The AI Selloff Repriced Legal AI Vendors, Not Risk

The immediate question after the February headlines was not whether the AI selloff’s impact on the legal AI market was real. It was real enough for anyone responsible for a tool shortlist, a renewal calendar, or a partner’s panicked “is this still safe?” email. The harder question is what kind of event it was. Anthropic’s legal workflow entry may have repriced incumbents and changed buyer leverage; it did not repeal the duties that attach when a lawyer files a brief, transmits client material, or relies on machine-generated legal analysis.

What actually moved in February

The timeline matters because several versions of the story now travel as if the market had discovered, in one trading session, that legal AI had suddenly become either safe or unsafe. Law.com Legaltech News reported Anthropic’s Jan. 30, 2026 announcement of a Claude Cowork legal plugin, with functions including document review, NDA triage, risk flagging, and compliance tracking.[1] Those are application-layer functions, not merely a model benchmark, which is why the announcement touched the legal-information incumbents rather than staying in the usual foundation-model lane.

By Tuesday, Feb. 3, Reuters reported that Thomson Reuters had fallen nearly 16% after seven straight losing sessions, while broader U.S. software and services stocks were also under pressure; Reuters also reported that the S&P 500 software and services index fell about 4% that Tuesday and another roughly 0.73% on Wednesday, erasing about $830 billion in market value since Jan. 28 and leaving the index about 26% below its October peak.[2] Legal.io’s one-month analysis put Thomson Reuters’ move at as much as 18% intraday, RELX at 14% — its steepest single-day fall since 1988 — and Wolters Kluwer at 13%.[3]

That was the procurement shock. It was not a court order, an ethics opinion, or an empirical finding about whether an AI-generated citation is more likely to be real this week than last week. Legal IT Insider captured the market shorthand by describing Anthropic’s legal plugin as causing a market meltdown, and Marketplace later framed the episode as part of a broader investor debate over whether a workflow update justified such a sharp reaction.[4][5] For legal teams, the useful part of the stock story ends there: the selloff tells buyers that vendor assumptions changed. It does not tell lawyers that their verification obligations changed.

Platform entry is a vendor risk before it is a legal-risk release valve

There is a real vendor story here. A foundation-model provider moving into document review, NDA triage, risk flagging, and compliance tracking pressures the familiar legal-information bundle. It raises the possibility that work once routed through research platforms, contract tools, or workflow overlays will be absorbed into general enterprise AI environments. It also gives procurement teams a reason to ask whether a product’s premium is based on legal content, workflow convenience, defensible verification features, enterprise controls, or simple inertia.

That is a different question from whether the output can be trusted. Market pressure can weaken a vendor’s roadmap, force repricing, invite consolidation, or make a long-term contract more dangerous if data export, audit logs, and transition assistance are poorly drafted. It can also make incumbents offer terms they resisted six months earlier. Those are reasons to reopen procurement files. They are not reasons to reduce review of an AI-produced case discussion, privilege call, or filing exhibit.

The tempting mistake is to treat “Anthropic entered the workflow layer” as if it answered the supervision problem. It does not. Anthropic’s own positioning for legal-plugin output still leaves licensed-attorney review in the chain, which is exactly where professional responsibility would put it even if the market had rallied instead of sold off.[1] The person who signs the filing, approves the production, or advises the business on privilege cannot outsource the last mile to a stock move.

The risk ledger did not move with the share price

The governing ledger still starts with lawyer duties, not vendor valuations. ABA Formal Opinion 512 ties generative-AI use to competence, confidentiality, candor to the tribunal, and supervision obligations under Rules 1.1, 1.6, 3.3, and 5.3.[6] That is the same basic map before and after the February rout: understand the tool well enough to use it competently, protect client information, do not mislead the court, and supervise nonlawyer assistance — including technological assistance — in a way that fits the risk.

The practical burden is sometimes called a “Verification Tax,” and the phrase is useful as long as it does not become a substitute for the rulebook.[3] It describes the work lawyers still have to perform after the model produces something plausible: checking cited authority, confirming quotations, matching facts to the record, reviewing procedural posture, and deciding whether the tool’s answer is appropriate for the client’s objective. The tax is not imposed by Thomson Reuters’ market capitalization. It is imposed by the court, the client relationship, and the lawyer’s signature block.

The sanctions record is the easiest place to see the separation. Damien Charlotin’s database of AI-hallucination matters had passed 700 cases and 128 lawyers by late 2025, with the reported pace at roughly two to three court cases per day.[7] Those figures do not prove that every legal-AI use is reckless; they prove something narrower and more useful: courts keep encountering filings and submissions where lawyers failed to catch fabricated or unsupported material before it reached the docket.

Accuracy studies point in the same direction without supporting lazy exaggeration. Stanford RegLab reported hallucination rates of 17% to 34% for legal-specific AI research tools in its assessment of leading systems.[8] That does not mean every answer from every legal tool is wrong. It means the error rate is too large to treat a polished answer as self-authenticating, especially when the task will be converted into a client recommendation, discovery position, or filed representation.

This is also why prior verification analysis on AI memory and legal duties remains more useful than a price chart for anyone revising an AI policy. A model’s ability to summarize, remember, or route work changes the workflow design. It does not erase the need to identify which human reviews which output before the output leaves the firm or legal department.

Privilege does not become safer because the vendor category changed

The privilege problem is equally resistant to market narration. In United States v. Heppner, a February 2026 Southern District of New York ruling rejected privilege and work-product protection where a client used consumer Claude under terms that disclaimed confidentiality.[9] The lesson is not that Claude is uniquely unusable, or that enterprise AI is categorically safe. The lesson is that confidentiality turns on the actual use, the actual terms, and the actual disclosure posture.

That distinction is where many rushed AI memos go soft. A team may have one analysis for an enterprise deployment with negotiated confidentiality, retention, audit, and data-use restrictions, and a very different analysis for a lawyer or client pasting protected material into a consumer interface. The February selloff may increase attention to Anthropic as a counterparty, and it may make vendor diligence more urgent. It does not make the confidentiality clause better, the data-flow map complete, or the waiver argument disappear.

For teams already tracking Claude-related counterparty issues, the better comparison is not “incumbent legal database versus exciting plugin.” It is the diligence framework used for any high-sensitivity AI provider: what data goes in, whether it is retained, who can access it, whether it trains models, how incidents are disclosed, and what happens at termination. Those questions belong in the same file as broader Claude vendor-liability diligence, not in a slide that says the market has voted for or against legal AI.

What legal buyers should change

The selloff is still operationally useful. It gives legal departments, firms, and legal-ops teams a reason to separate vendor-survival risk from legal-output risk in their procurement records. Those two risks often sit in the same evaluation spreadsheet, but they should not be scored as if they respond to the same controls.

• Recheck vendor viability for any tool that depends on a single product line, narrow funding path, or expensive third-party model access. The question is not whether the vendor’s share price fell; the question is whether support, security maintenance, contractual performance, and export assistance remain credible.
• Use changed leverage in renewals. Ask for stronger termination rights, transition services, data-return language, audit-log access, incident-notice commitments, and restrictions on model training or secondary use of client data.
• Do not weaken verification gates because a foundation-model platform now offers legal workflow functions. Citation checks, quotation checks, record checks, privilege review, and supervising-attorney signoff should remain attached to the task’s legal consequence.
• Document the human review path. If a plugin flags risk in an NDA, summarizes a production set, or drafts a compliance tracker entry, the file should show who reviewed the output and what sources or documents were used to confirm it.
• Treat vendor claims after the rout as advocacy, not evidence. Incumbents will emphasize trusted content and workflow depth; challengers will emphasize speed and lower friction. Neither position answers the duty-of-care question by itself.

This is also where the February event differs from a budget story. A buyer may delay a contract, demand better terms, or run a more aggressive proof of concept because the incumbent moat looks narrower. But the approval memo should still distinguish product-market risk from professional-risk controls. The same discipline applies to newer procurement questions around emerging AI applications, including the evidence-gap issues raised in OpenAI Astra legal applications and Astra legal analytics: a tool’s strategic promise does not answer what evidence supports the output, what records are retained, or who is accountable for review.

The commentary can be true and still not answer the lawyer’s question

There is a respectable argument that the market reaction overshot. Legal.io summarized commentary from JPMorgan and Nvidia describing the panic as illogical, and also noted more cautious assessments from Legal.io, The Orange Rag, Harvey CEO Winston Weinberg, and Legora CEO Max Junestrand that disruption was neither immediate nor inevitable.[3] Marketplace likewise treated the episode as a debate about how investors should value AI’s movement into software workflows.[5]

That debate is useful for finance teams and vendor boards. For a litigator or in-house risk manager, it answers only the outer question: which providers may gain leverage, lose leverage, or have to defend their pricing. The inner question remains stubbornly legal. If the tool produces a case citation, who checks it? If it summarizes privileged material, what terms governed the upload? If it flags a contractual risk, who decides whether the flag is right and whether the business can accept it? If the output enters a filing, who is prepared to defend it when the judge asks where it came from?

The February selloff repriced vendors and bargaining power. Court orders, ethics opinions, privilege rules, and verification obligations still price the professional risk.

References

1. “Anthropic Releases Legal Plugin in Cowork Among Other Extensions for Enterprise Work,” Law.com Legaltech News, Feb. 2, 2026.
2. “US software stocks hit by Anthropic launch,” Reuters via Yahoo Finance.
3. “Anthropic's Claude Legal Plugin One Month On: the Market Fallout and What It Means for Legal Teams,” Legal.io.
4. “Anthropic unveils Claude legal plugin and causes market meltdown,” Legal IT Insider, Feb. 3, 2026.
5. “Why did an update to Anthropic’s AI spark a stock sell-off?,” Marketplace, Feb. 6, 2026.
6. “Formal Opinion 512: Generative Artificial Intelligence Tools,” American Bar Association, July 29, 2024.
7. “AI Hallucination Cases,” Damien Charlotin.
8. “Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools,” Stanford RegLab, 2024.
9. “United States v. Heppner,” U.S. District Court for the Southern District of New York, February 2026.