AI is already in your practice. The workflow isn't.
LexForward AI Lab closes the gap — with structured prompts, verified outputs, and documented workflows built from real in-house practice.
Seven enterprise tools. One curriculum.
AI is already in legal practice. The question is whether you have a structured methodology — or whether you're improvising every time.
Law school taught doctrine. The job market wants a lawyer who can also design AI workflows, quality-check AI outputs, and explain the professional obligations that govern both. LexForward AI Lab gives you that — plus a portfolio of documented deliverables to prove it on day one.
Sometimes excellent. Sometimes alarming. You are not sure why the quality varies. You have no verification methodology — just a habit of re-reading and hoping. LexForward AI Lab replaces hope with a five-point verification framework and workflows that produce consistent output.
Some are using free tools with confidential client data. Some are submitting unverified AI output. None have a written acceptable use policy. LexForward AI Lab is a complete, credible training programme — with ethics coverage grounded in ABA Model Rules and professional responsibility standards.
Clients expect AI efficiency. Your new associates have the tools but not the workflows, verification standards, or professional responsibility training to use them safely. LexForward AI Lab gives every associate a consistent, firm-grade AI foundation — in one week, at scale.
LexForward AI Lab's Ethics module covers ABA Formal Opinion 512 and its rule-by-rule application to AI in legal practice. Available as co-brandable content for bar associations and law schools seeking AI curriculum grounded in professional responsibility. CLE accreditation application in progress — please contact us to discuss partnership timing.
The gap is not about intelligence or effort. It is about having a structured methodology. Every item below takes similar time — but only one produces professional-grade, defensible output.
"ABA Formal Opinion 512 (2024) clarifies that the duty of competence under Model Rule 1.1 extends to AI tools. Lawyers must understand both the capacity and limitations of any AI tool they use." LexForward AI Lab teaches exactly this — woven into every module, not added as a footnote. CLE accreditation application in progress.
No two lawyers on your team use AI the same way. No verification standard. No acceptable use policy. No documented workflows. If something goes wrong — a hallucinated citation submitted to a court, confidential data in a free AI tool — you have no evidence that adequate steps were taken.
LexForward AI Lab changes that. One course. Every lawyer working from the same verified methodology, professional obligation framework, and documented standards.
A new grad builds personal AI fluency. A mid-career in-house lawyer builds team systems. A GC builds an AI-governed legal function. One curriculum. Your pace.
Download the 20-prompt legal starter library. Use it on your next contract review. If the output isn't better than what you were getting before, you've lost nothing. If it is — you know what to do next.
10 modules + Ethics CLE. Personal AI fluency from the tools you already have. A working AI agent and documented portfolio by the time you finish.
One payment gives you ongoing access to the course content as it exists at purchase and as updated at our discretion. We intend to keep the course current and include updates at no additional charge, but reserve the right to modify content. Tier 1 purchasers will receive a preferential rate when Tier 2 launches.
None. This course assumes no prior AI experience and no technical knowledge. It assumes you are a qualified lawyer. Everything is built from that baseline. The most technically demanding module (Module 8 — building an AI agent) requires only a browser.
Approximately 7 hours of video content across 10 modules plus the Ethics module. Most lawyers complete it over two to three weeks alongside normal workload. The downloads are designed to be used immediately — the 20-prompt library and verification checklists apply from day one.
A CLE accreditation application is in progress. Until confirmed, we issue completion certificates documenting the content, hours, and subject matter — suitable for self-reporting under your state bar's applicable rules. We do not guarantee CLE credit. Enrolled students will be notified when accreditation is confirmed.
Refund requests submitted within 14 days of purchase, before completing more than two modules, will be considered at our discretion. Submit requests to wei@lexforward-ai.com with purchase details. Refunds are not available solely on dissatisfaction with content. Technical access failures are handled separately — contact us within 14 days and we will resolve the issue or refund.
The course is built from in-house commercial practice — contracts, M&A, and compliance. The methodology is designed to apply to any practice area involving document review, written analysis, or client advice. Individual applicability will depend on your specific practice context. No specific professional outcome is guaranteed.
Individual access is $599 — one payment, all content, all updates included. For teams of five or more, institutional plans are priced on request and are always cheaper per seat. Start with the free taster if you want to see the course before committing.
Enquire about institutional pricing →Every workflow in this curriculum comes from active in-house legal practice — using the full stack from free-tier consumer AI to legal-native enterprise platforms — on real commercial contracts, M&A, and compliance matters.
When generative AI started appearing in my daily legal work, I did what most in-house lawyers did — I asked general questions and got general answers. I re-read them, felt uncertain, and did most of the work manually anyway. Pasting client documents directly into a consumer AI tool was never an option: the confidentiality obligations were obvious from day one.
It took six months of daily experimentation to understand why. The problem was never the tool. It was the prompt — and the workflow around the prompt. Once I understood that AI is a pattern completion engine, not a reasoning engine, everything changed.
What makes LexForward AI Lab different is not the content alone — it is the experience behind it. As a New York-barred attorney in active in-house practice, with work experience spanning New York, Taiwan, and Switzerland, I apply the full AI stack daily to real commercial contracts, M&A matters, and compliance work — after redacting or anonymising confidential information in accordance with my professional obligations. The before-and-after numbers in this curriculum come from that real work: real task types, real document structures, real time measurements.
Most legal AI courses are built by academics, consultants, or people who have seen AI work in someone else's practice. This one was built because I needed it — and when I looked for something like it, it did not exist.
Almost no practising lawyer in the world has simultaneous hands-on experience across the complete AI stack — free-tier consumer tools, enterprise platforms, and legal-native specialist systems — plus prior CLM experience. Most legal AI adopters pick one product. The curriculum teaches which category of tool is right for which task, drawn from direct daily practice across all of them.
With work experience in New York, Taiwan, and Switzerland, and current in-house practice advising on cross-border commercial matters, the professional obligation framework in this curriculum is drawn from real multi-jurisdiction practice — not a single-bar perspective.
Not tools observed or researched from the outside. Tools used in real legal matters — in active current use or prior production experience. The workflows are tested on real documents. The failure modes come from real failures.
This list reflects current and prior experience and is updated as tools are adopted, retired, or replaced. Status indicators show current vs prior use.
The Ethics module (Module E) provides a full examination of how ABA Formal Opinion 512 applies to AI use in legal practice — rule-by-rule, with specific application to the tools and workflows covered throughout the curriculum. Not a compliance checklist. Professional content built into every module from the start. CLE accreditation application currently in progress.
45 minutes. A live indemnification clause demonstration — Round 1 versus Round 2, side by side. The complete Five-Part Framework. And 20 production-ready legal prompts you can use tonight.
Get instant free access →The 20-prompt library is sent immediately on sign-up. No drip campaign. No sequence. Just the prompts and an invitation to watch the taster module.
We review the same indemnification clause twice — using the kind of prompt most lawyers actually use, then using the Five-Part Framework. The gap between the two outputs is the core argument for this curriculum.
Notes from daily in-house legal work — what works, what fails, and what the professional obligations actually require. Written by a lawyer who uses the full range of AI tools — free-tier to legal-native enterprise — in active in-house legal practice.
The most common AI contract review mistake is asking a question that produces a vague answer. A live before-and-after on an indemnification clause — Round 1 versus Round 2, same tool, completely different output.
A plain-language breakdown of the six Model Rules that apply to AI use. The competence obligation is more demanding than most lawyers realise — and less demanding than the anxious ones fear.
As someone who uses all three simultaneously, the question is never “which is best?” — it’s which is best for this specific task. A task-by-task breakdown from active daily use.
A pattern I have started seeing in commercial negotiations — where both parties use AI, neither understands what it is doing, and the deal quietly stops moving. What causes it and why a lawyer who understands the mechanism is more valuable than ever.
Most lawyers who say they verified their AI output did not verify it — they re-read it. Re-reading is not a verification methodology. This is what one actually looks like, applied to three of the most common legal tasks.
These notes are published when there is something genuinely worth saying — not on a schedule. Subscribe below for the highlights, or follow on LinkedIn for the full stream.
Monthly digest of what's worth knowing about AI in legal practice. No filler. Published when there's something worth saying.
The gap between a vague prompt and a structured one is not about the AI — it is about what you give it to work with. A live before-and-after on an indemnification clause.
I want to start with a question. The last time you used AI to review a contract clause — what did you type?
If the answer is something like "review this clause" or "is this reasonable?" or "what are the issues here?" — you are in good company. This is how the overwhelming majority of lawyers use AI for contract work. And it is why the overwhelming majority of lawyers are getting approximately 20% of the value that AI can deliver.
The problem is not the tool. The problem is the question. Let me show you exactly what I mean.
Here is a standard indemnification clause from a commercial services agreement. The prompt most lawyers type:
And here, approximately, is what comes back:
Vague in. Vague out. The AI told you what you told it — that you wanted a general assessment — and produced a general assessment. This is not the AI failing. This is the prompt working exactly as designed.
Look at what this output cannot tell you. It does not know which party you represent. It does not know your standard positions. It cannot tell you whether the consequential loss exclusion has a carve-out for wilful misconduct — because you did not tell it whether you need one. And it does not give you a single actionable next step.
Here is the same clause, reviewed with a structured prompt. The prompt takes about 60 seconds to build once you have done it a few times.
Wilful misconduct carve-out: DEVIATES — SIGNIFICANT. Clause excludes consequential loss "regardless of theory of liability" with no carve-out for wilful misconduct. Redline: insert "provided that this exclusion shall not apply where the relevant loss arises from wilful misconduct or fraud."
Third-party scope: DEVIATES — SIGNIFICANT. Clause covers all third-party claims — significantly broader than IP infringement only. Redline: replace with "third-party claims alleging infringement of intellectual property rights by the indemnifying party's deliverables."
Liability cap carve-outs: SILENT — SIGNIFICANT. 12-month cap applies without carve-outs. Redline: add "excluding claims for data breach, death or personal injury, fraud, and wilful misconduct."
The AI is not smarter in Round 2. You are. You gave it the context it needed — your role, your positions, your specific numbered tasks, your constraints, and the output format you can use. The intelligence went into the prompt, not just into the review of the output.
The redlines in Round 2 are usable right now. Copy them into a tracked-changes version of the contract and send. That is not something the Round 1 output could do for you.
ABA Formal Opinion 512 (2024) requires lawyers to verify AI output before relying on it — specifically, every legal citation and every suggested redline. The Round 2 output above is a starting point. The five-point verification checklist must follow before anything goes to the counterparty. Speed without verification is not efficiency. It is liability.
The Five-Part Framework — Role, Context, Task, Constraints, Output Format — is the consistent structure behind every high-quality AI output. It is learnable. It is fast once you have done it a few times. And it applies to every legal task where you are currently getting generic output from AI.
The 20-prompt legal starter library (free download) contains 20 fully written versions of this structure across six task categories. The taster module shows the full live demonstration. The full curriculum teaches you to build your own prompt library calibrated to your specific practice and standard positions.
20 structured prompts across 6 task categories. Use them tonight.
The duty of competence now extends to AI tools. This is less alarming than it sounds — and more demanding than most lawyers have considered.
In July 2024, the American Bar Association issued Formal Opinion 512 — the most comprehensive guidance on lawyer AI use from any major bar association in the world. Since then, I have seen it mischaracterised in two equally unhelpful directions: either dismissed as aspirational guidance with no real teeth, or described as creating sweeping new obligations that will make AI use impossibly risky.
Both readings are wrong. Let me tell you what it actually says.
Opinion 512 does not create new professional obligations. Every obligation it identifies existed before generative AI existed. What it does is clarify how existing Model Rules apply to AI — and in doing so, it makes explicit what was previously implicit. Six areas. Six Model Rules.
The Opinion states that the duty of competence extends to AI tools. Lawyers must understand both the capacity of the AI tools they use — what they do reliably — and the limitations — where they fail and why.
What does "understand" mean in this context? The Opinion is specific: you do not need to understand large language model architecture. You need to understand that AI is a pattern completion engine, not a reasoning engine. You need to understand the three types of AI error (factual hallucination, plausible distortion, omission). You need to understand what a knowledge cutoff means for legal research. The obligation is also ongoing — as tools evolve, the understanding must be updated.
"The duty of competence requires understanding the capacity AND limitations of AI tools — not just the ability to use them."
ABA Formal Opinion 512, 2024
Before entering client information into any AI tool, lawyers must assess the potential that the information will be disclosed to or accessed by others. Free AI tools without enterprise data agreements are high-risk for confidential information. Most bar guidance on this point is consistent: where you have no formal data processing relationship with a vendor, avoid putting confidential client data into their tools.
The solution for free tools: anonymise before pasting. Replace client name with Party A, counterparty with Party B. The AI does not need to know whose contract it is to analyse the clause. This takes 30 seconds and eliminates the primary confidentiality risk.
All AI-generated legal citations submitted to a court must be independently verified. This is the absolute version of the obligation — no exceptions.
Mata v. Avianca, Southern District of New York, 2023. Lawyers submitted a brief containing citations to cases that did not exist. The citations were generated by ChatGPT. The sanctions were $5,000. Written apology letters to the named judges were required. "The AI produced it" was not a defence. It was not even argued as one.
For court submissions: every citation, every statutory reference must be verified against a primary source before submission. Every time. No exceptions for time pressure, document volume, or confidence in the AI's output. ABA MR 3.3 does not have an AI exception.
Supervision of non-lawyer staff extends to AI tools and AI output. A lawyer who reviews work product that was AI-generated is responsible for the adequacy of the verification, not just the quality of the final output. "I reviewed the memo" is not adequate supervision if the memo was AI-generated and you did not verify that adequate verification was done.
The alarmed reading: Opinion 512 means using AI is now legally perilous. This is wrong. The Opinion specifically notes that failing to use AI when it would produce better results for a client may itself raise competence questions under MR 1.1. The duty of competence runs in both directions.
The dismissive reading: Opinion 512 is just guidance. This is also wrong. The Opinion applies existing Model Rules — rules with real disciplinary consequences — to AI use. Submitting unverified AI citations is a MR 3.3 violation. Entering client information into a free AI tool without adequate data protection assessment is a potential MR 1.6 violation.
What the Opinion actually requires is a structured, documented approach to AI use. Understand the tools you use. Assess data protection before entering client information. Verify every output before relying on it. Supervise AI-assisted work at every level. Document your process. That is achievable. And it is considerably more than most lawyers are currently doing.
Rule-by-rule application of Opinion 512 to AI in legal practice. Included in the full Tier 1 curriculum. CLE accreditation in progress.
The question is never "which AI tool is best?" The question is "which AI tool is best for this specific task?" A task-by-task breakdown from someone who uses all three daily, simultaneously.
Every few months a new "Harvey vs ChatGPT" comparison article appears. They almost always reach the same conclusion: Harvey wins on legal-specific tasks, ChatGPT wins on price and general use, Copilot wins on Microsoft integration. Correct. And almost entirely beside the point.
The useful question is not which tool is best in the abstract. It is which tool is best for the specific task in front of you right now. That question can only be answered from the inside — from someone who has run the same task across all three and measured the difference.
I have used the full range of AI tools in active in-house legal practice — from free-tier consumer tools to legal-native enterprise platforms, plus hands-on CLM experience in active practice. Here is what I have learned about which tool belongs where.
I think about AI tools in three stacks rather than three products. Stack A (free to personal paid): ChatGPT free/Plus, Gemini personal. Accessible to any lawyer today. No enterprise data protection. Requires anonymisation before any confidential content. Strong for prompt engineering practice and general drafting. Stack B (mid-market enterprise): ChatGPT Enterprise, Gemini Workspace, Microsoft Copilot M365. Enterprise data processing agreements. Confidential content can go in subject to your organisation's approval. Stack C (legal-native): Harvey and other legal-native platforms. Purpose-built for legal work. Legal-grade data protection by design. Higher cost. Appropriate for high-volume, high-stakes legal work.
Harvey: Strongest output for complex, multi-party NDAs with unusual structures. Understands legal concepts natively. Handles large documents without context window degradation. Consistently higher quality on complex agreements.
ChatGPT (with Five-Part Framework): Excellent for standard mutual NDAs under New York law. Like all current general-purpose AI tools, it does not produce in-document tracked-change redlines — it generates suggested language you paste in and format yourself. The structured prompt closes most of the gap with Harvey for straightforward agreements. Faster to iterate. Free tier works well for anonymised content.
Copilot Advanced (in Word): Strong for drafting, summarisation, and asking questions about a document within the Word environment. Like ChatGPT, it does not currently produce in-document tracked-change redlines — output is generated text you paste in and accept, not native tracked changes. The advantage over ChatGPT is that it can reference other Microsoft 365 documents as context without pasting. Best for teams with mature SharePoint environments.
All three tools have knowledge cutoffs. All three will hallucinate citations with complete confidence. The tool does not matter as much as the verification workflow for research tasks. That said: Harvey has been trained specifically on legal content and tends to produce more precise legal propositions with better jurisdiction specificity. ChatGPT is better for synthesising a research landscape quickly. Neither should be used for current law research without primary source verification.
"Harvey is not better than ChatGPT. Harvey is better than ChatGPT for specific legal reasoning tasks on complex documents. ChatGPT is better for iteration speed, general drafting, and tasks where the Five-Part Framework closes the gap."
Microsoft Copilot Advanced has a capability that Harvey and ChatGPT do not: it can reach into your organisation's Microsoft 365 environment — your SharePoint policies, your existing compliance frameworks, your internal guidance documents — and use them as context for analysis. For a compliance gap analysis comparing a new regulation against your existing internal policy, Copilot can pull both from your environment without requiring you to paste either. This is a meaningful advantage for in-house teams with mature document environments.
If you have free or personal-paid tools: ChatGPT with the Five-Part Framework delivers the most value. The structure of the prompt matters more than the choice of tool at this level. Start here. Master the framework. Then evaluate whether a tool upgrade actually changes your output.
If you have enterprise tools: Copilot for in-document workflows and internal policy work. ChatGPT Enterprise for prompt-heavy tasks and agent building. Harvey for complex legal reasoning and large-document analysis where the volume and stakes justify the cost.
If you have a CLM system: the entire conversation about Harvey vs ChatGPT is addressing a different problem. CLM handles the operational layer — intake, approval, renewal, obligation tracking. Harvey and ChatGPT handle the intelligence layer. These are not competing; they are complementary. The CLM content in this curriculum is built from direct hands-on production experience — not speculation about what a CLM does in practice.
Stack A tools require anonymisation before any confidential content. Stack B tools require your organisation's data processing agreement and approval. Stack C tools have legal-grade data protection built in — but you still need to confirm your organisation's AI acceptable use policy covers the tool and the data category. "Enterprise-grade" does not mean "unlimited use of any data."
LexForward AI Lab teaches all three stacks — from free to enterprise — applied to the same workflows. Start free.
A pattern I have started seeing in commercial negotiations — where both parties use AI, neither understands what it is doing, and the deal quietly stops moving. What causes it, what it looks like from the inside, and why a lawyer who understands the mechanism is more valuable now than they were before AI existed.
There is a reasonable-sounding assumption spreading through commercial legal practice right now: that AI tools, by making both sides faster and better-informed, will make negotiations more efficient. Both parties draft smarter documents. Both parties review more thoroughly. Deals close faster.
The assumption is wrong — or at least, it is only true under conditions that most people using AI in negotiations do not currently meet. Under the wrong conditions, AI does not make negotiations more efficient. It makes them slower, more circular, and harder to diagnose. I have started calling the specific failure pattern “AI ping-pong.” This article explains what it is, why it happens, and what it means for lawyers who understand the technology versus those who do not.
The setup is straightforward. Party A — the company sending out contracts — has started using an AI tool to draft. Party B — the counterparty receiving and reviewing those contracts — has also started using an AI tool to redline. Neither party is doing anything unusual by the standards of current commercial practice.
Step one. Party A uses AI to draft a services agreement from their standard template. The output is clean, well-formatted, and internally consistent. It reflects Party A’s positions — but only because those positions are embedded in the template, not because the AI understood them, weighed them, or made any judgment about them. The AI completed a pattern. The template was the pattern.
Step two. Party B’s lawyer receives the draft. They open their AI tool, paste in Party A’s document, and ask it to review against Party B’s standard positions. The AI flags deviations, rates risk, and produces redlines. Some are accurate. Others reflect generic market-standard positions that may or may not correspond to where Party B actually stands. The lawyer reviews the output — quickly, because that is the point — accepts the redlines that look reasonable, and sends the document back.
Step three. Party A’s lawyer receives the redlined document. The natural thing to do is feed their original template and Party B’s redlined version into the tool and ask it to reconcile the two, or review the changes and recommend which to accept.
Here is what some tools actually do at this step. The AI has been given two documents: Party A’s original template — the “correct” version from the tool’s perspective, because it was the starting point — and Party B’s redlined version, which is a deviation from that starting point. The tool’s pattern-matching mechanism identifies the deviations and, depending on how the prompt is constructed, will often recommend reverting those deviations. Not because the deviations are legally wrong. Because they are departures from the template, and the template is the closest match to the pattern the tool is working from.
Party A sends the document back with most of Party B’s substantive redlines quietly reverted to Party A’s original language.
Step four. Party B’s lawyer receives the document back. They run it through their AI again. The AI identifies that Party A has rejected most of Party B’s redlines and produces a fresh set of recommendations — largely the same ones as before, because the underlying gap has not changed. Party B sends the document back again.
The deal is not progressing. Neither party understands why. Each side believes they are being commercially reasonable. Each side is being guided by an AI that is producing output calibrated to pattern-matching rather than negotiating judgment. The tool is not broken. No one is doing what a lawyer is supposed to do: exercising judgment about what matters, what does not, where to push, and where to concede.
A large language model does not read a contract the way a lawyer does. It predicts the most statistically likely next token given everything that preceded it — including the prompt instructions, the document content, and the training data. When a model is asked to “review this contract against my template,” it is performing pattern-matching. It identifies where the current document diverges from patterns common in its training data and from patterns in the template it has been given.
This mechanism is genuinely useful — fast, consistent, and catches things a tired reviewer might miss. But it has a specific failure mode: it cannot distinguish between a deviation that matters commercially and one that does not. It cannot weigh competing positions. It cannot assess whether a concession on one clause is worth a gain on another. All of that requires judgment — which requires understanding the commercial context — which requires being a lawyer who has been briefed on the deal.
In AI ping-pong, neither party has a lawyer doing that work. Both parties have lawyers who are reviewing AI output and accepting it at too high a rate. The AI is steering the negotiation, and the AI does not know what either party actually wants.
There is a related but simpler failure pattern that I see more frequently, and it does not require two AI-using parties to create the problem.
Someone — a business owner, a procurement manager, occasionally a junior executive at a company that has recently reduced its legal team — decides that AI can draft their contract. They do not need to instruct a lawyer because the AI will produce a professional-looking document. They are right that it will. They are wrong about what that means.
What AI produces without specific, detailed instruction about the drafter’s positions is a document calibrated to generic patterns in its training data. Relatively balanced. Commercially conventional. Not obviously unfair to either party. It is also a document that contains none of the drafter’s actual positions on the points that matter to their specific business, their specific risk appetite, and their specific relationship with the counterparty.
The document goes to the counterparty. The counterparty’s AI — or the counterparty’s lawyer — reviews it and responds based on their own positions. What follows is a negotiation conducted between two sets of generic positions and one set of specific ones. The party with a lawyer who understands what they actually need wins. The party who relied on AI to represent their positions — without telling the AI what those positions were — does not.
The lawyers called in to untangle an agreement that has been drafted and partially negotiated by AI typically spend more time, not less, than they would have if they had been involved from the start. The document has accumulated layers of competing revisions, some of which make no commercial sense because they were generated by AI responding to AI. The positions of both parties are buried under machine-generated language that nobody actually intended.
The conclusion some lawyers draw from this is that AI is dangerous and should be avoided. That conclusion is wrong.
AI ping-pong does not happen because AI is bad at contract review. It happens because AI is being used without the professional judgment that makes it useful. The same tool that produces AI ping-pong, used correctly — with a structured prompt encoding the drafter’s actual positions, followed by the lawyer’s critical review of the output, followed by judgment about what to push back on — produces faster, more consistent, and better-documented contract review than most manual processes.
The relevant skill is not deciding whether to use AI. The relevant skill is understanding the tool well enough to know when its output reflects your actual position and when it reflects a pattern that happens to look like your position. Those are very different things, and the gap between them is where the professional value of a lawyer now lives.
The lawyer who understands why AI reverted Party B’s redlines can prevent it from happening. They know to construct the prompt so the model understands it is reviewing the counterparty’s version against specific named positions, not reconciling deviations from a template. They know to read the AI’s recommendations critically rather than accept them at the rate that defeats the purpose of having a lawyer in the room.
The lawyer who does not understand the mechanism cannot diagnose the failure. They see the negotiation stalling, attribute it to commercial intransigence, and spend time trying to move positions that were never actually adopted in the first place.
The prompt is the position. If you want AI to represent your negotiating position, give it your negotiating position explicitly — in the Context section of the prompt, before the document. “My standard position on consequential loss exclusions is X — flag any deviation and rate its commercial significance” produces output calibrated to where you actually stand. Generic review prompts produce generic output. This is the most important thing in this article.
Verify the revert. Any time you use AI to review a counterparty’s version of a document you originally drafted, check manually whether the AI has recommended reverting substantive changes back to your original language. Read the AI’s recommendations alongside the counterparty’s redlined version, not just alongside your original. If you cannot explain why a specific change has been rejected on commercial grounds — as opposed to because it differs from your template — do not reject it.
Understand what the other side’s AI is doing. If you know counterparty counsel is likely using AI to review your draft, you can construct your draft in ways that are less likely to trigger unhelpful AI responses. A well-structured, clearly labelled contract with defined positions is less likely to produce AI-generated chaos from the other side than an ambiguous one.
AI has not made lawyers less necessary in commercial negotiations. It has changed what the necessary skills are. The lawyer whose value was in reading contracts quickly and producing first-draft redlines faces genuine pressure from AI that does those things faster. The lawyer whose value is in understanding what the tools are doing, directing them accurately, and exercising judgment about what the output means — that lawyer is more valuable now than they were before AI existed, because the consequences of not having that judgment in the room are visible in ways they were not before.
AI ping-pong is the clearest demonstration of this I have seen. It is not a technology failure. It is a professional judgment failure that technology has made more consequential.
ABA Formal Opinion 512 (July 2024) makes clear that competence under MR 1.1 extends to understanding the limitations of AI tools, not just their capabilities. A lawyer who uses AI for contract review without understanding the pattern-reversion failure mode described in this article is not demonstrating competence in AI-assisted legal work — they are demonstrating fluency with a tool whose failure modes they cannot diagnose.
Most lawyers who say they verified their AI output did not verify it — they re-read it. Re-reading is not a verification methodology. This is what one actually looks like, applied to three of the most common legal tasks.
After ABA Formal Opinion 512 was published in July 2024, I started paying attention to how lawyers described their AI review process. The most common answer, when asked whether they had verified an AI output before using it in work product, was some version of: "I read through it and it looked right."
That is not verification. It is an optimism check.
The problem is not that lawyers are being careless. The problem is that re-reading AI output feels like checking it — the output is coherent, professional in register, and structured in a way that signals reliability. The issue is that the things most likely to be wrong in AI output are not things re-reading can catch. You cannot spot a hallucinated case citation by re-reading. You cannot identify a clause that has been silently omitted by reading what is there. You cannot detect that an AI has applied the wrong jurisdiction's law by reading fluent legal prose that happens to be wrong.
Verification requires a methodology — a set of specific checks, applied in a specific order, that tests the things most likely to fail in AI output for a given task type. This article explains what that methodology looks like for three tasks: contract review, legal research, and client advice.
AI language models are, at their core, text prediction engines. They predict the most likely next word given everything that came before. This produces output that is stylistically consistent, grammatically correct, and coherent in structure — because coherence and consistency are exactly what the training data rewards. The model has been trained on enormous quantities of professional text. It has learned what professional text looks like. It produces professional-looking text.
The failure modes are not stylistic. They are factual, logical, and jurisdictional.
AI can produce a contract review memo that reads exactly like something you would write, citing a case that does not exist, applying a contractual interpretation that does not follow from the clause, and missing a deviation from your standard position that was present in the document all along. None of these failures are detectable by re-reading the AI's output. They are only detectable by checking the output against independent sources: the document, the primary law, and your own stated positions.
This is the fundamental insight behind any verification methodology. You are not trying to assess whether the AI output sounds right. You are trying to assess whether specific, testable claims in the output are accurate. Those are different questions, and only one of them can be answered by reading.
Every AI output you intend to rely on should pass five checks before it becomes work product. These checks apply regardless of the tool — free-tier or legal-native — and regardless of the task. The weighting of each check shifts depending on the task type.
Check 1 — Source verification. Does every factual claim in the output correspond to something you can independently confirm? For contract review: does the AI's description of a clause actually match what the clause says? For research: does the case it cites exist, say what the AI claims it says, and remain good law? For client advice: does the regulatory position described reflect the current state of the law?
Source verification is the most important check and the one most commonly skipped. It is not enough to know that the AI usually gets citations right. You need to know that this citation, in this output, is correct — because the only way to know that is to check it.
Check 2 — Logic verification. Does the reasoning in the output follow from the facts? In a contract review, does each risk rating have a specific, articulable reason — or has the AI applied a rating without explaining why? In a research memo, does the analysis follow from the cases cited — or has the AI used real cases to support a proposition the cases do not actually stand for? Logic failures are harder to spot than factual errors because they require you to think through the reasoning rather than just confirm a fact. They are also common: AI is very good at constructing plausible reasoning chains and very bad at knowing whether those chains are sound.
Check 3 — Completeness verification. Has the AI addressed everything you asked — and everything you should have asked? Completeness failures come in two forms. Explicit omissions: the AI simply did not address one of the numbered tasks in your prompt, which is easy to catch. Implicit omissions: the AI addressed the question you asked but not the question you should have asked. A contract review that flags every deviation from your stated positions but misses a governing law clause that gives the counterparty unilateral amendment rights has failed on completeness even though it appears thorough.
The best single technique for catching implicit omissions: after reviewing the AI's output, ask the AI explicitly "what have I not asked that a cautious lawyer reviewing this document would want to know?" Run this as a separate prompt. The answers are frequently useful.
Check 4 — Jurisdiction verification. Has the AI applied the correct jurisdiction's law? This is particularly important if you did not specify jurisdiction in your prompt constraints — AI defaults to patterns in its training data, which skew heavily toward US law and common law jurisdictions. But it also matters even when you did specify jurisdiction: AI can acknowledge the specified jurisdiction and still apply legal concepts from a different one, particularly for areas where the training data is thinner. Always read the analysis with jurisdiction in your mind as a separate question from whether the reasoning is correct.
Check 5 — The "so what" check. Could you act on this output? Not "is this output accurate" — but "is this output sufficient to do what I need to do with it"? A research memo that accurately summarises five cases but does not synthesise them into an answer to the business question has failed the "so what" check. A contract review that correctly identifies three deviations from your standard positions but does not tell you which one to push back on first has failed it. The "so what" check is a professional judgment call that AI cannot make for you — but it is easy to forget to make it yourself when the output looks polished.
The five checks apply to every task, but the weight and specific application varies.
Contract review. Source verification is the primary check: does the AI's description of each clause actually match the clause? Logic verification focuses specifically on risk ratings — does each rating have a specific commercial reason, or is the AI applying generic "significant / minor / acceptable" labels without reasoning? Completeness verification requires checking your stated positions against the output systematically, not relying on the AI to have done so accurately. Jurisdiction verification matters especially if your contracts span multiple governing law regimes. The "so what" check asks whether the output, as produced, could go into a negotiation memo without rewriting.
Legal research. Source verification is non-negotiable and irreducible. Every case citation must be confirmed in a primary source database before the research memo leaves your desk. This is not a proportionality question — you cannot decide that some citations are more likely to be hallucinated than others. ABA Formal Opinion 512 is explicit on this point, and Mata v. Avianca (SDNY 2023) demonstrates the professional consequences of skipping it. Logic verification focuses on whether cases are being used to support propositions they actually stand for. Completeness verification asks whether the research answers the business question or merely the legal question. The "so what" check asks whether someone who read the memo would know what to do next.
Client and stakeholder advice. Logic verification is the primary check for advice: does the recommendation follow from the analysis? Are the assumptions stated and confirmed? Completeness verification asks whether the advice covers the question the client actually has, not just the question you answered. Source verification applies to any regulatory or statutory position described. The "so what" check is critical: client advice that is technically accurate but fails to tell the client what to do has not done its job.
ABA Formal Opinion 512 and Model Rule 5.3 both contemplate that supervision of AI output is a professional responsibility — which means the adequacy of your verification is not just a quality question but a professional conduct question. Documenting your verification is how you demonstrate adequacy.
Documentation does not need to be elaborate. For each significant AI-assisted piece of work product, a one-paragraph note in the matter file is sufficient: what tool was used, what tasks the AI performed, what verification steps were completed, and which lawyer signed off. Something like: "Contract review: AI (Gemini) used for initial risk extraction. Source verification completed against document clauses. All citations confirmed. Risk ratings reviewed and confirmed. Jurisdiction check completed — New York law applied as specified. Output used as basis for negotiation brief after substantive review. [Initials] [Date]."
This note takes ninety seconds to write. It is the difference between a documented professional process and an undocumented assumption that things were fine.
The five-point framework is not a bureaucratic add-on to AI-assisted legal work. It is the thing that makes AI-assisted legal work professional rather than speculative.
AI produces confident-sounding output. Confidence is not accuracy. The only way to know whether AI output is accurate is to check it — specifically, systematically, against sources that do not depend on the AI's own reasoning. That is verification. Re-reading is something else entirely.
The lawyers who build this discipline now — who develop a verification habit that is fast, consistent, and documented — will produce better work product than those who do not. They will also be demonstrably more compliant with the professional conduct obligations that ABA Formal Opinion 512 makes explicit. In 2024, that matters. In 2027, when AI is embedded in every major legal practice, it will matter more.
Model Rule 5.3 requires that supervising lawyers ensure work product from non-lawyer assistance meets professional standards. AI output is non-lawyer assistance. A supervising lawyer who accepts AI-assisted work product without adequate verification has not met the MR 5.3 standard — regardless of whether the output happens to be correct. Correctness is not the test; adequacy of supervision is.
For law firm and in-house team licences, CLE provider partnerships, bespoke playbook engagements, and tool company partnerships. We respond within 48 hours.
Wei Lee is a New York-admitted attorney. LexForward AI Lab is an educational product and does not constitute legal services or legal advice. No attorney-client relationship is created by purchase of or enrollment in this course. Attorney Advertising. Prior results do not guarantee a similar outcome.