points towards a more useful executive dashboard.

Instead of reporting only seats and monthly prompt volume, management should be able to see:

1. Adoption breadth: What percentage of intended users are active?
2. Usage intensity: Which teams use AI repeatedly rather than occasionally?
3. Task concentration: Which business activities account for recurring usage?
4. Workflow maturity: Is the activity still ad hoc prompting, or has it become a controlled process?
5. Human review: Where is approval still required and how much review time remains?
6. Exception rate: How often does AI produce incomplete, unusable or escalated work?
7. Outcome improvement: Has cycle time, administrative effort, quality, conversion, response time or project delivery actually improved?
8. Risk exposure: What customer, employee, financial, privacy or operational information is involved?

The objective is not to maximise every measure.

It is to understand whether employee experimentation is becoming dependable organisational capability.

What Sydney Businesses Should Learn From 17 Million Messages

The most important conclusion from the research may be what it does not show.

There is no single enterprise AI workflow.

There is no single worker profile responsible for adoption.

There is no reliable reason to equate heavy usage with productivity.

And there is no evidence that purchasing access automatically reorganises the company.

Instead, enterprise AI appears to diffuse through existing work. Different employees discover different uses. Certain roles use it more intensively. Organisations deepen their usage after initial adoption. The task mix remains broad.

That makes the next stage an operating challenge.

Sydney and NSW businesses need to observe where AI is already helping, identify where employees are repeatedly compensating for weak systems, formalise the workflows worth keeping and apply controls proportionate to the information and decisions involved.

The competitive advantage is unlikely to come from accumulating the largest number of prompts.

It will come from converting useful employee behaviour into faster, clearer and more reliable organisational processes.

AI Adoption & Operational Review

Review What AI Usage Is Actually Changing Inside The Business

Review employee use cases, workflow dependencies, data access, approval points, privacy considerations, performance measures and implementation priorities before informal AI activity is scaled into business-critical operations.

Request A Project Review

Real AI Adoption Starts After The Software Is Purchased

The new ChatGPT Enterprise study provides a rare view inside workplace AI usage at scale. Its message is more complicated than the idea that businesses are simply adopting AI faster.

Adoption is deepening, but it is uneven. Usage extends across the organisation, but intensity differs substantially. Workers use the same technology for different tasks. Early-career employees can become some of the heaviest users. Large organisations may buy first without necessarily achieving the highest use per employee.

The research is also a working paper, its results may change, its public-company financial analysis is US-focused and it does not establish Sydney-specific adoption rates or prove productivity gains.

What it does provide is a better way to frame the enterprise AI transition.

Access is not adoption. Activity is not productivity. Automation is not transformation.

The real test is whether an organisation can recognise the useful patterns emerging across its workforce and turn them into controlled, measurable and dependable ways of working.

Sources And Further Reading

• How Organizations Use AI: Evidence from ChatGPT
• Digital NSW: NSW AI Operational Policy
• Digital NSW: NSW AI Assessment Framework
• OAIC: Guidance on privacy and the use of commercially available AI products
• Elyment: AI Readiness Assessment Sydney
• Elyment: Workflow Automation Sydney
• Elyment: AI Consulting Sydney
• Elyment: Business Process Automation Sydney
• Elyment: Contact