Guest post by Dr Peter Fields and Michaela Hardwick

Peter Fields
A study published in Harvard Business Review on 23 July 2026, by Ashish Agarwal, Anitesh Barua, Anu Puvvada, Fangchen Song and Wen Wen, has attracted quite a bit of attention. It is worth the attention.
KPMG and the University of Texas at Austin ran a field study of 523 early-career professionals at KPMG in the US, each with 18 months or less at the firm. Every participant completed tasks built to mirror real client work, using an AI agent designed for their domain.
The researchers first let the AI complete the work entirely alone, to set a baseline. They then measured how each person performed once given the same AI to work alongside.
The results split into three groups. Just over half of participants, 50.1%, outperformed the AI working alone. The researchers called this group ‘AI Amplifiers’. A further 25.8%, the ‘AI Delegators’, produced work roughly equivalent to the AI baseline. The remaining 24.1%, the ‘AI Apprentices’, performed below what the AI achieved on its own.
Here is the part we think should matter to every one of us, whatever stage we are at in our career. The researchers checked whether critical thinking, domain knowledge or AI literacy explained which group someone fell into.
On their measures, foundational capabilities did not clearly predict it. Apprentices matched Amplifiers and scored higher than Delegators on the foundational measures, yet still performed below the AI baseline. And it was not simply that Apprentices failed to challenge the AI. They did challenge it, but their critiques often chased irrelevant issues or steered the AI the wrong way.
Delegators, who scored lowest on foundational skill, largely accepted what the AI gave them and so stayed close to the AI baseline. Amplifiers were the ones who orchestrated the workflow: they framed the problem, anchored the work in relevant domain knowledge, evaluated what the AI produced and refined it across multiple rounds.
The distinction matters. Good AI use is not simply about challenging the machine; it is about knowing when, why and how to intervene.
KPMG and UT Austin’s earlier study of more than 1.4 million real workplace AI interactions points in the same direction. The more sophisticated users were distinguished by behaviour as much as technical knowledge: they were persistent, ambitious in the tasks they gave AI and deliberate about how they refined its output. That matters because these are behaviours which can be observed and developed.
There is an obvious caution: the July study was of early-career KPMG professionals in the US, not lawyers, and it did not test senior professionals. We do not think we should pretend that the findings transfer directly. But the question they raise certainly does.
We want to draw two things out of this, and neither is really about junior staff alone.
On competence
Competence is never only a question of what you know. Anyone who has spent time with the Solicitors Regulation Authority’s (SRA) Statement of Solicitor Competence or the Bar Standards Board’s (BSB) Professional Statement for Barristers will recognise this immediately.
Section A5 of the former already requires solicitors to apply understanding, critical thinking and analysis, including assessing information, recognising gaps, evaluating reliability and reaching reasoned decisions supported by relevant evidence. A2 also requires solicitors to reflect on and learn from practice and to adapt their practice to developments in the delivery of legal services.
Competences 2.5 and 2.6 of the BSB Professional Statement already require barristers to adopt a reflective approach to their work and to “practise with adaptability and flexibility, by being self-aware and self-directed, recognising and acting upon the continual need to maintain and develop their knowledge and skills”.
The BSB’s new guidance on the use of AI, valid from 18 May 2026, makes the connection even more explicit. It says barristers should maintain sufficient competence in technology and AI to understand how those technologies may affect their practice, whether or not they adopt them themselves.
Barristers who do use AI remain personally responsible for the work, should be able to justify their decisions and should verify references or conclusions through independent alternative sources. That is remarkably close to the behaviour the KPMG study identifies in its Amplifiers.
The SRA has now made much the same connection explicit, in a warning notice which it says it will have regard to when exercising its regulatory functions.
Its warning notice on the misuse of AI, published in August, states that solicitors and regulated individuals who use AI in the course of delivering legal services remain accountable for their work and outputs, regardless of how that work has been prepared.
The notice is anchored in the competence and supervision provisions of the code of conduct: paragraph 3.2 on providing a competent service, paragraph 3.5 on remaining accountable for work carried out through others and effectively supervising it, and paragraph 3.6 on ensuring that those you manage are competent and keep their professional knowledge and skills up to date.
It spells out the consequences for supervisors in unusually direct terms. Those who supervise junior or non-authorised colleagues may themselves be found to have breached regulatory requirements and professional duties where false citations reach the court without adequate review.
The regulator’s own summary is that appropriate human oversight, informed professional judgement and a proportionate, risk-based approach are essential.
That summary is where this research bites. Appropriate human oversight is not a condition a lawyer is in; it is something a lawyer does, and the KPMG findings suggest that doing it well is unevenly distributed and is not predicted by how much a person knows.

Michaela Hardwick
Both regulators now expect oversight of AI-assisted work, of supervisors as well as of the people they supervise. What neither regulator yet provides is a practical way of assessing whether a lawyer’s oversight of AI-assisted work is actually effective.
The underlying competence requirement has not changed. What has changed is the workflow in which that competence has to be exercised. If anything, this research suggests that the gap between knowing and doing, which reflective practice has always tried to close, becomes particularly visible once AI is involved.
The July study did not test senior professionals, so it cannot tell us that experience makes no difference. But nor should we assume that seniority itself solves the problem. A senior partner, KC or judge may have formidable professional knowledge and still use a new technology passively or uncritically.
Experience is valuable; it is not a substitute for learning how to exercise that experience in a changed workflow.
There is a related trap here. ‘Human in the loop’ sounds reassuring, but it means very little if the human merely approves a fluent-sounding answer. Research with knowledge workers has found an association between greater confidence in generative AI and lower self-reported critical-thinking effort.
That does not prove AI makes us less thoughtful, but it is a useful reminder that oversight has to be active to be worth anything.
On recruitment
We would be cautious about saying that foundational skill no longer predicts performance. That is not quite what the study shows, and the Amplifiers themselves had strong foundational skills. The more useful point is that those skills, on their own, did not identify who would add value when working with AI. That should make us think about recruitment and assessment.
Academic results, legal knowledge and the ability to produce good work without assistance remain important, but if an AI tool can itself produce a competent first draft, assessing only the finished answer tells us less about the person behind it than it once did.
What distinguished the Amplifiers in this study was a set of behaviours that can be assessed directly: orchestrating the work, framing the problem, testing the output and exercising judgement about what to accept, reject or change.
More importantly, these behaviours can be developed through training. KPMG’s own response gives a sense of what that training could look like: encouraging staff to record why an AI output was accepted, modified or rejected, and shifting some assessment away from the final answer alone and towards how a candidate reasoned through the task, tested assumptions and explained their decisions.
There is a regulatory echo here too: paragraph 7.2 of the SRA code requires solicitors to be able to justify their decisions and actions. Being able to explain why an AI output was accepted, changed or rejected is therefore not just a useful training exercise; it is closely aligned with an existing professional obligation.
There is also now enough legal-sector research to make this more than an analogy. Choi, Monahan and Schwarcz found that GPT-4 made participants substantially faster on realistic legal tasks, but produced only slight and inconsistent improvements in the quality of legal analysis.
A newer 2026 randomised trial, using a legal RAG system and an AI reasoning model, found improvements in both quality and productivity. And, in a separate study of 164 law students, Chen and Bao found that access to an LLM without training did not improve performance, while a brief training intervention did.
These studies are not directly comparable, but perhaps that is the point: outcomes depend on the task, the tool and what the human actually does with it.
Those ideas could readily be built into pupillage, training contracts and continuing competence records. A legal exercise could, for example, give a candidate an AI-generated piece of research, advice or drafting and ask them to work with it. What did they trust? What did they check? What did they reject? What had the AI missed? Can they explain why?
Recruiters and those responsible for training across the legal professions need to ask whether their selection and development processes currently test or build any of this at all.
A call for collaboration: education and practice need to talk to each other
We have noticed that new concepts or requirements can be introduced into learning pathways without carrying the same weight or emphasis once they reach practice; the competence statements are a case in point.
This can leave a gap between what the regulatory framework says and what happens day to day, and that can create unnecessary friction in the workplace, particularly in supervisory and mentoring relationships.
The timing is telling. Three days after publishing the AI warning notice, the SRA published its 2026 assessment of continuing competence. It found that some solicitors still focus their learning and development on technical legal knowledge rather than the wider skills needed to maintain competence.
That is almost the same problem viewed from the other end of the telescope: possessing knowledge, or recording that training has taken place, does not necessarily tell us whether the right professional behaviours are being developed.
Training and the use of AI is an area where we think it is now vital that education and practice are on the same page. The regulatory frameworks themselves are moving; the question is whether education, recruitment, supervision and continuing competence are translating those requirements into the way lawyers actually work.
If you train, recruit, supervise or educate the next generation of lawyers, we would like to hear how you are building this into what you do, and where you think the gap between statement and practice is widest. Comment below or get in touch directly.
Peter Fields is a practising barrister and FCIArb, with a particular focus on complex commercial disputes, effective dispute resolution and the use of AI in legal practice. Michaela Hardwick is a solicitor with over 30 years in legal practice and is founder of Beyond Compliance Ltd and Beyond Competence Ltd.