Why a form beats a chat for AI legal review

A founder pings the GC at 4:47 on a Tuesday. The launch is Thursday. The product is a customer-success agent that listens to sales calls in three jurisdictions and writes coaching notes about each rep. The GC has not seen the spec until now.

She has three options, and none of them is good. She can improvise a memo in Word, citing what she remembers about the EU AI Act and GDPR Article 22, hoping she did not miss a state biometric law. She can open Claude or ChatGPT, paste the spec, and ask "what regulations apply," which will produce a confident answer she will not have time to verify. Or she can ask outside counsel for a rush opinion, which will land Friday afternoon and cost roughly nine thousand dollars.

In all three cases, when the launch ships and a law changes six weeks later, nobody remembers what was decided or why.

This is the actual texture of AI legal review at most companies in 2026. Improvised. Undated. Unverifiable. Unowned. Anteroom exists because that texture is wrong.

The case against improvisation

Senior counsel time is the wrong tool for the wrong shape of problem. AI launches are repeatable. The 41st launch a company ships should not require the 41st blank Word document. The inputs are structured: use case, jurisdictions, data subjects, deployment context. The relevant law is finite. The EU AI Act has 113 articles, not 113,000. The output should be the same shape every time so it can be compared across launches, audited by a regulator, and forwarded to a PM without a translation step.

A blank Word doc has none of those properties. Neither does a Slack thread.

The case against the chatbot

This is the more interesting failure. Lawyers reach for Claude or ChatGPT because the marginal cost is zero and the answer is fast. The problem is what the answer actually contains.

In June 2023, two attorneys at a New York firm submitted a brief to the Southern District of New York citing six precedents. The judge could not find any of them. Asked for copies, the attorneys produced PDFs that looked like real court opinions, with case names, docket numbers, and reasoned analysis. The cases did not exist. ChatGPT had written them. The court sanctioned both attorneys under Rule 11. Mata v. Avianca, No. 22-cv-1461, S.D.N.Y. June 22, 2023.

This was not an isolated incident. In January 2024 the Second Circuit referred Jae Lee for discipline after a brief in Park v. Kim cited a hallucinated case. In February 2024 the Massachusetts Superior Court sanctioned Pamela Farley under Rule 11 in Smith v. Farwell for the same reason. In May 2025 the Utah Court of Appeals issued an order against Richard Bednar of the Durbano firm for filing a brief with non-existent citations, finding a violation of the duty of candor.

Four reported sanctions in twenty-three months, in matters where lawyers chose to use general-purpose AI for legal research. The dollar amounts are not the point. The point is that the output of a chatbot, on a legal question, looks correct in a way that does not survive contact with verification.

This is the structural property of LLM output that nobody markets. A chatbot generates text that is locally fluent and globally unverified. For a casual question this is fine. For a legal question where the citation must be real, the deficit is fatal. There is no chatbot user-interface fix for it. The fluency and the hallucination are the same property of the system, viewed from two angles.

What we are building instead

Anteroom is a structured place where the legal terrain for an AI launch is watched, cited, and shared. Three words, each doing work.

Watched. Every provision in the Anteroom corpus is a row in a curated table. EU AI Act Article 14. GDPR Article 22. Illinois BIPA Section 15. NYC Local Law 144. Each row carries the primary-source URL, the named reviewer who loaded it, the date it was last verified, and the effective date. When a regulator amends the underlying text, the row updates, and we know which saved analyses referenced it. The users who subscribed get an email.

This is what no chatbot can do. A chat session has no memory of what you asked three weeks ago and no notion of when the underlying law moved. Anteroom does. The saved analysis is a watched object, not a frozen snapshot.

Cited. Every claim in an Anteroom analysis points to a corpus row. Every corpus row points to its primary source. Click any citation in any saved analysis and you reach the actual EUR-Lex entry, the actual eCFR page, the actual state code. The citations are real because we curated them. They were verified by a named reviewer on a specific date. When you forward the analysis to a colleague, the colleague can check our work without trusting us.

This is the second property no chatbot can offer. A chatbot can produce a citation, but you cannot tell from the output whether the citation exists. With Anteroom, the trail is auditable end to end.

Shared. Every saved analysis lives at a permanent URL. Forward the URL. The recipient opens it and sees the same analysis you saw, with the same citations, against the same corpus version. The recipient does not need an account. The recipient does not need to recreate the wizard inputs. The artifact is portable.

A Claude chat is not portable in this way. You can share a link to a transcript, but the recipient sees your conversation, not the work product. Anteroom's saved analysis is the work product itself.

The receipt drawer

If the chatbot path is so flawed, why does it persist? Because the failure mode is invisible until it lands in a sanction order or an enforcement action. So here are some of the orders.

The FTC has been the most active. In July 2024 it sued NGL Labs and two of its principals, alongside the Los Angeles District Attorney, for using an AI chatbot to send fake threatening messages to minors as a growth tactic. FTC v. NGL Labs, No. 2:24-cv-5753, C.D. Cal. The proposed consent order cited Section 5 of the FTC Act, the Restore Online Shoppers' Confidence Act, the COPPA Rule at 16 C.F.R. § 312, and California's UCL and FAL. In September 2024 the FTC's Operation AI Comply settled with DoNotPay over deceptive claims about its "AI lawyer" service. Final order entered February 11, 2025. In January 2025 the FTC finalized a consent order with IntelliVision over false claims about the bias-free performance of its facial recognition AI. In April 2025 the FTC moved against Workado, the company behind a "content detector" AI, for false claims about detection accuracy.

The European DPAs are equally active. In April 2024 the Dutch Data Protection Authority issued a final decision against Clearview AI for facial recognition data processing without a lawful basis. €30.5 million fine. GDPR Articles 5(1)(a), 6, 9, 12-15, and 27. In December 2024 the Italian Garante issued a €15 million fine against OpenAI for inadequate lawful basis for processing personal data and for transparency failures.

State AGs are joining. In September 2024 the Texas Attorney General settled with Pieces Technologies, an AI healthcare startup, under the Texas Deceptive Trade Practices Act, over claims about the accuracy of its AI summaries of patient charts.

These are not the cases the company believed it was building toward. None of these companies thought they were going to be the cautionary tale in someone else's essay. The point of the receipts is not to argue that AI is risky. The point is that the law is in motion right now, and a tool that watches it for you is not a luxury.

How the workflow looks

You describe a launch in six structured fields. Use case. Jurisdictions. Data subjects. User population. Foundation model. Deployment context. Plus a free-text description if the structured fields cannot carry the nuance.

You submit. Anteroom runs a classifier against the corpus. Within seconds you see which dealbreakers fire (provisions a thoughtful counsel would flag as blocking and require remediation on before launch in the named jurisdictions), which obligations apply (specific operational, documentation, or notice work counsel would expect to see completed), and the provisions worth watching even if they do not strictly fire today. Each provision card carries a verified-by chip, the source URL, the plain-language reading, and a per-launch analysis written by an LLM but grounded in the corpus row, not pulled from training data. The final shipping determination is always counsel's, not the tool's.

The saved analysis is at a permanent URL. Share it. Subscribe an email address so you get notified when any of the underlying provisions shifts. Generate downstream artifacts, a DPIA under Article 35 GDPR, an application-layer model card, a board-level AI risk memo, each of which carries a footer pointing back to the saved analysis and to the audit trail of corpus rows that produced it.

The audit trail is its own page. It shows the saved profile as submitted, the classifier flags that fired, every corpus row consulted with reviewer and last-verified date, every generated document with its model and prompt and corpus-version stamps. A reviewing counsel can verify the system instead of trusting it.

This is the structure that does not exist anywhere else.

Who this is for, and what it is not

Senior in-house counsel at companies shipping AI products. PMs and founders who want a credible artifact to attach to a launch checklist and forward to outside counsel without losing a week. Outside counsel themselves who want a structured first-pass on a client matter before they bill into it.

Anteroom is not a better chatbot. We are not competing on output fluency. We will lose that comparison every time. Anteroom is not certification. We do not certify compliance and we do not establish an attorney-client relationship. Anteroom is not exhaustive. The corpus is curated, currently across ten lenses, twelve jurisdictions, and a few dozen provisions, growing weekly. We will keep adding rows. The corpus version is stamped on every analysis so you know what you got and when you got it.

Anteroom is not a substitute for a licensed attorney. Read the disclaimer on every output. We mean it.

What you can do today

Run the wizard at anteroom.so. Pick a launch you are actually thinking about. Submit. The whole loop is under thirty seconds. The saved analysis is yours. Forward it to your team. If a provision shifts that affects your launch, we will email you. That is the entire promise.

Browse the corpus directly. Every row visible, reviewer-attributed, source-linked, with effective and last-verified dates. If you find anything wrong, anything missing, anything sloppy, write to hello@anteroom.so. The corpus is human-curated. Our errors are correctable.

A note on what this is for

I built Anteroom because I run AI legal reviews at a public company and I do not have a better tool. Every senior counsel I have talked to either improvises or pays outside counsel or guesses with a chatbot. None of those is what we should be doing.

A small structured tool that watches the law, cites it correctly, and produces a portable artifact is not exotic. It is what lawyers have for every other domain. We have not had it for AI yet because the law moved too fast for the usual vendors to catch up. So I built it. The code will be open. The corpus is hand-curated. It is free to use. It is not legal advice. Use it the way you would use any tool a thoughtful colleague handed you: skeptically, with verification, and as a starting point for your own judgment.

---

Primary-source links to every cited matter live in the public corpus. When the underlying laws move, the corpus moves with them, and every saved analysis surfaces the diff.