Legal Services11 minNovember 17, 2026

AI ROI for Law Firms: The Gilbert + Tobin Case

AI ROI for law firms, measured: how Gilbert + Tobin runs agentic workflows, hit 87% adoption, and what CEOs can copy to cut routine work and approval cycles.

AI ROI for Law Firms: The Gilbert + Tobin Case

Why AI ROI for Law Firms Is Not About Fewer Lawyers

The cleanest proof that AI pays off in a law firm is that the firm's stated goal was never to cut people. The standard pitch says AI ROI for law firms comes from shrinking headcount and squeezing the billable hour. One of Australia's leading corporate firms says its CEO sees the value somewhere else entirely.

The numbers that follow come from the firm's own account of the rollout. They include an adoption rate that beats its usual software by a wide margin, a leadership review job that cost about $62, and a governance rule that decides whether any of it can scale. Read the case before you build your own business case.

What Gilbert + Tobin Actually Did

The source here is OpenAI's published interview with Sam Nickless, Partner and CEO of Gilbert + Tobin. (cite index="1-4">The firm is one of Australia's leading corporate law firms, with work spanning capital markets, mergers and acquisitions, disputes, and technology. It is a Sydney practice with, according to Wikipedia, more than 118 partners and about 1,100 employees. That is a mid-sized professional organisation, not a giant with an unlimited innovation budget.

Three phases, not one big launch

Nickless describes the journey in three phases: (cite index="1-11">experimentation, personal productivity, and organisational productivity. The sequence matters. The firm (cite index="17-5">began by letting staff experiment with the consumer version of ChatGPT under a strict rule that no client or confidential information be entered. Only then did it move to a secure enterprise environment. OpenAI says the firm has moved from open experimentation to (cite index="12-6">defined, multistep agentic workflows built on ChatGPT Enterprise, ChatGPT Work and Codex, with people still defining the task, checking the output and taking responsibility for the final work.

That last clause is the whole design philosophy. An agent does the legwork. A named human defines the task, checks the result and owns it.

The leadership move that no software can replicate

Nickless (cite index="10-5">introduced ChatGPT to employees through examples from his own work. Other senior leaders followed and shared what they did with their teams. The firm did not set usage quotas. It showed people what the tools could do in their own jobs, and its business transformation team built workflows for individual functions.

The result: (cite index="10-7">as of June 2026, 87% of enabled seats were active, more than twice the adoption rate the firm typically sees for other tools. The internal message was blunt. "AI is not cheating," Nickless told staff, according to coverage of the interview.

A tool nobody opens has an ROI of zero, whatever the vendor's slide deck promises.

Which tasks agents already took off human desks

The reported examples are operational, not courtroom drama:

  • Partner reviews. A confidential annual review process for all 120 partners was prepared in minutes with ChatGPT Work. Nickless put the cost at about $62 (the currency was not specified): the system updated files from spreadsheet data, produced PDFs and bundled them. He said the work would otherwise have required several people over multiple days. This figure comes from the CEO's own account in the OpenAI interview, as relayed by trade coverage.
  • Recruitment, audit and file work. Adoption spread into recruitment, finance, technology and business transformation, and Codex is used to move from individual assistance to defined multistep workflows. The firm has not published reliable before-and-after figures for these tasks in the source interview, so measure your own.
  • A CEO "digital twin." A custom GPT built from approved examples of Nickless's writing, priorities and feedback helps executives pressure-test ideas, without making decisions on his behalf or speaking for him.

Adoption spread from there into marketing, business development, recruitment, finance, technology, business transformation and parts of the legal practice.

Notice what is missing from the list: a robot arguing a case. The agents took the routine work that sits around legal advice. Partners and lawyers keep the judgment.

How AI ROI for Law Firms Gets Calculated Without Cutting Staff

Here is the part that breaks the usual spreadsheet. Nickless told OpenAI the value so far has not been about reducing cost or needing fewer people, but about freeing people up to do the more human side of their work. The firm's three stated objectives are (cite index="17-4">better client service, a stronger place to work and a leading position in its market.

Those goals sound soft. They are not. Each one converts to a number a CFO can audit.

Four lines of the return calculation

  1. Time released. Hours per task before and after, multiplied by volume. The partner-review run is this line: minutes of work and about $62 instead of several people over multiple days.
  2. Capacity redirected. What did the freed hours go into: billable work, client development, quality review? Time that disappears into nothing is not a return.
  3. Cost of errors avoided. Missed checks, wrong figures in a partner file, a conflict that slipped through. Onboarding and compliance checks are the obvious place to measure this.
  4. Retention and adoption. An 87% active-seat rate against a typical rate less than half that is a direct signal that licences are not shelfware. Idle seats are the quiet killer of any AI budget. Note that this figure counts active users among enabled seats, not necessarily daily or high-value usage.

For general AI ROI for law firms, the denominator matters as much as the numerator. The $62 run is a vivid number, but the full cost includes licences, governance work, training and the hours reviewers spend checking output. Count all of it, or your payback figure will flatter you and then embarrass you.

Why the billable hour is the open question

Not everything is settled. The effect of AI efficiency on the billable-hour model is not addressed in the source interview. If a task that took days now takes minutes, who captures the difference: the firm, the client, or both through pricing? The case does not answer that, and no honest ROI model can skip it. Your own model needs an explicit assumption about who keeps the savings.

The cheapest line in this rollout was the $62 run. The expensive one was earning permission to touch confidential data.

Comparison: three stages of AI maturity

The table below maps the firm's reported journey to the metric that belongs at each stage. The right-hand column is our suggested measure, not something the firm published.

Stage What the firm did Data rule Metric to track
Experimentation Staff tried consumer ChatGPT freely No client or confidential data allowed Number of use cases surfaced, ideas collected
Personal productivity ChatGPT Enterprise rolled out with CEO-led examples Enterprise environment, Australian data residency Active-seat rate, hours saved per task
Organisational productivity Multistep workflows in ChatGPT Work and Codex, human sign-off Contractual protections, role-based access, admin controls Cycle time per process, error rate, reviewer hours

The pattern: the further down the table, the more value per task, and the more you depend on controls. That is why the governance work is not overhead. It is the entry ticket.

Governance: The Limit That Decides the Return

The firm's boundary is confidentiality. It handles confidential client and business information, including information subject to different client requirements. Early experiments excluded client data altogether. Scale arrived only after the firm moved to an environment with Australian data residency, contractual protections, role-based access and admin controls.

Moving to that environment gave the firm greater confidence to expand access while meeting internal requirements and client expectations.

Two practical lessons for any executive:

  • Governance sets the speed limit. You can only widen access as fast as you can prove control. Firms that skip this step end up pulling tools back after the first incident.
  • Human accountability is the safety design, not a courtesy. If every output has a named owner, an agent's mistake becomes a review finding rather than a client crisis.

If you manage legal risk on the client side, the same logic applies in-house; we looked at it in what AT&T is building for its legal department. For the access layer, see how secure access for AI agents is being handled, and for oversight at scale, managing AI agents in the enterprise.

Be realistic about reliability, too. Agents still fail in messy real-world conditions, which is why the real-world performance benchmark for AI agents is a useful reality check before you promise the board anything.

A Step-by-Step Guide to Building Your Own Rollout

You do not need to be a law firm to use this. Any business with confidential files, approvals and compliance checks can follow the same arc. Treat it as a working checklist.

Step 1: Pick the sponsor and use the tool yourself

Gilbert + Tobin's adoption started with the CEO showing his own work. Choose one executive who will use the tool weekly and share real examples. Assign a single owner for the programme.

Step 2: Open a sandbox with a hard data rule

Allow experimentation, but write down what may never be entered: client data, personal data, contracts under confidentiality. Collect use cases for a fixed period, and use showcases or similar formats to surface ideas.

Step 3: Move to an enterprise environment

Before any sensitive work, check four things: where data is stored, whether your inputs are used for model training, how access is assigned by role, and what the admin can see. Get the contractual protections in writing.

Step 4: Select processes with a clear start and end

Good candidates are repetitive, document-heavy and checkable: periodic reviews, onboarding and compliance checks, audit packs, file preparation, recruitment research. Avoid starting with anything where the right answer is a judgment call.

Step 5: Define the workflow, then the human checkpoint

Write the task as a sequence: inputs, steps, outputs. Name who reviews the result and what "approved" means. The agent prepares; the person signs.

Step 6: Measure a baseline before you switch on

Record current time, error rate and cost for the process. Without a baseline, every later claim is an anecdote. After launch, track active seats, task time, review time and exceptions.

Step 7: Expand access only as controls mature

Gilbert + Tobin intends to extend access to ChatGPT Work and Codex as it puts in place the controls required for broader use. Follow that pace. Widen in rings: one team, then a function, then the organisation.

Done in this order, a leader gets something that rarely comes with technology projects: calm. You know who approved what, what the agent touched, and what it cost. That is control you can feel on a Monday morning, not just see in a dashboard.

FAQ

How do law firms calculate AI ROI? They compare hours per task before and after, then check where the freed time went, what errors were avoided, and whether people actually use the tools. Gilbert + Tobin's CEO frames value as freeing people for more human work, not as fewer staff, so capacity and quality count alongside cost.

What AI tasks can agents already do in a law firm? According to the reported case, they prepare partner review packs and support work across marketing, business development, recruitment, finance and technology. Codex is used for defined steps in larger workflows. Lawyers and staff still review and own the output.

Is client data safe in ChatGPT Enterprise? Safety depends on configuration and contracts, not on the product name alone. Gilbert + Tobin scaled use only after moving to an environment with Australian data residency, contractual protections, role-based access and admin controls. Check these same points for your own jurisdiction and clients.

Do usage targets help adoption? The firm reports that it did not mandate them. It relied on leaders showing examples and on role-specific workflows, and reached 87% active seats. That suggests relevance drives usage better than quotas.

Does AI replace lawyers? The firm's stated aim is better service and a better workplace, with people defining the task, checking the output and taking responsibility. The case gives no evidence of staff cuts as the goal.

What is the biggest risk when scaling AI in a professional services firm? Confidentiality and accountability. If you cannot show where data goes and who signed off each output, you cannot widen access safely, and the return stalls.

From Pilot to Predictable Profit

Gilbert + Tobin's story is a useful corrective to the cost-cutting script. The firm got its return by doing three unglamorous things in order: leaders using the tools, controls built before scale, and a human owner on every result. Your version will look different, but the sequence holds.

There is also a reputational payoff. A CEO who arrives at the board with an adoption rate, a baseline, a governance model and a cost-per-process figure is seen differently from one who arrives with a pilot and a promise. Investors read that as a leader who turns experiments into predictable economics, and your team reads it as someone who gives them better tools rather than a reason to worry.

If you want to see where agents would pay back first in your own approvals, compliance checks and reporting, book a 15-minute consultation and we will map the first three workflows together.

Have questions? Ask the AI agent right now

Responds in seconds, knows everything about our services and will help with your situation