A Mississippi federal judge cancels a trial and disqualifies all four attorneys after both sides submitted AI-fabricated citations — the first time opposing counsel have been sanctioned simultaneously for the same hallucination failure. Harvey integrates live deal room data from Datasite and SS&C Intralinks directly into its AI workflows in 72 hours, making the virtual data room AI-native. Legora opens offices in Madrid, Milan, and Paris and establishes a London engineering hub, targeting 700 EMEA staff within a year. And Superlegal expands its licensed AI law firm into U.S. construction at $117 per attorney-certified contract — concrete proof that supervised AI can reprice the commodity end of legal services today.
After this briefing first went out, the story that has taken over every legal-AI and infosec channel is a U.S. export order barring non-U.S. citizens from accessing Anthropic's Mythos and Fable models. Because the order reaches any foreign national anywhere — including Anthropic's own non-citizen staff — the only way to comply was to switch the models off for every customer on the planet. Anthropic is calling it a misunderstanding and is working to get access restored, and the picture is moving by the hour. The structural point below holds either way.
The reflex take is already everywhere: this proves we need "sovereign AI." We think that's the wrong lesson. Start with what actually got fenced off — the model is the most replaceable layer in the entire stack. Anthropic themselves are more or less conceding that the capability the government is worried about can probably be reproduced from other public models that aren't controlled at all.
So when one model gets walled off, there are alternatives. Where today's alternatives don't quite match Fable, an export ban is precisely the pressure that closes the gap: open models improve and "sovereign" models race to catch the state of the art. A month, not a decade. Picture a European bank a year from now — cut off from the U.S. frontier, it runs a domestic model that's good enough for the work, puts a flag on it, gets a minister to the launch. Sovereign AI, job done. It's a relatively simple problem to solve.
The real issue sits underneath. That sovereign model still trains on American chips, still runs on Korean memory, and still gets fabbed and packaged in Taiwan. You can win the model and still rent every layer beneath it from the people you were trying to get free of.
So the ban exports the cheapest, most copyable layer and leaves the moat — compute, memory, packaging — completely untouched. It doesn't contain anyone's advantage. It's painting the bike shed.
The useful lesson here is narrower than the geopolitics. The supervised-AI thesis running through the rest of this briefing — Mississippi, the deal room, the $117 contract — has never depended on which model happens to be best this quarter. It depends on the things that don't move: your templates, your terms, your escalation rules, and a documented human review gate before any output leaves the system.
Anchor to what's hard to replace. Never to this quarter's best model — and certainly not to its continued legal existence. A model that can be switched off worldwide by a single export order is, by definition, not the thing your operation should be built around. The teams that come through this calmly are the ones whose infrastructure treats the model as a swappable component behind their own templates, terms, and supervision — so that when a model disappears overnight, the workflow keeps running on the next one.
Every AI hallucination sanction since Mata v. Avianca in 2023 has followed the same basic pattern: one attorney, one filing, one supervision failure. The Mississippi case breaks that pattern entirely. Judge Aycock's order disqualifies all four attorneys across both sides of a dispute because both sides' filings contained AI-fabricated citations in the same proceeding — a structural contamination of the record that left the court with no way to proceed on the merits.
The underlying dispute was a commercial case: attorney Tom Withers alleging unpaid legal fees from the City of Aberdeen, Mississippi. Routine in every respect except that when the court reviewed the filings, it found hallucinated case citations in submissions from both parties' counsel. Lawyers on both sides had used AI tools to generate their filings and had not verified the citations before signing and submitting them.
Judge Aycock's response was the most sweeping judicial response to an AI supervision failure in U.S. legal history. The trial was cancelled. All four attorneys were disqualified from the case. Two are barred from the court for two years. Fines range from $1,000 to $3,500, graded by whether the attorney drafted AI-hallucinated content or merely failed to verify it in a document someone else produced.
The legal logic behind the disqualification is important. When one side has AI hallucinations, a court can issue sanctions and continue. When both sides do, the court faces a record that may be systematically unreliable. The integrity of the proceeding itself is compromised, not just one party's filing. Disqualification of all counsel isn't punitive — it's the only mechanism available to reset a record that can't be trusted.
The Mississippi case changes the risk calculus for enterprise legal teams in two ways. First, it confirms that AI supervision failures are no longer limited to the professional consequences for the attorney — they can now destroy a client's legal position mid-proceeding. The disqualified clients in Aberdeen are now without counsel in an active case through no fault of their own. That risk is now real and documented.
Second, the both-sides dynamic removes the last implicit safety net. In prior hallucination cases, an attorney could assume that an opposing counsel mistake would, at worst, level the playing field. Mississippi shows that when both sides fail at the same time, the court has no good outcome available — only ways to reset a contaminated proceeding at everyone's expense.
Four weeks into the summer, the enforcement pattern has moved from "career disruption for one attorney" to "entire proceeding collapsed for all clients." The architectural implication is the same one that's been true since 2023, but the magnitude has changed: every AI output that reaches a court, a counterparty, or a regulator requires a documented verification step before it leaves the system. Not a checkbox. Not a disclaimer. A genuine human review of the substance. Any system that doesn't build that review in structurally — as a required gate, not an optional step — is one supervision failure away from the Mississippi outcome.
Virtual data rooms are where M&A deals live. The documents that define a transaction — financial statements, material contracts, regulatory correspondence, IP schedules — exist inside a permissioned VDR for the duration of the deal. Until this week, accessing them inside an AI tool meant downloading, re-uploading, and managing a separate permission layer. Harvey's integrations with Datasite (June 9) and SS&C Intralinks (June 11) change that: the two largest independent VDR platforms now feed live, permissioned transaction data directly into Harvey's AI workflows.
The integration works through permission inheritance. When a deal team connects a Datasite or Intralinks room to Harvey, the existing access controls follow. A document you're not permitted to see in the data room is a document you can't query in Harvey. The permission boundary is maintained at source, not manually replicated in a second system. For law firms and investment banks managing complex, multi-party deals with tiered document access, that's a meaningful security assurance.
Practically, it changes the shape of M&A diligence work. Deal teams can now run AI-powered diligence queries, issue spotters, and risk summaries against the live transaction record — without leaving Harvey's workflow environment, without document transfers, and without the version-control risk that comes with downloading and re-uploading materials from a live room.
Two tier-one VDR providers integrating with Harvey in the same week isn't a product coincidence — it's a signal that the AI workflow layer in M&A has reached the security and compliance maturity threshold that tier-one financial services firms require. Datasite and Intralinks serve the bulge-bracket banks and Magic Circle firms that set the security standards for the rest of the market. Their willingness to integrate is an implicit endorsement that Harvey's data handling meets those standards.
It also narrows the gap between AI tools and the actual transaction record, which is where the work lives. The gap was, until now, the main practical obstacle to AI-assisted diligence at speed — you couldn't trust the AI's document references if the documents it was working from were a day-old download from a room that had since been updated.
Harvey's deal room integrations make M&A diligence faster for the lawyers doing it. The question every enterprise legal team should ask their outside counsel — and their own legal ops function — is the same one Ironclad's data raised in May: if AI is accelerating the diligence workflow by a factor of two or three, where does the time saving go? Does it reduce the deal timeline? Does it reduce the fee? Or does it increase the number of deals the same team can handle, capturing more revenue at the same billing rate? In every major law firm AI announcement this year, the answer has been the third option. The efficiency gain accrues to capacity, not to cost. That's a choice — but it needs to be explicit in your outside counsel conversations.
Superlegal's expansion into construction this week is interesting not because AI contract review is new, but because of the precise terms: $117 per commercial contract, attorney-certified, returned in under 24 hours, through a law firm that is licensed to practice — not a software tool that helps a lawyer practice. That is a fundamentally different price signal from anything the legal market has seen before.
The mechanism matters. Superlegal operates under the Utah Supreme Court's Legal Services Innovation Sandbox — a regulatory programme that authorises the practice of law through technology-based services and permits non-lawyer ownership. It is a licensed law firm, not an AI tool sold to lawyers. A licensed attorney signs off on every contract it reviews. The liability sits with the firm, not with the client's procurement team or in-house counsel. That's the same professional structure as hiring outside counsel — at a fraction of the fee.
The construction sector is a particularly revealing test case. Commercial construction contracts are voluminous, repetitive, and high-stakes — subcontracts, lien waivers, change orders, performance bonds, indemnification clauses. They are exactly the kind of work that ends up at law firms not because the underlying issues are novel, but because the volume and the liability exposure make doing it in-house feel risky. Superlegal is built specifically for that gap.
The AGC partnership is the distribution story. The Associated General Contractors of America is one of the largest construction trade associations in the country, with members ranging from small regional contractors to national infrastructure firms. Superlegal's agreement with AGC makes attorney-certified AI contract review available to that entire member base — tens of thousands of businesses that currently have two options for legal work: pay law firm rates, or go without. The Superlegal model is a third option that didn't exist six months ago.
The significance of Superlegal isn't the Utah sandbox — it's the pricing proof point. For years, the argument for AI in legal services has been made in percentage terms: AI can reduce legal costs by 30%, 50%, 70%. Superlegal translates that into a dollar figure attached to a specific deliverable in a specific practice area. When a construction company can compare $117 for a Superlegal-reviewed contract against the invoice they received last quarter for the same type of work, the efficiency argument stops being theoretical.
The construction companies paying law firm rates for repetitive commercial contract reviews are doing so not because the work requires a partner's judgment, but because the routing infrastructure — the thing that identifies routine work, handles it at the right price point, and returns it with appropriate certification — didn't exist. Superlegal is that routing infrastructure, in licensed law firm form. The structural inefficiency it's solving is the same one that shows up across every sector of corporate legal work: inexpensive work done by expensive resources, at the price the expensive resources command.
Superlegal is building for SMEs and construction companies. But the pricing logic it demonstrates applies to every enterprise legal team that is still routing standard commercial contracts, NDA reviews, and procurement agreements to outside counsel. What is the dollar cost, per contract, of the routine work your outside firms handle today? And what would that figure need to be for you to rethink the routing? The Superlegal number — $117 for an attorney-certified review — tells you where the ceiling is for supervised AI on commodity work. If your current cost per contract is materially above that, the routing question isn't academic. It's a budget conversation you're already losing.
Three stories this week — a judicial disqualification order, two data room integrations paired with a pan-European expansion, and a $117 contract review from a licensed AI law firm — all describe the same market but from different ends of the value chain. Legal AI is bifurcating: the premium end is getting deeper and more secure; the commodity end is getting cheaper and more certified.
On the premium track: Harvey is embedding AI into the most complex and sensitive work in legal — live M&A transactions — with two tier-one VDR integrations that bring permissioned deal room data directly into AI workflows. Legora is opening four locations simultaneously to chase enterprise law firm demand that has outpaced its EMEA capacity, which is what a $5.6B company does when the market is moving faster than its salesforce. The premium track is consolidating around platforms with the security credentials and deal-workflow depth to operate where lawyers are already priced for the work.
On the commodity track: Superlegal's $117 attorney-certified contract review is the sharpest pricing signal the legal AI market has produced. It is not a discount product. It is a licensed law firm using a supervised AI workflow to deliver at the natural price of the underlying work — the price you get when the AI is doing the labour and an attorney is doing the verification, rather than the reverse. The construction sector is the entry point, but the pricing logic is sector-agnostic.
The Mississippi disqualification order sits across both tracks. It is simultaneously the most sweeping judicial response to AI supervision failures in U.S. legal history and a confirmation that supervision architecture is non-negotiable at every price point. Every $117 Superlegal contract has a licensed attorney sign-off. Every Harvey M&A diligence query in a permissioned deal room has an access-controlled document trail. The Mississippi outcome — four lawyers disqualified, one trial cancelled, both sets of clients left without counsel mid-case — is what happens when AI output reaches a legal proceeding without a credible human review gate in the architecture.
The bifurcation playing out this week is the structural insight Flank is built on: some legal work requires expensive expertise and earns the fee it commands; the rest is inexpensive work that keeps being routed to expensive resources because the routing infrastructure didn't exist. Harvey and Legora are racing to own the expensive end, building faster tools for lawyers already billing at the rate the work warrants. Superlegal is demonstrating that the commodity end can be attorney-certified, licensed, and delivered at $117. The enterprise legal teams that do well here are the ones that have categorised their work explicitly enough to know which track each item belongs on — and who have built the supervised agent infrastructure to handle the commodity work before it ever reaches a lawyer priced for the other one. Mississippi tells you the cost of not building that infrastructure. Superlegal tells you the price of building it correctly.