When AI should stay silent: Preserving judgement in M&A

Large language models are designed to generate outputs. They are fluent by default, capable of producing polished, structured and commercially plausible responses even when the underlying analysis is incomplete or based on unstated assumptions. This fluency is superficially attractive in diligence, disclosure review and transaction management. But it also creates a critical risk: that readability is often mistaken for reliability.

A failure to produce an answer is not necessarily a failure of the system. In many cases, it is the most responsible outcome.

In practice, the most problematic failure mode is not obvious error. Obvious errors tend to be caught. The greater risk is confident error, where an answer appears coherent and well-reasoned but is built on gaps, assumptions or missing inputs that were never properly tested.

In M&A, that risk can have significant downstream consequences. In due diligence, confident but incomplete outputs can reduce materiality, obscure risk, and create a false sense of coverage across large document sets. In risk allocation and disclosure, they can subtly shape judgement in ways that are difficult to unwind later.

There is a natural tendency to treat failure as an answer to weakness, or as evidence that the system has failed to perform the task it was given. In practice, however, the opposite is true, particularly in professional settings where incomplete information, ambiguous drafting and evolving deal structures are common rather than exceptional.

This shifts the focus for lawyers. The objective is no longer to get better answers from AI, but to design systems and workflows that align with professional standards of judgement and restraint.

VIEW ALL

Among more sophisticated users, prompting is already changing. The emphasis is moving away from generating more content and towards shaping behaviour, particularly in relation to how systems respond to ambiguity, incomplete inputs and conflicting signals that are typical of M&A transactions.

The task for lawyers, particularly in M&A, is not to make these systems more fluent, because they already are, but to design and deploy them in a way that preserves judgement, deal discipline and accountability. Prompting, done properly, is one of the best ways that discipline is imposed. Not by encouraging the system to do and say more, but by ensuring it answers only when it should.

However, even well-structured prompts do not eliminate the need for scrutiny. A critical question in AI-assisted workflows is not simply whether an answer is wrong, but whether it should’ve been provided at all.

This is where structured testing becomes essential. Techniques such as “red teaming”, deliberately attempting to break or disprove an output, can expose hidden assumptions, alternative interpretations and failure points. By contrast, “blue teaming” requires the model to justify its conclusions using only the facts explicitly provided, without filling in gaps. Where those two approaches produce materially different results, it is often a signal that the original output relied on unstated assumptions.

Adversarial prompting takes this further by requiring the model to assume the opposite conclusion or identify what would need to be true for its answer to be misleading. Similarly, rerunning prompts with controlled changes to variables, such as evidentiary thresholds or tolerance for uncertainty, can reveal whether an answer is stable or whether it shifts under pressure.

These techniques do not replace legal judgement. They are tools to support it.

The future of AI-assisted M&A will undoubtedly involve better models and better workflows. It will also involve workflows that are deliberately less confident, particularly as those workflows become more complex.

As AI becomes more embedded in transactional workflows, the differentiator will not be how effectively lawyers can deploy these tools to generate answers, but how effectively they can ensure those tools stop, before the evidence does. Techniques such as red teaming, blue teaming, adversarial prompting and controlled reruns are practical ways to supplement human judgement when working with models in these environments.

Ultimately, generative AI is indifferent to risk and has no inherent sense of consequence. Nothing will surpass human judgement in deciding whether an output is to be relied on. Used well, AI can materially improve speed, consistency, and organisation. Used without restraint, it risks eroding the very judgement it is meant to support.

Steve Johns is a partner, Eliza Unger is a senior associate, and Lisa Ziegert is the director of client solutions at Hall & Wilcox.