The hardest room to pitch AI to is a law firm partnership meeting. The hardest partner in that room is the one who's been practicing for 30 years, survived WordPerfect, Summation, Westlaw, and two generations of contract review software that promised to replace associates — and is still billing $800/hour because none of those things did what they promised.

That skepticism is earned. And in 2026, it's also driving real decisions that cost firms measurable time and money.

Here's the critique I take seriously: Claude Mythos Preview demonstrated that a Claude-based system can find OS vulnerabilities autonomously. That's impressive. It doesn't demonstrate that Claude understands the strategic posture of a commercial dispute, reads judges, or knows when not to make an argument. Lawyering is 20% research and 80% judgment. Mythos addresses the 20%.

That critique is correct. Where it breaks down is in the conclusion. The answer to "Claude doesn't have lawyer judgment" isn't "don't use Claude." It's "use Claude for the 20% that doesn't require lawyer judgment, and reallocate those hours to the 80% that does." The senior partner running this objection is still framing AI as a replacement. It's a force multiplier for the work that was never worth $800/hour in the first place.

"I've Seen Five Technology Revolutions — Show Me the Malpractice Case First"

This is the most rational objection a senior partner can make, and it's not a rhetorical trick. The malpractice exposure from AI errors is real and the case law is beginning to accumulate.

Mata v. Avianca (SDNY 2023) is the canonical case: attorneys submitted a ChatGPT-generated brief containing entirely fabricated case citations. The court sanctioned the attorneys, not OpenAI. The holding established what the senior partner already suspects: the tool doesn't carry the professional responsibility; the attorney does.

That's not an argument against Claude. It's an argument for correct workflow structure. Claude outputs are first drafts that require attorney review before becoming work product. The senior partner knows how to review first drafts — that's what first-year associates have been producing for 30 years. The question is whether Claude first drafts are better than first-year associate first drafts on research and drafting tasks. The answer, on average, is yes on research and competitive on drafting.

The malpractice risk isn't in using Claude. It's in not reviewing Claude output. That's a supervision failure, which is a workflow problem, not a tool problem.

The Accountability Gap: Who's Liable When Claude Gets It Wrong?

The senior partner's accountability objection is: "If Claude makes an error that harms the client, who pays?" The answer is clear and the partner already knows it: the attorney pays. Anthropic is not a licensed professional. The attorney's professional responsibility doesn't transfer to the AI vendor regardless of which tool was used — Harvey, CoCounsel, Claude, or a first-year associate.

Specialized legal AI tools address this partially by providing contractual indemnification clauses and documented review workflows that create a paper trail showing the firm exercised appropriate supervision. That documentation has value in a bar disciplinary proceeding: it shows the firm had a process, the process required review, and the attorney signed off on the work product.

Claude at $20/month doesn't generate that documentation automatically. You build it yourself: a simple internal protocol that records what prompt was used, what output was generated, and who reviewed it. Takes a day to set up. Not as convenient as a managed legal AI tool's built-in audit trail, but functionally equivalent in terms of what a bar committee would look for.

The accountability gap is real. It doesn't resolve by paying more for specialized tools — it resolves by building the supervision structure that makes Claude's outputs defensible work product.

Mythos-Level Capability Doesn't Mean Mythos-Level Legal Judgment

Claude Mythos Preview's April 2026 demonstration established that a Claude-based system can autonomously reason through thousands of complex technical iterations. That capability is real. It maps to cybersecurity vulnerability research in a controlled environment. It doesn't map to legal judgment.

Legal judgment requires: knowing which arguments will persuade this specific judge on this specific issue given the current circuit precedent; reading whether opposing counsel is bluffing or has a real case; deciding whether to file a motion that's legally sound but strategically counterproductive; understanding what the client actually needs versus what they're asking for. None of those are in Claude's capability set, regardless of the Mythos architecture.

The senior partner is right to make this distinction. Where the distinction fails is in the implication. The fact that Claude lacks legal judgment doesn't mean Claude's research and drafting output is useless. It means Claude's output needs judgment applied to it before it becomes work product. That judgment is what the partner provides. Claude changes what the associates are doing before the partner applies judgment — it doesn't change the partner's role.

Where the Skeptic Is Right — and Where the Conclusion Breaks Down

The 30-year partner's objections that hold:

Where the conclusion breaks down: the implicit assertion that these objections justify not using Claude at all. The objections describe constraints on how Claude should be used, not reasons to avoid it. A tool that requires attorney review before use and can't exercise independent legal judgment is still enormously useful if it compresses 10 hours of research and drafting into 2 hours, freeing the attorney to spend the other 8 on work that actually requires judgment.

The senior partner billing $800/hour has a finite number of hours. Every hour spent on work that doesn't require $800/hour judgment is a misallocation. Claude is the mechanism for reallocating those hours. The objections are about how to deploy it safely, not whether to deploy it at all.

The One Test That Actually Moves Skeptical Senior Partners

Abstract arguments don't move senior partners. Evidence from their own practice does.

The test: take a research memo from a first-year associate on a topic the partner knows cold. Ask Claude to produce the same memo on the same topic. Print both, remove the attribution, and show the partner both documents side by side. Ask which is better.

In my experience running this test, partners are typically not impressed because Claude is categorically better. They're moved because the gap between Claude and a first-year associate on research tasks is narrower than they expected. Sometimes Claude is better. Sometimes the associate is better. Sometimes they're competitive. But the exercise breaks the categorical assumption that AI output is obviously inferior to junior associate output — and once that assumption breaks, the cost calculus changes.

Claude Pro at $20/month versus first-year associate billing rates. If Claude produces research that's competitive with a first-year on a significant fraction of tasks, the economic math is not subtle. The senior partner doesn't need to be enthusiastic about AI to see that the numbers don't favor ignoring it.

My take: The 30-year partner's skepticism is legitimate on judgment and accountability — those critiques are correct and shouldn't be dismissed. Where it breaks down is the implicit conclusion that the right answer is waiting. Claude deployed with appropriate supervision handles the research and drafting load that doesn't require senior partner judgment. That's the test: not whether Claude replaces the partner, but whether it changes what the associates are doing while the partner is focused elsewhere.

AI-Assisted Research. This piece was researched and written with AI assistance, reviewed and edited by Manu Ayala. For deeper takes, email me directly.