AI training for the whole firm

AI training in a firm can start with a few people. The person who looks after IT goes along to a vendor session, or a colleague who has been experimenting shows the team what AI can do with a file note or a client email. Those first sessions are generous, and they show people what is possible with their own kind of work.

What we tend to see is that the confidence stays with the people closest to the tools. The rest of the firm has seen a demonstration, and a few months on, it may have faded. AI training brings the whole firm along, from the front desk to the partners, so everyone starts from the same place.

1What AI training covers

AI training is about the tools: how to use the ones the firm has chosen, and how to get better with them. It sits alongside the wider work of building AI skills across the firm, which covers understanding AI and where it fits.

The Law Society of NSW's guide to responsible use of AI, published in January 2026, asks firms to consider ongoing professional development so that "all practitioners have the skills and knowledge to" prompt the tool appropriately, evaluate its common limitations, and supervise its outputs and their accuracy.

In a firm, that covers which tools people can use and for what work, how to ask for what they need with the right context from the file, and what stays out of the tool. It also covers how to read what comes back, and what to do when it is wrong. The tools keep changing, and people get better with them through use, so AI training needs to be revisited.

2Levels of training

In the 2025 Victorian Lawyer Census, published by the Victorian Legal Services Board and Commissioner in April 2026, the lawyers who already used AI were more interested in applied skills: 73% of them wanted training in prompt writing, against 61% of non-users. Non-users were more interested in the foundations, such as understanding how AI works (65%, against 53% of users) and safe and ethical use (73%, against 66%). The census heard from 1,887 Victorian lawyers.

We usually set AI training up at three levels.

• A foundation for the whole firm. The firm's tools, what they do well, where they go wrong, what stays out of them, and how the work gets checked. It is the same for the receptionist and the managing partner.

• Applied training by team. How AI fits the work a team does every day. A property team, a tax team and the accounts team may use different tools, or the same tool in different ways.

• Deeper training for the people who want it. Prompting for longer jobs, building templates the team can reuse, and helping colleagues along.

In the Australian Public Service, the Digital Transformation Agency announced in January 2026 that "Foundational AI training will now be mandatory for all staff across the APS, building a consistent baseline of understanding of responsible AI use." Its staff training guidance encourages agencies to add training for staff "in consideration of their roles and responsibilities".

3Who gets trained

In the 2026 ALPMA and Dye & Durham survey of 555 law firms in Australia and New Zealand, 44.5% said their firm has people who can use AI effectively. (Dye & Durham, a legal software provider, co-produced the survey.)

In the Victorian census, half of the lawyers who responded had received AI training in the past year: 72% of those who used AI and 38% of those who did not. The census can't tell us which came first, the training or the use.

A 2026 trial at the University of Hong Kong looked at what happens when people have AI without training. Chen and Bao gave 164 law students an issue-spotting exam. Some had no AI, some had AI with no training, and some had AI plus a brief training session. Access without training "proved counterproductive": those students wrote shorter answers and misstated more cases, although most of those differences fell short of statistical significance. The trained group took up the tool more often (41% against 26%), scored higher and stated the rules more accurately. Without training, the authors found, the stronger students tended to opt out, while the weaker ones used the tool without much benefit. The study has not yet been peer reviewed.

In Xero's 2026 survey of 524 senior accountants and bookkeepers in Australia, the top-performing practices were more likely to treat high tech fluency as a mandatory baseline (51% against 28%). It is Xero's own vendor survey, and part of the sample came from Xero's customer list.

With a shared foundation, the person who has been using AI for a year and the person who is new to it both know the firm's tools, what goes in them and how the work gets checked. From there, the people who want more can go further, and their colleagues can follow and use what they build, so the benefit spreads across the whole team.

4Knowing when to check

In a study published in Organization Science in March 2026, Dell'Acqua and colleagues gave 758 BCG consultants a set of realistic tasks. On one task chosen because AI handled it poorly, the group without AI got it right about 84.5% of the time. The groups using AI scored 70.6% and 60%, and the 60% group was the one that had also been given an overview of prompt engineering. That is why the foundation needs to cover where the firm's tools are weak, what a wrong answer can look like, and when to stop and check.

The Tax Practitioners Board said in July 2026 that "It is important for tax practitioners to understand the capabilities and limitations of AI tools, and ensure AI outputs are assessed and supplemented by professional judgement, before being relied on." The federal government's Guidance for AI adoption asks organisations to make sure people using or overseeing AI understand "each system's capabilities, limitations and failure points, and when to intervene."

In the Victorian census, 89% of those who used AI said they took steps to check its accuracy, and 76% cross-checked the output against other sources. Human in the loop covers who checks, and at which point.

5The forgetting curve

A single session can introduce all of this, and then people have to remember it.

In the 1880s, Hermann Ebbinghaus measured how fast he forgot lists he had learned, by timing how long it took to learn them again. Murre and Dros repeated his experiment in 2015 and got a similar curve, with the steepest drop inside the first day. It is foundational research with one person learning nonsense syllables, so treat it as an illustration of how forgetting works.

In a foundational review of 317 experiments, mostly on remembering words and facts, Cepeda and colleagues found that only 12 of 271 comparisons showed no benefit from spacing practice out over time, against doing it all at once. The longer something needed to be remembered, the wider the best gap between sessions.

In a 2026 study of workers in the United States and Europe, Bick and colleagues found that encouragement from the employer was the strongest predictor of whether people used AI, and training did not predict higher use once encouragement and access to tools were taken into account. In Xero's survey, the most common answer on how best to build capability was self-led experimentation during work time (56%), ahead of formal training courses (39%) and peer-led coaching (35%).

The evidence points towards revisiting the learning, giving people time to use it on their own work, and making AI use something the firm actively supports.

6Setting up AI training that lasts

7Where we come in

Our investigation looks at how your people use AI today and measures how long your everyday work takes, so you start from a baseline of your own. The findings belong to your firm whether or not we go any further together.

From there, every person in the firm completes AI Foundations, the short training that gives the whole firm the same starting level. Cohort workshops follow: a full day, with no more than 10 people in a room, grouped by team or skill level so the work in front of them is their own. AI Champions, one person per team, can keep it going after we leave. Through the first year we stay in touch with check-ins and a resource library, and task times are measured again at three months against the baseline from the investigation.

If you would like to talk through AI training for your firm, reach out to our team. The first conversation works out whether your firm is the right fit for the AI Uplift Program, and whether we are the right people to help.

Setting up AI training that lasts