In June 2023 a federal judge in Manhattan fined two lawyers $5,000 for citing six cases that never existed — Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. 2023). Since then I've had at least a dozen managing partners open our first call the same way: 'We're not doing that.' They mean AI drafting. What they're actually describing is fear of one documented failure mode, not a considered decision about whether the tool works.
I'm William Vásquez. Before HODOS, I ran a law firm for fifteen years and drafted the same motions, retainer letters, and immigration petitions your associates draft every week. A computer science degree came first, then a J.D. and fifteen years of practice, an M.Div. somewhere in the middle, seven years in the Air Force before any of that — none of it taught me how to run a business. I learned Scaling Up the hard way, after I'd already made most of the expensive mistakes a firm makes when nobody's watching the P&L.
The question in front of you this week isn't whether AI drafting software works. Vendors have public numbers on that. It's whether your firm has the Execution discipline to put a verification step around it before your name goes on the filing.
AI legal drafting software uses large language models to generate first-draft contracts, pleadings, and correspondence from firm precedent and prompts, then routes that draft through human attorney review before filing. It does not replace an attorney's independent judgment or the bar-admission requirement; FRCP 11(b) still holds the signing attorney personally responsible for every citation.
The Hard Truth
Will AI Drafting Software Get My Associate Sanctioned?
Not on its own. It gets an associate sanctioned when nobody reads the citations before they go out under a signature. That's the naive version of the Execution decision most firms default to: draft it, have someone 'review' it, file it. That workflow assumed the failure mode was a typo or a misquoted holding — the kind of error a normal read-through catches. Generative drafting tools break that assumption because the failure mode changes shape. A hallucinated citation reads exactly like a real one: same format, same confident tone, same plausible court and year. Readability review doesn't catch it. Existence review does, and that's a different task requiring a different tool, not a sharper associate.
A case tracker maintained by legal researchers, widely referenced in legal press since Mata v. Avianca (No. 22-cv-1461, S.D.N.Y. 2023), had logged well over a hundred documented instances of fabricated case citations in U.S. court filings by 2024, spanning federal and state courts, plaintiff and defense counsel, solo practitioners and firm associates alike. FRCP 11(b) puts personal responsibility on the signing attorney regardless of what tool produced the draft — the rule has no software carve-out. ABA Formal Opinion 512 (July 2024) says the same thing in ethics terms: a lawyer using generative AI must still exercise independent professional judgment consistent with Model Rule 1.1, Comment 8, on technology competence.
Utah went further than bar guidance. Its House Bill 452, the Artificial Intelligence Policy Act (2024), became the first state statute imposing AI-disclosure obligations directly on regulated professions, legal services included — worth watching if your firm operates in more than one state, since it's unlikely to stay a one-state rule.
Meanwhile adoption is real but partial. The ABA's 2023 Legal Technology Survey Report found only about one in nine responding lawyers reported using generative AI tools in their practice, which means most firms are making this decision later than they think, without the two years of case law the early adopters already had to learn from.
What Happens If You Wait
What Happens If My Firm Waits to Adopt AI Drafting Tools?
Waiting doesn't freeze your risk in place. It changes which risk you're carrying.
Malpractice underwriting is starting to ask. ALPS Corporation, one of the larger legal malpractice carriers, published risk-management guidance in 2024 telling policyholders to expect underwriting questions about AI tool use and firm-level verification policies — the same way carriers already ask about conflicts-checking software and trust accounting controls. A firm with no answer to 'how do you verify AI-drafted citations' looks worse on that questionnaire than a firm with a documented policy, regardless of whether either firm has actually had a problem.
Confidentiality exposure doesn't wait for a decision meeting. Model Rule 1.6 and its state analogs require reasonable efforts to prevent unauthorized disclosure of client information — pasting a client's settlement demand into a consumer-grade chat tool with no data-processing agreement is a disclosure problem the moment it happens, not a hypothetical for later. SOC 2 Type II certification (AICPA Trust Services Criteria) and the newer ISO/IEC 42001:2023 AI-management-system standard exist so a firm can point to a third-party audit instead of taking a vendor's word for how client data is stored, trained on, or retained. Ask any vendor for the actual report, not the badge on their homepage — badges get left on homepages long after certifications lapse.
Time saved compounds against firms that don't capture it. Vendors report drafting-time reductions in the range of 25-50% on first-draft contract and pleading work (per published case studies from Ironclad and CoCounsel, 2024), and that differential shows up downstream in your cash conversion cycle — faster first drafts mean faster client turnaround, faster billing, faster realization. A firm still drafting everything from a blank page in 2025 isn't only slower on the document. It's slower to bill, which is a Cash decision wearing an Execution costume.
Step-by-Step Process
How Do I Vet an AI Legal Drafting Vendor Before Rollout?
Court rules on AI disclosure aren't uniform, which is exactly why a vetting process has to be built once and reused, not improvised per filing.
| Court | Requirement | Since |
|---|---|---|
| N.D. Tex. (Judge Brantley Starr's chambers, Dallas Division) | Mandatory Certification Regarding Generative Artificial Intelligence — counsel must certify either that no portion of a filing was drafted by generative AI, or that a human verified every citation against a reporter or database | June 2023 |
| Other federal districts | No uniform circuit-wide rule as of this writing; a growing number of individual judges have issued similar standing orders since Judge Starr's — check each assigned judge's standing orders page before every filing | Ongoing since 2023 |
| Most state courts | No AI-specific standing order identified as of this writing; general state Rule 11 analogs and bar ethics guidance (California's COPRAC, the NYSBA Task Force on Artificial Intelligence's 2024 report) govern instead | N/A |
One. Confirm the vendor's grounding method before anything else. Ask whether outputs are checked against a live citation database — the way CoCounsel grounds against Westlaw, or Lexis+ AI validates against Shepard's Citations — or whether the model generates from training data alone with no real-time verification layer. That answer eliminates half the vendor field for litigation drafting.
Two. Get the actual SOC 2 Type II report and ISO/IEC 42001:2023 certificate, not marketing copy referencing them. Read the exceptions section of the SOC 2 report — a clean report has no client-data exceptions relevant to your practice area.
Three. Confirm Model Rule 5.3 coverage. Non-lawyer assistance rules require a lawyer to make reasonable efforts to ensure a tool's output conforms to the lawyer's professional obligations, which means a written internal policy assigning a specific person, not 'the team,' to sign off on AI-drafted filings.
Four. Check integration against what your firm already runs. Whatever practice-management system you use, confirm a real API integration exists rather than manual export and import — manual handoffs are where version-control errors creep in.
Five. Run a 90-day pilot on a bounded document type. Immigration petition drafting (I-130 supporting statements, I-589 declarations) is a reasonable pilot category because the factual record is well-defined and the citation burden is lower than litigation briefing.
Six. Before firm-wide rollout, notify your malpractice carrier in writing of the tool and your verification policy. This is the Execution decision made visible: a documented procedure, not a verbal understanding that 'someone checks this.'
This is where the Witness pillar of the Fifth Decision does actual work, not as a slogan. Witness means every AI-generated clause or citation carries a signed, timestamped audit chain back to its source document and the specific human reviewer who approved it — a log entry a malpractice carrier or a court can request, not a screenshot. The risk it protects against is specific: at a bar inquiry or a Rule 11 hearing, 'someone checked that' is not an answer. A timestamped reviewer sign-off tied to a specific citation is. Judge Starr's certification requirement and FRCP 11(b) are both, functionally, asking for a Witness record — neither rule cares what you call your internal process, both care whether you can produce one when asked.
A Real-World Example
A Composite Rollout: What a 90-Day Pilot Actually Looks Like
Consider a composite, twelve-partner firm — not a real client, a composite built from patterns common across firms this size — piloting AI drafting software for its immigration practice group in early 2024.
Week one: the managing partner and one immigration partner selected a single document type for the pilot, I-130 supporting statements, specifically because the underlying facts are bounded and the citation burden is lower than motion practice. They did not pilot litigation drafting first.
Weeks two through four: the firm requested the vendor's SOC 2 Type II report and confirmed no data-processing exceptions touched client PII retention. They wrote a one-page internal verification policy assigning citation review to a specific senior associate by name.
Weeks five through ten: associates drafted first-pass petitions using the tool, then ran every factual assertion and cited authority through the assigned reviewer before a partner's final read. Reviewer sign-off timestamps were logged in the practice-management system rather than tracked through email.
Weeks eleven through thirteen: the firm compared drafting-time logs from the pilot period against the same period the prior year, reviewed how many revision rounds each draft required, and evaluated whether the verification workload was sustainable at current staffing before deciding whether to extend the pilot to a second document type.
That last step is the one most firms skip. A pilot isn't complete when the software produces a usable draft. It's complete when the firm has actual data on whether its verification process holds up at volume, which is an Execution question, not a technology question.
About the Author
William J. Vásquez built HODOS after fifteen years practicing law, several of them as managing partner responsible for his own firm's P&L, staffing, and technology decisions. His academic background is a BS in Computer Science followed by a J.D. and, later, an M.Div.; he served seven years in the Air Force before any of it. He holds no formal AI or data-science certification, and this piece does not claim one on his behalf — the operational discipline described here (Scaling Up's four decisions, OPSPs, weekly scorecards) came from running a real firm imperfectly for a while, then correcting it deliberately.
Disclosure: HODOS is the software platform William built from that experience, and this article discusses AI legal drafting tools as a category, including HODOS's own approach and named competitors, to help managing partners evaluate the category, not as an independent, vendor-neutral product review. Get pricing and technical specifics, including SOC 2 reports and model-grounding methodology, directly from each vendor before deciding.
Bar admission and firm-specific details available on request through the consultation link below.
Key Terms Explained
Glossary
FRCP 11(b) — Federal Rule of Civil Procedure provision requiring an attorney who signs a filing to certify, after reasonable inquiry, that legal contentions are warranted and factual contentions have evidentiary support. Applies regardless of whether a human or an AI tool drafted the language.
Hallucination — An AI output that states a fact, citation, or holding that does not exist or does not say what the output claims, presented with the same confidence and format as accurate output, which is why a normal readability review misses it.
SOC 2 Type II — A third-party audit report, issued under the AICPA's Trust Services Criteria, verifying that a vendor's data-security controls operated effectively over a period of time, commonly six to twelve months, rather than a single point-in-time snapshot.
ISO/IEC 42001:2023 — The first international standard for AI management systems, covering governance, risk management, and ongoing monitoring of AI tools inside an organization.
Unauthorized Practice of Law (UPL) — State-specific statutes restricting legal advice and representation to licensed attorneys; relevant here because the software cannot hold a license and cannot substitute for the licensed attorney's review and sign-off.
Model Rule 5.3 — ABA Model Rule requiring a supervising lawyer to make reasonable efforts to ensure a non-lawyer assistant's conduct is compatible with the lawyer's professional obligations; ABA and several state ethics opinions extend this framework to AI tools functioning as non-lawyer assistance.
Standing Order — A judge's binding procedural requirement applying to all filings before that judge, separate from federal or state rules of civil procedure. Judge Brantley Starr's Mandatory Certification Regarding Generative Artificial Intelligence (N.D. Tex., June 2023) is the most cited example.
Witness (Fifth Decision sub-pillar) — A signed, timestamped audit chain linking an AI-generated draft element back to its source and the human reviewer who approved it, built to be producible at a bar inquiry, malpractice claim, or Rule 11 hearing.
Frequently Asked Questions
Frequently Asked Questions
Isn't AI drafting software just the first step toward replacing lawyers? The objection has a real observation inside it — document assembly that used to take an associate two hours can now take twenty minutes for a first pass. Where the objection breaks is the leap from 'drafts faster' to 'replaces judgment.' Nothing in FRCP 11(b), ABA Formal Opinion 512, or any state bar opinion issued so far treats AI output as self-certifying. Each one puts the licensed attorney's signature, and the review behind it, at the center of the filing. A faster first draft still needs the same judgment call it always did about whether the argument fits this client, this judge, this record — that call doesn't get faster because the typing did.
What does AI legal drafting software typically cost? Entry-level, template-based drafting tools typically list per-seat monthly pricing in roughly the $99-$400 range depending on plan (per public vendor pricing pages, 2025). Enterprise platforms like CoCounsel and Ironclad generally quote custom enterprise contracts rather than public per-seat pricing, often running into five or six figures annually depending on firm size and document volume. Get a quote in writing tied to your actual seat count and document volume before comparing vendors on price alone.
How long does implementation actually take? A bounded pilot on a single document type typically runs 60-90 days from vendor selection to a firm decision on whether to expand. Firm-wide rollout across multiple practice groups, with integration into existing practice-management software, commonly takes two to four additional months beyond the pilot, depending on how much of your document library needs preparing as reference material.
Does it integrate with the practice-management software we already use? That depends on the vendor, and it's worth confirming before signing anything. Several drafting tools offer native integrations with common practice-management platforms for document storage and matter linking; others require manual export and import, which reintroduces the version-control risk the tool was supposed to reduce. Ask for a live demonstration of the specific integration, not a slide describing it.
Who's liable if the AI drafts a bad citation and my associate misses it? The signing attorney, under FRCP 11(b), regardless of what tool produced the draft. The rule has no software exception. Firm-level liability can also attach under negligent-supervision theories if the firm had no verification policy at all — the gap a written policy and audit trail are built to close.
Do we have to disclose AI use to the court or the client? To the court: depends on the assigned judge's standing orders — check each judge's individual page, not just district-wide local rules. To the client: ABA Formal Opinion 512 (July 2024) and a growing number of state ethics opinions lean toward disclosure as good practice for billing transparency and informed consent, even where no mandatory-disclosure rule exists yet.
Does using AI to draft privileged documents waive privilege? Not inherently, but it depends on the vendor's data-handling architecture. If a vendor trains its shared model on client inputs, or a third party without a proper confidentiality agreement can access draft content, that can create a waiver argument that wouldn't exist with an internal drafting process. Confirm contractually that your firm's documents are never used to train a vendor's shared model.
What Firms Report After Adopting a Verification-First AI Drafting Process
Independent survey data, not vendor marketing, gives the clearest read on where the profession stands. The ABA's 2023 Legal Technology Survey Report found generative AI use still concentrated in larger firms, with solo and small-firm adoption trailing behind despite those firms often having the most to gain from time savings on routine drafting. The NYSBA Task Force on Artificial Intelligence's 2024 report found similar hesitation tied directly to verification uncertainty, not skepticism about whether the tools work. Stanford Law School's CodeX center and Georgetown Law's Center on Privacy & Technology have each published research tracking adoption and risk patterns independently of vendor marketing, and both point the same direction: the technology outpaces firms' internal verification policy, not the other way around.
Firms that go through a structured rollout, with a written verification policy and vendor SOC 2 documentation on file, consistently report the conversation with their malpractice carrier and their more risk-averse partners gets shorter once there's a document to point to instead of a verbal assurance. That's a modest claim, and it's the honest one — the tool doesn't sell itself into a firm on speed alone. The paperwork around it is what gets a skeptical name partner to sign off.
Where This Fits Into Your Firm's Broader Operating System This Week
AI drafting software is one decision inside a larger one. The question underneath 'should we adopt this tool' is really 'does our firm have an Execution system that can absorb a new tool without losing control of it' — the same question that governs your meeting rhythm, your KPI scorecard, and your cash conversion cycle. Firms already running a documented OPSP and a weekly scorecard tend to fold a new drafting tool into an existing verification checkpoint. Firms without one tend to bolt the tool on with no checkpoint at all, which is where the sanctions cases come from.
If your firm hasn't formalized that operating rhythm yet, that's a separate, earlier conversation than which drafting vendor to pick.
Talk to Us Before You Sign a Vendor Contract
If you're evaluating AI drafting software for your firm, sequencing matters more than the vendor name. Schedule a HODOS Fifth Decision consultation and we'll walk through your firm's specific Execution and Witness requirements — court standing orders in your districts, your malpractice carrier's current questionnaire, your existing practice-management integrations — before you sign anything with a drafting vendor. Book the consultation at hodos360.com/consultation and bring your current practice-management setup and your top two vendor candidates; that's the information we need to make the session useful.
- /fifth-decision/execution
- /fifth-decision/witness
- /fifth-decision/cash
- /fifth-decision/people
- /fifth-decision/strategy
- /pillars/ai-legal-drafting-software
- hodos360.com/consultation