How law firm Gilbert + Tobin governs and scales AI with OpenAI

This is OpenAI's own customer story, published by OpenAI with a Contact sales button on it, so the numbers are the firm's and the framing is the vendor's. Read accordingly. What it describes is a large Australian corporate law firm putting ChatGPT Enterprise into operations first — marketing, business development, recruitment, finance, technology, business transformation — and only then into parts of the legal practice.

The claimed results: 87 percent active usage among enabled seats as of June 2026, more than twice the firm's typical adoption for other tools; one recruitment research and data-extraction workflow down from roughly four hours to 20 minutes; selected conflict, KYC, AML and related checks down from three to eight hours to five minutes; and a day of manual work avoided preparing audit reports across 300 entities.

Strip the vendor packaging and the adoption mechanics are the story. The firm did not mandate usage targets. Its business transformation team went into individual team meetings and demonstrated workflows specific to each function rather than running generic training. CEO Sam Nickless introduced the tool with his own examples and the line "AI is not cheating." That sentence is doing more work than any feature in the product.

The deliberate sequencing matters too. Operations first, legal practice later, with the stated position that the goal is the work around legal advice rather than the advice itself. For a profession where a hallucinated citation is a career event, starting where errors are recoverable is the correct order of operations, and it is the part most firms get wrong when they lead with a drafting pilot.

Signal Desk's read: the 87 percent figure is the least transferable number here and the most quoted. Active usage among enabled seats measures who was given access and used it, not whether the work improved. The time savings are more interesting but they are self-reported single-workflow comparisons with no error rate attached — four hours to twenty minutes is meaningful only if someone checked what the twenty minutes produced. The firm's own framing, keeping people accountable for the result, implies that check exists. The case study does not show it.

What is genuinely portable: role-specific enablement beats generic training, visible executive use beats a policy memo, and no usage quota beats a quota. Those cost nothing and are the reason this deployment landed differently from the average enterprise rollout. The Codex step — moving from assisting individuals to completing defined steps across workflows — is where the governance claims will actually be tested, and there are no numbers on that yet.

Legal is among the most cautious enterprise AI buyers for obvious reasons, and Australian data residency plus a governance structure is what made the internal argument winnable here. Case studies like this one are marketing artifacts; their value is in the operational detail a vendor includes almost by accident, not in the headline metrics.