Artificial intelligence is rapidly becoming embedded in professional practice, with lawyers, tax advisers and accountants increasingly using it to support everything from contract reviews and legislative analysis to reporting and advisory work. The debate is no longer whether these tools will be used, but how they should be used.
Much of the discussion surrounding AI risk has focused on hallucinations: from fabricated authorities to confidently stated but incorrect propositions. These incidents understandably attract attention. They are relatively easy to spot once discovered and provide clear examples of how reliance on AI can go wrong.
However, for legal, tax and accountancy professionals, the more significant long-term liability risk may be less obvious. It is not necessarily that AI produces an incorrect answer, but that AI-assisted drafting and redrafting may gradually alter the substance of professional advice in ways that are difficult to detect.
The issue is particularly relevant because AI is increasingly being used throughout the lifecycle of professional work rather than simply at the research stage. A technical note may be drafted using AI assistance, summarised for internal review, converted into client-facing language, revised following comments and ultimately condensed into a board paper or executive summary. At each stage, the technology may perform exactly as intended. The difficulty is that qualifications, assumptions and caveats can gradually disappear along the way.
This concern has been reflected in recent commentary on the rise of so-called “vibe lawyering”, a phrase used to describe an increasing tendency to rely upon AI-generated outputs because they appear coherent and professionally drafted. The real concern is not simply that professionals stop exercising judgment, but that a polished AI-generated note can make weaknesses in the underlying analysis harder to spot.
That risk is reinforced by emerging research. Microsoft researchers recently described large language models as potentially acting as “unreliable delegates”, capable of introducing isolated but significant errors that can accumulate through repeated interactions. In practice, advice rarely emerges fully formed and instead develops through multiple drafts and iterations. A minor omission introduced at an early stage may therefore survive every subsequent revision and ultimately become part of the final advice delivered to the client.
In contract drafting, the risk may arise where an AI-generated comparison fails to highlight the commercial significance of a modified liability cap or a termination provision that has changed between drafts, with technology identifying the textual amendment without adequately conveying its importance. In tax advisory work, the challenge is often more subtle as many tax conclusions depend upon carefully defined assumptions, which may gradually be lost through successive rounds of summarisation. A conclusion that was originally contingent and fact-sensitive may therefore gradually appear more definitive than intended.
The courts have already indicated where responsibility is likely to remain. In Cork & Anor v Smith [2026] EWHC 1199 (Ch), the High Court criticised a law firm after AI-generated material was included in documents presented to the court without adequate verification. The decision focused not on the use of AI itself, but on the failure to properly review and verify the material before relying upon it. The underlying principle is familiar: professional responsibility remains with the individual or firm providing the advice.
A related challenge is the growing prevalence of AI-generated analyses produced by clients themselves. These outputs are often heavily influenced by the factual descriptions contained within each client’s prompts, from which they are derived. Where those assumptions are incomplete or inaccurate, the resulting analysis may be equally misleading despite appearing persuasive. Ultimately, however, the obligation to test and verify the underlying reasoning remains unchanged.
Against this background, firms will benefit from focusing less on whether AI should be used and more on the stage at which AI introduces risk, as not all use cases carry the same exposure. Producing an internal meeting summary is fundamentally different from drafting a tax opinion. The challenge is therefore to identify the stages at which professional judgment cannot safely be delegated.
Perhaps the most interesting question is whether AI will ultimately raise the standard of care expected of professional advisers. Historically, certain inconsistencies or omissions may have escaped detection despite reasonable review processes. As AI tools become increasingly capable of identifying anomalies, clients, regulators and insurers may begin to expect those capabilities to form part of ordinary professional practice. The future negligence question may therefore not simply be whether a professional relied excessively on AI, but whether a reasonably competent professional making appropriate use of available technology would have identified the issue before the advice was delivered.
The developing liability landscape is therefore more nuanced than the current focus on hallucinations sometimes suggests. The greater risk is probably not the obvious AI failure, but the routine use of AI, where small changes accumulate over time without anyone noticing. As AI becomes embedded in advisory processes, the challenge for professionals will be to ensure that efficiency gains do not come at the expense of analytical precision.
Used properly, AI has the potential to improve the quality of professional work. Used uncritically, it may create a different category of risk: advice that remains persuasive and largely accurate, but which no longer fully reflects the reasoning on which it was originally based.
First published in Business & Accountancy Daily on 3 September, 2026.
