OpenAI Reports Codex Use Grew 108x in Legal Teams: AI Agents Have Moved Beyond Coding

OpenAI says weekly active enterprise Codex users in legal grew 108x since February 2026, far faster than engineering. For Sydney and NSW organisations, the significance is operational: agentic AI is moving from software development into document-heavy, approval-heavy knowledge work. Legal teams can now use agents for structured analysis, document handling, workflow execution and internal tooling, but confidentiality, human review, court rules, privacy and auditability remain non-negotiable.

The most interesting number in OpenAI's latest enterprise data is not coming from software engineers.

According to OpenAI's Enterprise Signals data, the number of weekly active enterprise Codex users working in legal functions has increased 108 times since February 2026. Sales and recruiting each grew 41 times, marketing grew 26 times and engineering, the function most closely associated with Codex, grew five times.

That does not mean legal productivity has increased 108 times. It does not establish that AI-generated legal work is more accurate, nor that legal judgement is being automated. The figure measures growth in weekly active users and almost certainly reflects expansion from a smaller starting base.

Yet the direction of travel is difficult to dismiss.

Codex began as a coding agent. It is now becoming an execution environment for people whose jobs are built around documents, systems, analysis, approvals, research and repeatable workflows.

For businesses buying AI services or comparing AI consulting firms in Sydney, that changes the implementation brief. The next useful enterprise agent may not be a chatbot and may not sit inside the technology department. It may work beside a solicitor, conveyancing team, compliance manager, property operator, finance analyst or project coordinator.

The 108x Figure Matters Because of What It Measures

OpenAI's latest enterprise reporting describes a wider transition from conversational AI towards delegated work.

As of June 2026, OpenAI says Codex accounted for 64 per cent of combined Codex and ChatGPT output tokens among enterprise customers. OpenAI defines the distinction operationally: conversational AI helps a worker ask questions and develop ideas, while agentic systems can use tools, work with files and carry out multi-step tasks under varying levels of supervision.

Separate OpenAI research into agentic work found that adoption has also moved rapidly beyond developers. Within OpenAI itself, legal, finance and recruiting teams shifted towards Codex as a primary AI working environment during 2026, while non-developer adoption across organisational users grew substantially faster than developer adoption.

The important development is therefore not simply that lawyers discovered a coding product.

It is that the boundary between coding and ordinary knowledge work is beginning to disappear.

A professional no longer needs to think, "I should commission a software application for this problem." They may instead describe the problem, provide structured context and ask an agent to manipulate files, interrogate data, build a small internal interface, transform information or execute a repeatable process.

That changes who can create operational software and how quickly a business can move from a recurring frustration to a working internal tool.

Why Legal Work Is Becoming an Agentic AI Test Case

Legal work looks highly specialised from the outside, but much of its operating environment contains exactly the characteristics that suit controlled automation.

• Large volumes of documents must be collected, named, classified and compared.
• Multiple versions of agreements need to be tracked accurately.
• Information moves between email, document systems, spreadsheets, matter-management software and external portals.
• Checklists need to be completed in a defined sequence.
• Missing information must be identified and chased.
• Standard information needs to be transformed into different internal formats.
• Exceptions must be escalated to a qualified person.
• A defensible record of what occurred can be more important than speed alone.

None of those activities requires an AI system to make the ultimate legal judgement.

They do, however, create significant scope for an agent to prepare the working environment around that judgement.

A Sydney conveyancing operation, for example, may receive a sale contract, strata documents, identity material, finance correspondence, agent emails, inspection information and settlement instructions across several systems. A carefully constrained agent could help organise those inputs, identify missing items, compare versions, prepare an internal chronology and update the matter workflow.

The solicitor still determines the legal significance.

That distinction is likely to become central to enterprise AI design: automate preparation aggressively where it is safe, but preserve accountable professional judgement where it matters.

The AI Consulting Brief Has Changed

Many early enterprise AI projects started with a technology question: which model should the organisation buy?

Agentic adoption creates a different set of questions.

Which model should we use?

• Agentic AI implementation question: Which part of the workflow should the agent be authorised to perform?

Can it answer questions?

• Agentic AI implementation question: Which systems, documents and tools can it safely access?

How good is the prompt?

• Agentic AI implementation question: What business context, rules and permissions persist across the task?

Can it draft the output?

• Agentic AI implementation question: Can the result move safely into the next operational step?

Does a human review it?

• Agentic AI implementation question: Exactly which events require approval, escalation or rejection?

Did the demonstration work?

• Agentic AI implementation question: Can the workflow be monitored, audited, recovered and improved in production?

This is why organisations evaluating AI consulting in Sydney increasingly need more than model expertise.

An implementation team needs to understand process mapping, permissions, system integration, identity, document control, exception handling, operational ownership and measurable business outcomes.

The intelligence layer matters. The workflow around it often determines whether the system survives contact with the real organisation.

Where Sydney Legal Teams Could Use Agents Without Automating Legal Judgement

The strongest early opportunities are likely to sit one or two steps away from the final professional decision.

1. Matter Intake and Information Preparation

An agent can check an incoming matter against a defined intake structure, organise supplied documents, identify incomplete fields and prepare an internal summary for review.

The productivity gain comes from reducing the time professionals spend turning unstructured material into a usable matter file.

2. Document and Version Comparison

Legal operations frequently involve determining which document is current, what changed between versions and whether supporting records correspond with the latest instructions.

Agents can assist with structured comparison and produce an issue list, while the responsible practitioner decides whether any change has legal significance.

3. Due Diligence Preparation

An agent can assemble a review pack, index documents against predefined categories, flag missing categories and prepare a structured evidence map.

This changes the economics of document-heavy work because the expensive professional is presented with a better prepared environment rather than spending the first part of the engagement constructing it.

4. Internal Operational Tools

The expansion of Codex beyond programmers is particularly important here.

A legal operations team may not need to wait for a conventional software development project to create a small reconciliation tool, document processor, matter dashboard, data transformation utility or internal workflow interface.

Agent-assisted development can compress that distance, provided the software still goes through appropriate security, testing and change-control processes.

5. Transaction Coordination

In property and conveyancing environments, administrative coordination can involve clients, agents, brokers, banks, strata managers, inspectors and settlement platforms.

An AI agent can help prepare status summaries, track outstanding inputs and route administrative follow-up. It should not independently authorise exchange, provide unreviewed legal advice or make a legal commitment simply because the underlying technology is capable of taking actions.

Organisations considering this type of integration should assess the underlying process first. Elyment's approach to workflow automation for Sydney operations teams starts with the work moving between people and systems rather than treating the AI model as the project itself.

NSW Creates a Clear Boundary Between Operational Assistance and Professional Responsibility

The legal sector is useful precisely because it demonstrates why greater technical capability does not remove professional responsibility.

The Law Society of NSW maintains dedicated AI resources for legal professionals, including responsible-use guidance, court protocols and law-practice resources. Its guidance continues to emphasise confidentiality, competence, accuracy and the practitioner's responsibility for work produced with AI assistance.

NSW courts have also established formal protocols. The Supreme Court of NSW's generative AI framework includes Practice Note SC Gen 23 and related guidance dealing with the use of generative AI in court proceedings.

That means a legal AI implementation cannot be designed solely around whether a model is technically capable of producing an output.

The workflow must recognise what happens to that output next.

A system used for internal document organisation is different from a system whose output may enter evidence, a court document, legal advice or another professionally significant process.

The control boundary should follow the consequence of the action.

Confidentiality Turns Data Architecture Into an Operational Issue

Legal teams routinely handle personal, commercial and confidential information. Agentic AI therefore cannot be treated like another isolated productivity application.

The Office of the Australian Information Commissioner's guidance on commercially available AI products advises organisations to conduct due diligence, understand how personal information will be handled, embed appropriate human oversight and consider privacy throughout the system lifecycle.

For an agent, that review needs to extend beyond the model itself.

The organisation should understand:

• which repositories the agent can search;
• which matters or clients it can access;
• whether permissions mirror the underlying user's permissions;
• what information can leave the organisation's controlled environment;
• which external tools or plugins can receive data;
• how long logs and outputs are retained;
• how incorrect outputs are identified and corrected;
• how access is revoked when staff roles change; and
• what evidence exists after the agent performs an action.

Agentic AI turns identity and access management into part of workflow design.

Cyber Security Becomes More Important When AI Can Act

The security model also changes once AI moves from generating text to interacting with operational systems.

In May 2026, the Australian Signals Directorate and international cyber-security partners published guidance on the careful adoption of agentic AI services. The guidance emphasises constrained privileges, strong identity controls, monitoring, human oversight and incremental deployment rather than broad access to sensitive systems.

That principle is particularly important for legal, property, finance and other regulated workflows.

An agent does not need unrestricted access simply because it occasionally needs to retrieve information from several places.

Good system design gives it the smallest useful operating envelope.

The Most Important Design Question Is Where the Agent Stops

AI consulting firms working with Sydney organisations should be able to answer a deceptively simple question before production deployment:

What is this agent explicitly not allowed to do?

That answer should be operational rather than philosophical.

• It may prepare a contract comparison but cannot determine whether the client accepts the amendment.
• It may identify an incomplete matter file but cannot invent the missing information.
• It may draft a client update but cannot send specific categories of communication without review.
• It may update internal workflow status but cannot authorise settlement.
• It may retrieve approved knowledge but cannot browse unrestricted repositories.
• It may create an internal utility but cannot deploy production code without testing and release controls.
• It may recommend an escalation but cannot override the accountable professional.

These boundaries are not evidence that the technology is weak.

They are evidence that the organisation understands where authority sits.

A Practical Sydney Rollout Should Start With One Workflow

The 108x growth figure may tempt organisations to expand quickly. A more defensible approach is to treat agentic AI as operational infrastructure and scale it through evidence.

1. Map the current process.
2. Record how work actually moves through email, documents, matter systems, spreadsheets, portals and staff rather than how the procedure manual says it moves.
3. Measure the existing bottleneck.
4. Establish baseline cycle time, queue time, manual touches, rework, errors and escalation volume.
5. Separate preparation from judgement.
6. Identify the steps that can be delegated and the decisions that remain with an accountable professional.
7. Define the agent's operating boundary.
8. Specify approved data, systems, actions, approval points and prohibited activities.
9. Test realistic exceptions.
10. Do not test only clean examples. Use duplicate files, outdated documents, conflicting instructions, missing data, unusual matter types and unavailable systems.
11. Launch with observability.
12. The business should be able to see what the agent did, why a task stopped and when a human intervened.
13. Measure the whole workflow.
14. Faster AI output is irrelevant if it creates a larger human review queue downstream.

Businesses at an earlier stage can use an AI readiness assessment for Sydney operations to identify which processes have the data quality, repeatability and ownership needed for a serious deployment.

The Metric That Matters Is Not How Much AI Gets Used

Enterprise AI reporting naturally focuses on adoption because adoption can be measured across very large populations.

Individual businesses need a more demanding set of metrics.

End-to-end cycle time

• What it reveals: Whether the complete process became faster, not merely the AI step.

Professional review time

• What it reveals: Whether the agent reduced or increased the burden on senior staff.

Rework rate

• What it reveals: Whether speed is being achieved at the expense of quality.

Exception rate

• What it reveals: How frequently the workflow leaves its intended operating envelope.

Human escalation quality

• What it reveals: Whether staff receive enough context to resolve the exception efficiently.

Cost per completed workflow

• What it reveals: Whether automation creates a genuine economic improvement.

Audit completeness

• What it reveals: Whether significant actions remain explainable after the event.

This is where many agent demonstrations encounter their first operational reality.

Generating a draft in 30 seconds does not create a 30-second process if a senior employee then spends 15 minutes validating it.

The objective should be to redesign the full workflow, not simply accelerate the visible AI moment.

What Buyers Should Expect From AI Consulting Firms in Sydney

The expansion of Codex into legal teams reinforces a broader change in the AI services market.

Model access is becoming easier. Implementation quality is becoming more valuable.

A serious enterprise engagement should be able to connect six disciplines:

• workflow design, understanding the actual sequence of work;
• AI capability, selecting the appropriate model and agent architecture;
• integration, connecting approved systems and data without creating another isolated interface;
• security and privacy, controlling identity, information access and external connections;
• operational governance, defining approvals, escalation, monitoring and change control; and
• business measurement, proving whether the completed process improved.

That is increasingly the difference between an AI demonstration and a production operating system.

Elyment's broader AI systems and software development services in Sydney are structured around that implementation problem: connecting AI, workflows, integrations, human review and operating controls rather than treating the model as a standalone product.

The Australian Government's current responsible-adoption framework points in the same direction. Its Guidance for AI Adoption treats governance as something that should apply both across the organisation and to individual AI systems according to the risks created by each use case.

Design the Workflow Before the Agent Gets Authority

Review workflow structure, data access, integrations, approval boundaries, compliance considerations and measurable delivery outcomes before moving an AI agent into production.

Request an AI Project Review

The Bigger Signal Is That Coding Is Becoming a Way to Operate, Not a Job Description

The 108x figure should therefore be read carefully.

It is not evidence that legal teams are being replaced by software. It is not proof of a 108-fold productivity gain. It is not evidence that professional judgement can safely be delegated to an autonomous system.

It does show that a tool originally identified with developers is being adopted rapidly by people whose primary work is not software engineering.

That may prove to be the more consequential shift.

For decades, most employees interacted with software that somebody else had designed around a predefined process. Agentic AI increasingly allows workers to instruct software to investigate, transform, build and execute around the work in front of them.

Legal teams are an unusually important test because their workflows combine valuable knowledge, sensitive information, strict professional obligations, complex documents and clear consequences when something goes wrong.

If agentic systems can become useful there while remaining properly constrained, Sydney businesses should expect the same operating model to spread through finance, property, procurement, project delivery, compliance, construction administration and other knowledge-intensive environments.

The enterprise AI question is consequently moving beyond, "What can the model answer?"

The more important question is becoming, "What work can we safely delegate, what authority does the agent need and where must a person remain responsible?"

Sources and References

• OpenAI: Enterprise Signals data
• OpenAI: How agents are transforming work
• Elyment: AI consulting in Sydney
• Elyment: Workflow automation for Sydney operations teams
• Law Society of NSW: AI resources for legal professionals
• Supreme Court of NSW: Generative AI framework
• Office of the Australian Information Commissioner: Privacy and commercially available AI products
• Australian Signals Directorate: Careful adoption of agentic AI services
• Elyment: AI readiness assessment for Sydney operations
• Elyment: AI systems and software development services in Sydney
• Australian Government: Guidance for AI Adoption
• Elyment: Contact