Law firm Gilbert + Tobin reaches 87% active ChatGPT use without usage KPIs

Most enterprise AI rollouts run into the same problem: licenses are purchased, but only a handful of people use them. Australian law firm Gilbert + Tobin, whose work spans IPOs, M&A and litigation, told OpenAI in an interview how it achieved 87% active usage among employees given access to ChatGPT. That is more than twice the rate for the firm's other tools. An important caveat: this account appeared on the vendor's blog, so the figures should be read as the firm's own claims.

Three phases instead of a single switch

The firm followed three stages: experimentation, personal productivity and organizational productivity. At the first stage, employees were allowed to try ChatGPT under a strict rule: no client data or confidential information. This created a safe sandbox without stifling curiosity.

The second step was a secure channel. The firm introduced ChatGPT Enterprise in a controlled environment and gradually expanded access. Australian data residency, contractual guarantees, role-based access and administrative controls proved decisive. For lawyers, these are not bureaucratic formalities: they are prerequisites for clients to accept AI in their work.

The third stage involved Codex and ChatGPT Work: a shift from one-off assistance to multistep agentic workflows, with a human responsible for review and approval. The principle is straightforward: AI is a tool, not cheating. But the employee defines the task, checks the output and remains accountable for it.

Why people actually use it

No usage targets were imposed. Instead, the business transformation team demonstrated workflows for specific functions, ran videos, demos and challenges, and even offered an early “bounty” for ideas. Leaders used AI themselves to set the tone.

For product teams, the logic is familiar: adoption grows when a product solves a problem for a specific role, rather than when management demands that people “use AI.” Showing a lawyer a ready-to-use technique for their working day is more effective than a general presentation about a model's capabilities.

A digital twin and a $62 bill

CEO Sam Nickless built a “digital twin” in a custom GPT, which he is now moving to Plugins. He uploaded his writing, feedback he had received, context about the firm and a description of his working style. It started as a personal tool, before access was extended to his direct reports. They test their ideas against his context before meeting him. The twin makes no decisions and does not speak for the CEO; it simply helps them enter the conversation better prepared.

Notice the sequence: he tested the assistant himself before sharing it with the team. The same approach works for product teams: build an assistant from real documents, test it on your own tasks, then share it.

The second example is an annual confidential review of 120 partners. ChatGPT Work updated files with data from a spreadsheet, created PDFs and assembled the final package. It took minutes. The usage dashboard showed a cost of around $62, whereas previously several people would have spent several days on the work. That amount excludes the time spent checking the output and setting up the process, and it is the CEO's own estimate.

Limitations and takeaways

This is a case study, not a research study. It provides no data on quality, errors, the proportion of work that had to be redone or changes in lawyers' workloads. The claim that AI does not reduce headcount is the leader's public position. The “active users” metric indicates frequency, not value. Furthermore, custom GPTs are being replaced by Plugins, so workflows will need to be rebuilt: dependence on a single vendor's features is a risk worth factoring into plans.

What to take away. First, start with a safe sandbox and clear data rules, then move to enterprise access with roles and auditing. Second, provide ready-to-use workflows tailored to roles instead of usage KPIs. Third, track the cost of agentic tasks in the dashboard and compare it with person-hours, making the return-on-investment discussion concrete. Fourth, keep a human accountable for the result, especially where mistakes are costly.