When Not to Use AI: The Cases Where Human Judgment Is Not Optional

There is a practical argument for AI in a lot of legal operations contexts. There is also a set of situations where reaching for an AI tool is a mistake, and being clear about the difference matters.

AI tools are good at pattern matching and language generation. They are not good at judgment under genuine uncertainty, relational intelligence, or the kind of contextual reasoning that comes from years of experience inside a specific practice area. When a task requires the latter, AI is not a useful shortcut. It is a liability.

Strategic Decisions

If you are deciding whether to file a motion, how to position a case, or what settlement recommendation to bring to a client, AI can help you research the landscape. It cannot make the call.

That decision depends on factors no language model has access to: the specific judge's tendencies, the client's actual risk tolerance, the credibility of a witness who has been in three depositions, the political dynamics of a multi-party negotiation. These are judgment calls that require someone who knows the file, knows the players, and is accountable for the outcome.

Using AI to shortcut that judgment is not efficiency. It is exposure.

Client Crisis Situations

When a client calls in distress, whether that is a denial notice, an enforcement action, or a sudden change in circumstances, the response is a human function. A client who receives an AI-drafted communication during a crisis will feel it. The generic language, the missing specifics, the absence of the particular detail that would have told them you understood their situation.

Prepare for those conversations using every tool available. Use AI to pull together the case history, the relevant precedents, the open questions. Then get on the phone.

Complex Credibility Work

In immigration practice, credibility is often the case. Whether a client's account is internally consistent, whether the documentary evidence supports the narrative, whether the declaration addresses the specific weaknesses the officer will look for. This is not pattern matching. This is a practitioner reading a file with an understanding of how adjudicators think.

AI can help draft declarations. It cannot evaluate them for credibility risk. That evaluation requires the kind of judgment that comes from reading hundreds of decisions and knowing what failed and why.

When AI Flatters Bad Work

This is the underreported risk. AI tools generate confident, professional-sounding output regardless of whether the underlying work is correct. A poorly constructed argument comes back from an AI as a polished, organized brief. That is the danger.

The review process has to be capable of catching errors that the AI presentation has already made harder to see. If the person reviewing the AI output does not know enough to evaluate it critically, the review is theater, not quality control.

The Practical Rule

Use AI where the task is defined, the output is reviewable by someone with domain expertise, and the cost of an error is recoverable. Do not use it where any of those three conditions are missing.