# The verification skill: legal AI's premium billable work **TL;DR:** A model can produce a competent first draft of most legal documents in seconds, which has pushed the price of drafting toward zero. What has not gotten cheaper is establishing that the draft is correct: that the cited case exists, is still good law, stands for what the draft says it stands for, and applies to the facts at hand. That judgment call has become the scarce, billable skill in Indian legal practice, and it is reshaping training, billing, and professional risk in ways most firms have not priced in yet. --- ## On this page - [The trade that already happened](#the-trade-that-already-happened) - [What verification actually is](#what-verification-actually-is) - [Existence: does the authority exist at all](#existence-does-the-authority-exist-at-all) - [Currency: is it still good law](#currency-is-it-still-good-law) - [Ratio versus obiter: what the case actually decided](#ratio-versus-obiter-what-the-case-actually-decided) - [Applicability: does it fit these facts](#applicability-does-it-fit-these-facts) - [The rung that came off the training ladder](#the-rung-that-came-off-the-training-ladder) - [The hour that carries the risk is the hardest hour to bill](#the-hour-that-carries-the-risk-is-the-hardest-hour-to-bill) - [The professional duty is already settled](#the-professional-duty-is-already-settled) - [What a firm should actually do about it](#what-a-firm-should-actually-do-about-it) - [Drafting skill against verification skill](#drafting-skill-against-verification-skill) - [Frequently asked questions](#frequently-asked-questions) --- ## The trade that already happened Ask an AI legal drafting tool for a first draft of a bail application, a reply to a legal notice, or a research memo on whether a High Court order survives a later Supreme Court ruling, and a competent draft appears in under a minute. It will cite cases. It will quote sections. It will read like the work of a diligent junior who spent an afternoon on it. None of that took an afternoon. It took the time it takes to type a prompt. That collapse in drafting time is real and it is not reversible. What did not collapse alongside it is the separate task of establishing that what came out is actually correct: that the cited case is a real, reported judgment; that it has not been overruled since; that the quoted paragraph says what the draft claims rather than something narrower or contrary; and that the proposition, even if accurately stated, actually fits the facts it has been dropped into. That task took a competent lawyer real time before AI drafting tools existed, and it takes a competent lawyer real time now. The skill that commands a premium in Indian legal practice has moved from the first task to the second, and the shift has been quieter than the drafting speed-up that caused it. The clearest evidence for how serious the failure to do the second task can be is [*Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. & Anr.*](https://indiankanoon.org/doc/113338666/), 2026 INSC 668, decided by the Supreme Court of India on 2 July 2026. The National Company Law Tribunal had relied on six citations in an insolvency order. Three did not exist under the case names given. Two were genuine Supreme Court judgments, correctly cited, with quoted paragraphs that do not appear anywhere in them. The bench of Justice Pamidighantam Sri Narasimha and Justice Alok Aradhe set aside the order and the appellate judgment that followed it, restoring a Section 7 application filed under the Insolvency and Bankruptcy Code, 2016 to its original number. An admission order and an appellate ruling, spanning more than a year, were undone because nobody had opened six citations before relying on them. The mechanics of what those citations got wrong, and the rule the Court laid down, are covered in [the Supreme Court's 2026 ruling on AI-generated case law](/blog/sc-ai-generated-case-law-2026). What matters here is narrower: producing those six citations took the Tribunal's own research process minutes. Discovering that three were fake and two more were misquoted took the Supreme Court a full hearing, on appeal. That asymmetry, not the existence of AI drafting tools by itself, is the argument of this piece. --- ## What verification actually is "Verify the citations" sounds like a single task. In practice it is four distinct judgment calls, and a lawyer who is good at one of them is not automatically good at the rest. Confusing the four is how firms end up training juniors to run a fast existence check and stop there, which is exactly the pattern that let two of the six citations in *Pooja Ramesh Singh* survive as far as a tribunal order. | Verification task | What it answers | What it is not | |---|---|---| | Existence | ✓ Is there a real, reported judgment matching this name, court, and year | ✗ Whether the paragraph quoted from it says what the draft claims | | Currency | ✓ Has this judgment been overruled, reversed, or superseded by a later amendment | ✗ Whether the case was ever correctly decided in the first place | | Ratio versus obiter | ✓ Is the quoted proposition the actual basis for the court's decision | ✗ Whether the sentence appears in the judgment at all | | Applicability | ✓ Does the ratio, correctly stated, actually govern these facts | ✗ Whether the citation is otherwise sound in the abstract | Each of the four rows below is a separate skill with its own failure mode, and each takes a different kind of expertise to run well. --- ## Existence: does the authority exist at all This is the cheapest of the four checks and the one most verification workflows stop at. It asks a narrow question: does a document matching the case name, court, and year the draft gives you actually exist in a primary source. For Supreme Court judgments, that means checking the [Supreme Court of India's e-SCR portal](https://scr.sci.gov.in/) or the [Court's own judgment search](https://www.sci.gov.in/), which carry the neutral citation format the Court assigns itself. For High Court and subordinate court decisions, the [National Judicial Data Grid's judgment search](https://judgments.ecourts.gov.in/) is the primary route. Existence checking is close to clerical work, and that is precisely why it is the least defensible thing to bill a senior lawyer's time for and the easiest thing to delegate to a junior, an assistant, or a well-built research tool. It is also, on its own, insufficient. In *Pooja Ramesh Singh*, three of the six disputed citations failed at exactly this step, either because they matched no real judgment or because a real reporter citation had been repurposed under the wrong cause title. Those three were the easy catches. The two that survived a name-and-year check, Everest Kento Cylinders Ltd. v. Union of India and Canara Bank v. N.G. Subbaraya Setty & Anr., were genuine, correctly cited Supreme Court decisions carrying invented paragraphs. A verification process built entirely around the existence question passes both. The full mechanics of running this check without missing the harder half of it are set out in [a five-step protocol for verifying AI output before it leaves your desk](/blog/legal-verification-workflow-ai-output), and the broader pattern of what gets fabricated and why is covered in [AI hallucinated citations in India](/blog/ai-hallucinated-citations-india). The distinction between a primary source and a search engine's summary matters here. [Indian Kanoon](https://indiankanoon.org/) indexes a wide range of Indian judgments and is a fast first pass, but a citation load-bearing enough to put in front of a bench needs confirmation against the court's own record, not against a third-party index alone. A tool that attaches a source link to every generated proposition, the way [Niyam](https://niyam.ai) does across its judgment corpus, turns this step into a click rather than a fresh search, though the click still has to be opened and read. The difference between primary and secondary legal sources, and when each is safe to rely on, is set out in [primary versus secondary legal sources](/blog/primary-vs-secondary-legal-sources). --- ## Currency: is it still good law A citation can be entirely real, correctly quoted, and still be dead law. This is not an AI-specific problem. It predates AI drafting tools by as long as Indian law reports have existed, and it is the check most likely to be skipped by a lawyer confident that a well-known case cannot possibly have moved. Currency checking asks whether the judgment has been overruled outright, reversed on appeal, referred to a larger bench that has since ruled differently, or so consistently distinguished on its facts that the proposition drawn from it no longer carries the weight the draft assumes. Within the Supreme Court, bench strength governs: a two-judge bench decision yields to a conflicting three-judge bench decision on the same point, and a lawyer relying on the smaller bench without disclosing the conflict has a candour problem, not just an accuracy one. The full taxonomy of overruled, distinguished, doubted, and statutorily superseded, and how to check for each, is set out in [good law checking in Indian legal research](/blog/good-law-checking). This check has a second, India-specific layer that did not exist before 1 July 2024. On that date, the Bharatiya Nyaya Sanhita, the Bharatiya Nagarik Suraksha Sanhita, and the Bharatiya Sakshya Adhiniyam replaced the Indian Penal Code, the Code of Criminal Procedure, and the Indian Evidence Act. A model trained substantially on material predating the transition can produce a section number that was correct under the old code and means something different under the current one, without any hallucination in the ordinary sense: the number is real, it just governs a different offence now. Checking a statute reference for currency means confirming which code it belongs to and what the section currently says, not just that a matching number exists somewhere. The renumbering between the old and new codes is mapped in [the BNS to IPC section mapping](/blog/bns-ipc-section-mapping), [the BNSS to CrPC section mapping](/blog/bnss-crpc-section-mapping), and [the BSA to Evidence Act section mapping](/blog/bsa-evidence-act-section-mapping). Currency checking is slower than existence checking and requires more judgment, because a case that has been distinguished on facts close to yours is not a clean fail the way a fabricated case is. It is a soft fail: usable with disclosure, not usable as unqualified authority. Telling the difference is a substantive legal judgment, not a database lookup, which is why this task resists full delegation to a junior far more than existence checking does. --- ## Ratio versus obiter: what the case actually decided This is the hardest of the four checks and the one that separates a competent verifier from a merely diligent one. A judgment can exist, be correctly quoted word for word, and still be misused if the quoted sentence was never part of the court's actual reasoning for the outcome. Courts write broadly in the course of deciding narrowly. A line of general commentary, a hypothetical the bench raised to test an argument, or a passing reference to a point not argued before it can all be technically present in a judgment and still carry none of the binding weight the draft attaches to them. Distinguishing a holding from a passing remark requires reading the judgment as a whole, not the isolated paragraph a citation points to. It means asking what result the court actually reached, whether the cited proposition was necessary to reach that result, and whether the surrounding paragraphs support or undercut the reading a draft has given it. This is the step where a fabricated paragraph and a genuine but misapplied one start to look alike, because both present as a quotation attributed to a real case. The technique for reading a judgment this way is covered in [how to read a judgment](/blog/how-to-read-a-judgment), and the citation conventions flagging which part of a judgment a pinpoint reference draws from are in [how to cite Indian judgments](/blog/how-to-cite-indian-judgments). No existence check and no currency check catches a misapplied ratio, because a real holding and a stray remark pass both equally well. Only a lawyer who has read the reasoning, not just the extracted sentence, can tell the two apart. This is also the check an AI drafting tool is structurally worst at running on its own output, since a model trained to produce fluent, on-topic text has no independent way to weigh whether a sentence it generated was load-bearing in the source judgment or incidental to it. --- ## Applicability: does it fit these facts The fourth check is the one that turns a correctly verified proposition into usable legal argument for a specific matter, and it is the check most likely to be skipped once the first three have passed, because by that point the citation feels finished. It is not. A correctly stated ratio from a genuinely good-law judgment can still be the wrong authority for the facts in front of you, if the underlying facts diverge in ways that matter to how the court reasoned. This is where district-level and High Court research differs sharply from Supreme Court research, because factually close precedent is thicker at the district level and the discipline of matching facts before matching propositions matters more. The workflow a district court practitioner runs to keep this straight, rather than reaching for the first apparently on-point authority, is set out in [a district court lawyer's research workflow](/blog/district-court-lawyer-research-workflow). Applicability checking is also where a lawyer's judgment about the client's actual position, not just the abstract legal question, does work no verification tool can substitute for. A model can tell you a case exists and, with the right design, that it has not been overruled. It cannot tell you whether the facts of your client's cheque bounce case are close enough to a cited Section 138 precedent that the same reasoning should govern, or different enough that citing it invites a distinguishing argument that damages your position instead. --- ## The rung that came off the training ladder The traditional apprenticeship model for training a junior litigator or corporate associate ran through the search itself. A senior lawyer would assign a research question, the junior would spend hours or days finding, reading, and sorting authorities, and competence was built through repetition: enough searches, enough dead ends, enough judgments read cover to cover, until pattern recognition for what a strong authority looks like became instinct rather than a checklist. AI drafting tools remove the rung the apprenticeship was built on. A junior asked to research a question today can produce a plausible-looking memo with citations in minutes, without having done any of the searching that used to build the underlying pattern recognition. The memo can look like the product of a diligent afternoon of research and be, in substance, the product of five minutes of prompting followed by zero minutes of verification. A firm that measures junior output by how polished the memo looks, rather than by how the junior arrived at it, is training juniors to produce fluent-looking work without the underlying judgment that used to come bundled with the search process. The honest response is not to ban AI drafting tools from a junior's desk. It is to separate the two skills explicitly in how juniors are trained and evaluated: unaided research competence has to be built and tested before AI-assisted drafting is introduced, because a junior who has never run a manual existence check, currency check, and ratio-versus-obiter read has no internal benchmark for judging whether an AI-drafted memo's citations are behaving normally or are quietly wrong. A structured approach to building this, including what a junior should manage unaided before touching an AI tool and how to review a memo in a way that tests judgment rather than polish, is set out in [training junior associates for legal research in the AI era](/blog/legal-research-training-junior-associates). A junior trained the old way and one trained the new way can, at first glance, produce indistinguishable memos. The gap shows up only when something goes wrong: a citation turns out to be dead law, a quoted paragraph turns out not to exist, or an on-point authority rests on facts different enough that it does not govern the matter. A junior who built pattern recognition through unaided search catches these failures on sight, the way an experienced reader catches a sentence that does not fit its paragraph. A junior who has only ever reviewed AI output for surface polish does not have that instinct, because it was never built. --- ## The hour that carries the risk is the hardest hour to bill Clients have an intuitive sense of what drafting time is worth, because they can see the output: a document exists that did not exist an hour ago. Verification produces no new document. An hour spent confirming that six citations are real, current, correctly read, and applicable to the facts produces, if the work is done well, no visible change to the draft at all. The citations were already sitting in the document before the hour started and are still sitting there after it ends. To a client looking at an invoice, that hour looks like nothing happened. This is backward from where the risk sits. The drafting hour is the one AI has made cheap to redo if something is wrong with it. The verification hour is the one that, if skipped or rushed, produces exactly the outcome in *Pooja Ramesh Singh*: an order set aside, a proceeding restored to its starting point, and a client who paid for a resolution that has to happen again. The hour hardest to justify on an invoice is the hour that was actually protecting the client. Firms that measure this seriously find that verification time does not shrink the way drafting time did, and in some matters it grows, because the volume of AI-generated propositions needing a check has risen even as the time to produce each one has fallen. Treating the speed gained at the drafting stage as time saved on the matter overall is the mistake to avoid: the time has not disappeared, it has moved further down the workflow. A firm's own honest accounting of where the hours went, not a vendor's marketed time-savings percentage, is the only way to know whether AI use is a net gain on a given matter. The mechanics of running that measurement, including the counter-case where verification eats the savings entirely, are in [legal AI return on investment for an Indian law firm](/blog/legal-ai-roi-law-firm-india). Relabelling verification time as drafting time on an invoice does not solve the problem, because it invites exactly the scrutiny a client is entitled to ask for: what, specifically, did this hour verify. Naming the work for what it is, on the invoice and in the engagement letter, is the fix. A client paying for legal advice should understand that the hour spent confirming a citation is real and still good law is not overhead on top of the legal work. It is the legal work that used to be invisible inside a longer drafting process and is now visible on its own, because the drafting around it got fast enough to expose it. --- ## The professional duty is already settled None of the argument above depends on whether a firm decides verification is worth billing properly. The duty to verify sits with the lawyer regardless of what produced the draft, and Indian law has already stated this plainly rather than leaving it as an open question. The Supreme Court's holding in *Pooja Ramesh Singh* is direct on this point. At paragraph 7 of the judgment, the Court said it is necessary for courts to adopt a zero-tolerance mode for producing, citing, or using AI-generated precedents without verification, and that citing such material without verification is misconduct on the part of an advocate. It went further on the consequence: a decision resting on fake or hallucinated material is no decision in the eyes of the law, irrespective of whether the material had a direct or indirect bearing on the outcome, and such decisions are to be set aside even if only an iota of fabricated material entered the reasoning. Read the operative paragraphs in full at [the Court's own judgment portal](https://scr.sci.gov.in/) or the [reported text on Indian Kanoon](https://indiankanoon.org/doc/113338666/). Paragraph 16 of the same judgment records a detail that removes any comfort a lawyer might take from the source of the fabrication in that particular case: the party that benefited from the false citations filed an affidavit stating its own counsel had not cited them, and that the tribunal had generated them through its own research. The fabrication came from the adjudicator, not the advocate, and the Court still treated verification as a standing obligation running through every stage of the process, not a duty that shifts depending on who or what produced the error. That is the position worth stating plainly rather than hedging: a lawyer who files a fabricated citation is answerable for it regardless of what produced it. It does not matter whether a paralegal typed it, an AI tool generated it, or opposing counsel's filing planted the idea for it. The name on the filing carries the responsibility for what is in it. This is not a novel rule invented for the AI era; it restates, in the context of AI-generated material, a professional duty of competence and candour to the court that predates any AI drafting tool by decades. What has changed is the volume of citations a single lawyer can now generate without personally having searched for a single one of them, which raises the stakes on an old duty rather than creating a new one. The professional duty in full, including how it interacts with an AI tool's own claims about its accuracy, is covered in [a lawyer's duty to verify AI output](/blog/lawyer-duty-verify-ai-output). Bar disciplinary bodies exist to enforce this kind of duty, and the Supreme Court's judgment directed the Bar Council of India to constitute a committee on this question rather than treating a prohibitory declaration as sufficient on its own. A firm waiting for a formal, published rule with a specific number before treating verification as mandatory is waiting for something that arrives after, not before, the duty already binds it. --- ## What a firm should actually do about it The argument above points to a small number of concrete changes, not a wholesale restructuring of how a firm operates. Separate drafting time from verification time on the timesheet, not just in internal thinking about workflow. If an associate spends forty minutes drafting a memo with AI assistance and ninety minutes verifying six citations in it, both numbers should appear as distinct, named entries. This gives a firm honest data about where its hours go, and gives a client a legible account of what they are paying for, rather than one blended number. Write the verification obligation into a policy document the firm can point to, rather than leaving it as an unwritten expectation that varies by which partner is supervising a given matter. What belongs in that document, clause by clause, including mandatory human verification before filing, disclosure obligations to courts and clients, and sanctions for a breach, is set out in full in [a law firm AI use policy for India](/blog/law-firm-ai-use-policy-india). Choose a research tool for its verification affordances, not for how fast it drafts. A tool that links every generated proposition back to the specific judgment it drew from shortens the existence and ratio checks, because the source document is one click away instead of a fresh search. A citator that flags adverse subsequent treatment automatically shortens the currency check, provided the flagged treatment is still read rather than accepted on faith. A firm evaluating any legal AI vendor should ask specifically how the tool supports verification rather than how quickly it drafts; [Niyam](https://niyam.ai), for instance, attaches a source citation to every retrieved proposition, which is the affordance that matters for this purpose, not the drafting speed. The comparative mechanics of the major Indian legal research platforms on this question are covered in [choosing an Indian case law search engine](/blog/choosing-indian-case-law-search-engine) and [SCC Online versus Manupatra versus Indian Kanoon](/blog/scc-online-vs-manupatra-vs-indian-kanoon), and the limitations a free-tier search tool carries into a verification workflow are set out in [the limitations of Indian Kanoon for legal research](/blog/indian-kanoon-limitations-legal-research). Train juniors on unaided research before AI-assisted drafting, in that order, and evaluate a memo on how the citations were checked, not only on how the final draft reads. A solo practitioner without a training pipeline to redesign faces a narrower version of the same problem, running all four checks personally rather than delegating any; the adjustments that make that workable without a full associate team are in [legal AI for solo practitioners in India](/blog/legal-ai-for-solo-practitioners-india). None of this requires waiting for a further judgment, a further circular, or a further vendor feature. The duty already exists. The only open question for a given firm is whether its billing, training, and tooling decisions currently reflect where the actual legal work, and the actual risk, now sits. --- ## Drafting skill against verification skill The table below states the shift plainly, stripped of hedging, because the two columns describe genuinely different work rather than two speeds of the same work. | | Drafting skill | Verification skill | |---|---|---| | What it produces | ✓ A new document that did not exist before | ✗ No new document; confirms what is already there | | Time trend since AI drafting tools arrived | ✗ Falling, sharply and continuously | ✓ Flat or rising, as volume of propositions to check has grown | | Where the risk sits if the work is skipped | ✗ A weak first draft, cheap to redo | ✓ A filed or advised-on error, expensive and sometimes irreversible | | Client's visibility into the work | ✗ High; a document appears | ✓ Low; nothing visibly changes if done correctly | | Whether AI can do it unsupervised | ✓ Largely yes, for a first pass | ✗ No; requires reading primary sources and legal judgment | | Whether it builds junior competence through repetition | ✗ Less so; a prompt substitutes for the search | ✓ Yes; each check is a small, repeatable exercise of judgment | | Professional duty attached | Standard duty of competence in drafting | Duty of candour to the court, restated directly by the Supreme Court in *Pooja Ramesh Singh* | --- ## Frequently asked questions ### Is verification really a separate skill from research, or just a slower version of it? It is a separate skill because it asks a different question. Research asks what authority exists and might support a proposition. Verification asks whether a specific, already-drafted proposition is actually true: real, current, correctly quoted, and applicable. A lawyer can be excellent at finding authority and weak at checking whether authority someone else, including an AI tool, has already assembled is sound. ### Does this apply equally to litigation and transactional practice? The four checks apply differently but not less. A transactional lawyer verifying a due diligence memo checks statutory currency and applicability to the deal's facts at least as heavily as a litigator checking case law; a contract clause drafted against a superseded regulation is the transactional equivalent of a citation to overruled case law. The applicability check in particular carries more weight in transactional work, where deal-specific facts determine whether a precedent structure actually fits. ### Why did drafting get cheap but verification did not? Drafting is pattern generation: producing text that resembles competent legal writing, which is exactly what large language models are built to do well. Verification requires checking a specific claim against an external, authoritative source and exercising legal judgment about whether it holds, which is a different kind of task that speed in text generation does not shorten. ### How does the BNS, BNSS, and BSA transition specifically affect verification workload? It adds a currency-check failure mode unique to the post-1 July 2024 period: a section number correct under the old codes and wrong, or governing something entirely different, under the current ones. This is not a fabricated citation in the ordinary sense, since the number is real; it is a real number attached to the wrong statute, and it requires checking against the current text rather than trusting that a familiar-looking number is still correct. ### Can a firm bill the verification hour at the same rate as the drafting hour? There is no rule against it, and the argument in this piece supports doing so, since the verification hour is where the legal judgment and the risk actually sit. The practical obstacle is client expectation, not any rule, which is why naming the hour explicitly on an invoice, rather than folding it into a blended drafting entry, matters more than the rate charged for it. ### What is the difference between a citation being distinguished and being fabricated, for verification purposes? A distinguished citation is real and correctly quoted; a later court has held it does not apply to different facts, so it remains good law on its own facts and fails only the currency or applicability check, often usable with disclosure. A fabricated citation fails the existence check outright and cannot be used in any form. ### Does a research tool that shows sources solve the verification problem on its own? It shortens the existence and ratio checks by putting the source document one click away instead of requiring a fresh search, but it does not remove the obligation to read the source and exercise judgment about currency and applicability. A tool that shows a source and a lawyer who does not open it has the same exposure as a lawyer working from an unsourced draft. ### Who is responsible if an AI tool's citation turns out to be fabricated and it reaches a court filing? The lawyer who signed the filing, without exception. The Supreme Court's holding in *Pooja Ramesh Singh* treats citing unverified material as misconduct on the part of the advocate regardless of whether the fabrication originated with counsel, an assisting AI tool, or, in that case, the adjudicating tribunal's own research process. ### Is this a uniquely Indian problem, or does it show up in other jurisdictions too? Courts in multiple jurisdictions have addressed sanctions or adverse findings arising from AI-fabricated citations in filings, generally on the same underlying pattern: a model produces a plausible but non-existent or misquoted authority, and counsel files it without independently checking a primary source. The specific Indian development is the Supreme Court's explicit zero-tolerance rule in *Pooja Ramesh Singh* and the direction to the Bar Council of India to build accountability mechanisms around it. ### How much of a junior's training time should go to unaided research versus AI-assisted work? There is no fixed ratio that fits every firm, but the sequencing matters more than the split: a junior needs to demonstrate unaided competence on the four verification checks before being trusted to review AI-assisted drafts for the same checks, because reviewing requires an internal benchmark for what correct looks like that only unaided practice builds. ### Does verification time ever go down as a lawyer gets more experienced with it? Yes, for the existence and currency checks on familiar categories of authority, where an experienced lawyer recognises quickly whether a citation looks routine or needs the full treatment. The ratio-versus-obiter check and the applicability check resist this kind of speed-up more, because each new matter presents a fresh set of facts and a fresh judgment to read, regardless of how many prior matters the lawyer has verified citations for. ### What should a client actually be asking a firm about its verification practice? Whether the firm has a written policy requiring human verification of AI-assisted drafts before filing, who performs that verification and at what seniority, and whether the invoice separates verification time from drafting time. A firm that cannot answer the third question clearly likely has not separated the two internally either. ### Is there a shortcut that makes the ratio-versus-obiter check faster without losing rigour? Reading the paragraph before and after a cited passage, rather than the cited sentence in isolation, catches most misapplications quickly, because a proposition lifted out of surrounding context usually reveals itself once the surrounding text is read. There is no shortcut that replaces reading the reasoning; the technique only shortens how much surrounding text needs to be read before the answer is clear. ### Does the professional duty to verify differ for a junior associate acting under a partner's instructions? No. The duty attaches to the person who signs or is responsible for the filing, but a junior who prepared a draft the partner relies on without independent verification is not shielded by having acted under instruction, and a partner who signs a filing without reviewing a junior's verification work carries the exposure that comes with the signature. Seniority changes who is ultimately accountable when something goes wrong; it does not remove the duty from anyone in the chain. ### If verification cannot be automated, is there any point using AI tools for legal research at all? The value of an AI tool in this workflow is compressing the existence and ratio checks, since a tool with source attribution puts the primary document in front of a lawyer faster than a manual search would, and a well-built citator can flag adverse subsequent treatment that a manual currency check might otherwise miss on a fast read. The point is not automating verification away; it is making the four checks faster to run without removing the requirement to run them. ### How does a firm know if it is actually charging for verification or just calling drafting time something else? Look at whether the invoice or internal time record distinguishes a distinct verification entry with its own description, separate from the entry describing the drafting work itself. If verification only ever appears as part of a longer drafting narrative, the firm has not separated the two in practice, whatever its stated policy says.