AI Can Read the Severance Agreement. It Can't Read the People Behind It.

Judith C. Wolff

AI reads documents fast, if not always correctly. It reads people only as accurately as they describe themselves.

These days, AI shapes the attorney-client relationship before I have even met a new client.¹ Clients find me through ChatGPT after being informed, also by AI, that they have solid claims for legal action. Their first email to me is often a narrative written for them by AI. When I open emails in the morning, it's pretty typical to see not a factual summary but a confident summation of legal theories and damages.

AI has become an essential tool in my practice, to the point that my engagement letters contain a disclosure that I routinely use AI during representation. It still makes mistakes, so I check everything it produces. But it works so quickly, and sounds so certain, that I can understand why a potential client would be more inclined to trust a machine over a human being with a financial stake in an outcome.²

It's already hard to remember life before AI. What bothers me most about it, though, is that it's just too darned agreeable with whoever is asking questions.³ Researchers call it sycophancy. Tell it your boss retaliated after you complained, and with nothing to check it against, it accepts what you say as unshakable truth. It becomes like a parent who cannot see any flaws in a child's behavior. Yes, it can argue the other side if explicitly asked to, but clients rarely ask, understandably. Everyone in a workplace dispute tells the story from their own side. That's human, not dishonest. People leave out the facts that don't reflect well on them, often without realizing those facts really matter to the outcome of the dispute. AI has told them what matters. And then, when I tell them something that counters their AI-generated conviction, they feel even worse than they did before.

Until AI routinely cross-examines casual users who want to know if they have a legal case, the questions that matter most will remain the ones AI doesn't ask:

What will they say is the reason you were let go?

What happened in the months you haven't mentioned?

Did you ever put that complaint in writing or were you just thinking about it?

And sometimes, silently: I'm noticing that you keep interrupting and talking over me. Did you do that at work?

The answers, and sometimes the silences, tell me the version of the story the employer probably already knows. Drawing out those facts without embarrassing the client is what an experienced attorney learns to do—something AI doesn't yet do reliably.

Do you even need an attorney to negotiate your severance these days?

For a simple separation agreement (standard form, formula severance, routine departure), AI can take an employee most of the way. It's mostly free and usually close enough, and for someone who either can't or doesn't want to afford a lawyer, it's a great solution.

Executive severance agreements are more complex: salary-based severance, equity, bonuses, "cause" and "good reason," restrictive covenants, how the exit is explained. AI can tell an executive whether the agreement is standard. But standard doesn't mean fair or final. The agreement is an opening offer, and the release is the company buying the executive's claims. Again, what the release is worth depends on things AI doesn't yet know to ask for:

Client’s Employment history. Length of tenure, performance reviews, working conditions, complaints made by or about them, who championed them, who they clashed with, employer promises not fulfilled, how others in their role have left.

The employer’s values and urgencies (and secrets). A reorganization that has to close by quarter's end. Leadership that wants to avoid a public dispute. Team morale. Other complaints. Knowing what the company needs, how soon, and what it wants to avoid is often where the leverage is.

What the client reveals. Interviewing clients is a lot like a first date. You learn an enormous amount by listening not just to what is said but also tone of voice, pauses, reticence or candor. You can listen and imagine how a client with unflappable common sense might intimidate a reactive supervisor, or how a client who drifts away from direct questions might come across at work. Both are useful information.

Settlement ranges. AI will produce a settlement range, and it will sound confident. But it draws mostly on published verdicts, a skewed slice of reality, because most severance deals and employment settlements are confidential.⁴ They exist only in the files and memories of the lawyers who negotiated them. That will probably change, but for now, experience knows more than the machine.

The rhythm. Every negotiation has a cadence, and usually the client sets it; they know better than I do, or than AI does, what tone will land with their company. Sometimes the best move is to be disruptive, sometimes to wait the other side out, sometimes to be completely straightforward. Knowing which comes from years of watching people and organizations, and even then, I don't always get it right.

AI is, to some extent, people-blind. It's incredibly smart, but it isn't yet savvy.

Changing norms

AI is changing more than how we bill. Practicing law has always been a somewhat isolated profession, and it's lonelier now.⁵ Alas, Claude never says, "I had a case like this once," or recounts the terrible mistake it once made during a trial. It never says, "Let's go play Wordle."

The prevalence of AI is also changing the rules of engagement between attorneys and their clients, particularly around fee structuring. The ethics rules have already caught up: a lawyer who bills by the hour may not charge for time that AI saved.⁶ By one industry estimate, up to 74% of the tasks lawyers bill by the hour could be automated.⁷ Corporate clients have noticed, and many now expect AI to lower what they pay outside counsel.⁸

My clients are individuals, not general counsel, but the question underneath is the same: if the machine did the reading, what exactly are they paying for? The answer, for now, is the part AI can't do—judgment, relationships, knowing which facts matter and which are background noise.

Empathy counts, too. People who have just been shoved out of a career don't need a machine to agree with them. They need a human who will tell them the truth—and help them close the chapter with kindness.

Notes

¹ Pew Research Center, "About half of U.S. adults now use AI chatbots, up from a third in 2024" (survey of 5,119 U.S. adults, February 2026); JUSTICE, "One in six people with a legal problem turning to AI for advice" (nationally representative UK survey of 3,287 adults, 2026).

² Jennifer M. Logg, Julia A. Minson, and Don A. Moore, "Algorithm Appreciation: People Prefer Algorithmic to Human Judgment", Organizational Behavior and Human Decision Processes 151 (2019): 90–103 (Harvard and UC Berkeley). See also Eike Schneiders et al., "Objection Overruled! Lay People can Distinguish Large Language Models from Lawyers, but still Favour Advice from an LLM" (2025), a UK study of legal advice specifically.

³ Mrinank Sharma et al., "Towards Understanding Sycophancy in Language Models", Anthropic (2023).

⁴ Minna J. Kotkin, "Outing Outcomes: An Empirical Study of Confidential Employment Discrimination Settlements", 64 Washington and Lee Law Review 111 (2007).

⁵ ABA Journal, "Lawyers rank highest on 'loneliness scale,' study finds" (reporting on Shawn Achor et al., "America's Loneliest Workers, According to Research," Harvard Business Review, March 2018); Daniel T. Lukasik, "Overcoming Loneliness in the Legal Profession", New York State Bar Association (July 22, 2026).

⁶ American Bar Association Standing Committee on Ethics and Professional Responsibility, Formal Opinion 512, "Generative Artificial Intelligence Tools" (July 29, 2024); see also State Bar of Texas, Opinion 705 (2025).

⁷ Clio, 2024 Legal Trends Report.

⁸ James Haggerty, Chambers and Partners, "AI and legal advice: Why client expectations are reshaping law firm strategy" (September 15, 2026); LexisNexis, "How generative AI will transform pricing in law" (52% of in-house teams expect bills to fall because of generative AI).