Time to Build

Tom Martin opened the conference by situating the moment. He referenced a leaked internal OpenAI memo outlining five levels of AI progress on the path to AGI, and noted how dramatically the pace of that progression had surprised even optimists. When he had given a talk on the same framework in Madrid the prior June, he had assumed certain milestones were years away. Many of them had already arrived.

The through-line he set for the two days: AI isn't something happening to the legal profession — it's something the legal profession has the opportunity, and responsibility, to shape. "We have a choice," he said. The conference was designed to help people exercise that choice more wisely.

Bridget McCormack opened with history. The American Arbitration Association is 100 years old — a non-profit dispute resolution provider that administered more than half a million disputes in just the last two years. That foundation, she explained, is precisely what makes the organization's AI ambitions coherent rather than opportunistic: the AAA/ICDR recently became the first ADR organization in the world to offer AI-native dispute resolution.

McCormack's framing of progress was notable. She argued that organizations navigating AI well are not the ones that move fastest, but the ones that connect what comes next to why they exist. The AAA's founding mission — to make dispute resolution accessible — is the same reason it is building AI into its infrastructure now. Without that tether to purpose, she suggested, new technology tends to amplify existing institutions rather than transform them.

When Sateesh Nori asked how the AAA's culture of incentivizing risk-taking might translate to other institutions, she was direct: institutional intransigence is often a structural problem, not a personal one. Leadership has to deliberately create conditions where trying something and failing is safe.

The panel's framing question resisted the obvious: rather than asking whether AI creates new ethical duties for lawyers, Gartner opened by asking whether existing duties are adequate, under-stressed, or already obsolete. The answer, across all three panelists, was: all of the above, depending on the duty.

Nori named a second misconception: "Lawyers think AI is unethical, that they can't use it. That it's cheating. That it's a shortcut that they shouldn't be taking." Palmer argued that the existing framework does not need a wholesale overhaul — existing duties still apply; what has changed is the how. "We don't need to silo AI into its own category." His concern was that stopping at the surface layer of rules misses the deeper purpose behind them — protecting the client.

On disclosure, Sanders-Reach walked through the growing patchwork of court orders. The Ropes and Gray AI order tracker has filled in so completely that Alaska is reportedly the only state without a court order regarding AI use. The ABA's opinion takes a functional three-part approach: disclosure is required when a client asks; when AI use is relevant to the reasonableness of the fee; or when a lawyer is relying on AI output in making an assessment.

Palmer offered a practical framework: his "four Ds" — data, dollars, decisions, and duties to the tribunal. These are the categories where a client's rights are likely to be impacted, and where the client needs to be part of the conversation. Nori sharpened the internal governance question: the most important line practitioners should draw is between internal use and external use.

On billing, Nori was direct: "If you are charging your client 10 times longer than it takes to do the work because you're using AI, that's unethical... the billable hour cannot last for very long." Sanders-Reach invoked the Westlaw and Lexis transition of the late 1990s — billing back for AI use as a disbursement is unlikely to land well. Palmer pointed to ABA Opinion 512's guidance that unless a tool has a specific, demonstrable client benefit, it belongs in overhead.

On privilege, Palmer highlighted two cases pulling in opposite directions: Heppner (NY) held that sharing information with an AI waived privilege; Warner (Michigan) went the other way — AI-assisted research was held to be protected work product. The doctrinal question remains open: is AI a third-party interlocutor, or just a tool like Westlaw?

On agentic AI, Sanders-Reach expressed measured skepticism: "I am a huge fan of deterministic automation, not so much with the probabilistic, because it is hard to check, and it has unintended consequences." Nori pushed back: "Lawyers tend to jump to the most complex case. But instead, we should be thinking about, like, what are the vast majority of legal cases about?" Palmer, channeling Susskind, drew the distinction that mattered: automation is doing something we already do faster; innovation is creating a whole new path.

Palmer identified deepfakes as the most underappreciated risk: "I think we're just still on the beginning of deepfakes, and how that's going to impact our justice system... it's going to be two-pronged: making fake bad evidence, but also unfairly challenging good evidence by the guilty."

Tom Martin closed by asking who ultimately sets AI limits for the profession. Palmer's answer shifted the frame entirely: "I'll say who it's gonna end up being is the client. I think the tail is gonna wag the dog in many of these respects."

This was among the most practically grounded panels of the conference, carrying a quiet urgency throughout. Gilchrist opened by describing a conceptual shift happening inside legal aid organizations — AI is no longer arriving as a tool to evaluate, it is becoming infrastructure to govern.

On internal use, Zarnow identified three underserved areas where legal aid organizations should be investing: data analysis (organizations are sitting on enormous case datasets they cannot yet efficiently interrogate), communications and impact storytelling, and the updating of internal procedures. Colpoys highlighted language translation as a high-return, lower-risk entry point — one LTF grantee used AI to translate its entire self-help materials library from English to Spanish, though "it always requires human review, especially when legal terms are being used." She described one grantee receiving 7,000 calls per month that now uses AI voice agents to triage who gets through.

Gilchrist described Legal Aid of NC's Justice Hub, a case priority analysis tool that filters incoming cases and surfaces a summary for staff intake workers. "It's not making the decision — but it's sort of helping us to filter and prioritize."

The panel's sharpest tension emerged around self-represented litigants. Colpoys put the scale plainly: "92% of people who are seeking help with a legal problem don't actually get a legal aid lawyer." Gilchrist pushed back against the idea that accuracy is the primary challenge: "It's not just accuracy, it's context. People don't come to us with clean legal questions. They come with stories, and stress, and incomplete information... we also have to design for real conditions, which is low literacy, high emotional stress."

Zarnow was the panel's most insistent voice on testing: "Whatever you're doing, you gotta test the hell out of it. Constantly, repeatedly, before you deploy it, after you deploy it, when the models change." He also named a structural risk: "At the risk of getting kicked out of the AI conference — the goal is not to use AI, the goal is to solve problems... if you're doing an experiment, you've got to be really careful, because in all likelihood, you're experimenting on someone."

Gilchrist's framework for responsible deployment covered four pillars: strong but flexible AI governance; elevating the human role rather than reducing it to oversight; maintaining professional responsibility and consumer protection; and cross-functional collaboration with continuous testing. Colpoys offered the simplest diagnostic: stop asking which AI tool to use. "What you should be asking is, what are the problems you're trying to solve?"

Ron Flagg built on what the AI and Legal Aid panel had laid out, but his framing was sharper and more structural. The justice gap, he said, cannot be closed with money alone — and it cannot be closed with technology alone, either.

Flagg argued that legal aid's most important asset is not the AI tools themselves but the community of innovators building with them. Decades of operating under resource constraints have produced institutional creative capacity not common elsewhere: "If you want to call it this, if there's a silver lining to operating for decades with far fewer resources than you need, it's that you learn to be creative." Legal aid is not just a user of technology — it is a proving ground for what high-stakes, constrained, ethically serious AI deployment actually requires.

On internal operations, Flagg named specific opportunities: reducing the administrative burden on staff; streamlining intake and triage; and using AI to increase organizational capacity at every level. "How do we use AI to increase capacity at every level of the organization?" — the question he posed was operational, not theoretical.

The funding context: Congress appropriated $540 million for LSC in FY2026 — a 3.5% cut from the prior year. The Trump administration's FY2027 budget proposal called for eliminating LSC entirely, though Flagg's read on the likely outcome was a Senate landing closer to flat or modestly increased, as in the previous cycle.

On evictions — LSC's single largest category — Flagg noted they represent approximately 35% of cases closed by LSC grantees. Notably, he pushed back on a simple tenant-versus-landlord frame: many landlords have come to recognize that working with tenants is often a better economic outcome than eviction. "I don't think throwing out your customers is a very good business model."

Flagg closed with a direct charge to builders: develop with the people who will use what you build. "If you're a developer, stay grounded in the real-world needs, and build with users, not in isolation. If you're a legal services advocate, the status quo is turning away half your eligible applicants — we need to lean into experimentation."

Reid opened by asking not just whether AI can make courts more efficient, but whether it can make them more understandable, more accessible, and more just. The three panelists pulled in different directions with equal conviction.

Tashea's first move surprised the audience. Asked what he would build in the next three years, he named cybersecurity — not court navigation tools, not AI judges, not smarter forms. "I see cybersecurity as inherently an access to justice issue. When a court gets shut down because of a ransomware attack... that's an opportunity for justice denied." In a court technology environment with highly fragmented systems and inconsistent standards, security infrastructure is the foundation everything else depends on.

Ebron offered the structural critique the panel needed. Courts are still designed for represented parties: "Courts don't see self-represented litigants as the norm. They're still built for represented people." The compounding barrier she identified is language — legal language is a foreign tongue to most court users, and "the closest native you have is opposing counsel, who isn't necessarily operating in your best interest."

Marz reported from the front line. At the most recent eCourts Conference, roughly 70% of sessions were AI-focused. Priority applications: chatbots for internal staff training and external self-help; AI video tools for training materials; AI transcription to address the court reporter shortage ("a dying industry — in some courts, they literally do not have a record"); and AI-assisted form processing. Alaska's chatbot project expected to take three months took fifteen, and maintenance is ongoing: "Hallucinations still come in... it's really constant, it's not a one and done."

Tashea gave the most skeptical assessment of the conference. His core argument: the justice technology community has done a poor job evaluating its own work for twenty years, and that failure is now built into the AI moment. On whether the access-to-justice gap has closed despite decades of technology investment: "No, it hasn't. The gap keeps... No matter all the projects we've done, all the good intentions, all the good work coming from the right people, the gap gets bigger."

His diagnosis: the case management systems underlying most court technology are bad enough that the data cannot be trusted. The only intervention he could point to with genuine RCT evidence: text message reminders.

Tashea named a specific and growing problem: AI is enabling debt collectors and other high-volume filers to generate more cases — not better ones. He coined the term "litigator psychosis" for a related phenomenon: AI tools being so sycophantic toward self-represented litigants that they generate unjustified confidence in weak claims.

Ebron pushed back firmly: "We have a constitutional right to access to the courts, and we can't stop people from filing what might be frivolous claims." By some estimates, there are nearly 300 million unmet legal needs in the country. Tashea's closing call was for the field to build its evidence base before building more tools: benchmarks for translation accuracy, technology audits for deployed court systems, and genuine impact evaluation. "I need to know which plants are poisonous at this point. Stop planting seeds."

Rishwain set the tone deliberately: he wanted to skip the hallucination discussion and get to practical reality. What are working attorneys actually doing with AI, and what has it changed?

Each panelist described their first real use. Elefant picked up ChatGPT in November 2022 to respond to a pleading: "It was kind of a very aha moment to see what it could do... it took the mental load off me." Jackson started sharing AI experiments on LinkedIn — analyzing Supreme Court briefs, having public conversations with early models: "All of my mistakes and all of my successes are on LinkedIn. I want people to learn from my mistakes." Hernandez brought the technology perspective: she had been building rules-based document automation at Gavel before the generative AI moment, and the key limitation was always the same — "it's not going to auto-update if anything changes in the law."

Jackson gave the most detailed picture of any panelist. His setup: Claude Projects with uploaded briefs and case materials; Perplexity and Claude browser plugins that "see what's on your screen"; a custom Claude skill that automatically responds to emails, logs to Notion, and sets to-dos; a sub-prompt that runs a fact-check on every prior response, catching mistakes "almost every single time." For client communication, he uses NotebookLM for research consolidation and 11 Labs — trained on his own voice in about 30 minutes — to generate audio summaries clients can listen to rather than read.

On billing, the three had distinct approaches. Elefant offers clients a choice — use AI and pay a few hundred dollars, or proceed at full rate. Hernandez described the rise of tiered packages: DIY document automation for a lower flat fee, full-service for more. Jackson is using transparency as a marketing tool: "We're actually pointing out — this research normally would have taken 4.5 hours, but we were able to do it in 15 minutes using our internal AI database, and we're passing the savings along to you. They love that."

On junior associates, the panel produced its sharpest disagreement. Elefant argued that AI can compensate for the mentorship gaps that have always afflicted solo and small firm practice. Jackson's view was expansive: "AI is going to take a lot of jobs from a lot of people, but it's also going to allow those of us that are interested in what we're doing to raise the bar." Hernandez was more cautious: if the mentoring Jackson and Elefant described does not happen, AI will hollow out exactly the deep expertise that took junior lawyers years to build.

This panel was one of the most substantively rich of the conference, and ran long enough that the moderator repeatedly noted time slipping away.

Brice opened with a question about whether all faculty need to engage with AI. Colin Lachance set the tone immediately with a demographic observation: there are graduating 3Ls today who had access to ChatGPT before they sat the LSAT. These students went through three-plus years of law school surrounded, in many cases, by professors who said no to AI from the beginning.

The obligation on faculty, he argued, is not to be champions — but to understand enough to bring even an accent of AI awareness into whatever they teach. He quoted a malpractice insurer: "For every ten things AI can do for you, eight will be amazing and two will get you disbarred." That framing, he said, is sufficient to work with.

Nicole Morris described what she's seeing at Emory: incoming 1Ls who are not just AI-aware but, in some cases, actively building tools. The question of how to scale faculty competence is real. Morris argued bluntly that until performance evaluations and institutional requirements create real stakes, adoption will be uneven: "You have to put some skin in the game for them."

Mark Williams made a subtler point. The productivity gap between faculty members who have adopted AI for research and those who haven't is becoming visible and will only widen. "The floor is just going to get raised, and the only way to catch up will be AI." He also noted that some of the most resistance-prone faculty he'd assumed would never engage were showing unexpected openness — partly because the tools have simply become good enough that the case for ignoring them has collapsed.

The panel turned to students. Williams described admitted students days at Vanderbilt where the volume of AI-related questions has roughly doubled each year. Morris added that her 2Ls and 3Ls this year were telling her she was often the only professor who allowed AI use in class. She said the word "only" with appropriate exasperation.

Lachance called for 2025–26 to be the year of faculty upskilling and 2027 to be the hard deadline for real curricular requirements. Morris pushed further: she has argued that law school should be two years on the business school model. AI makes the case stronger.

Nori offered the sharpest provocation: "What's more important in 2026 and beyond — Contracts, or client counseling?" Morris's answer was careful: the doctrinal core still teaches something AI can't easily replicate, particularly in torts, where students learn to think about policy. "I don't think you can get policy learning from an AI tool. Yet."

Williams closed with a forward-looking note: AI agents are already beginning to contract with each other, creating novel rights and responsibilities that no existing framework cleanly handles. The foundational reasoning taught in a first-year contracts course is going to be more necessary, not less.

Moon opened by recounting two-plus years of "women are falling behind in AI" headlines — and her response: convene people, talk about it, take action. The Women in AI Summit has now happened twice, in 2025 and 2026. The same narrative keeps arriving.

Anastasia Boyko responded with sixteen years of legal tech experience: when ChatGPT launched, she was genuinely excited — finally a tool to reduce inefficiency. Then she started seeing error rates, seeing use cases built around interrelational dynamics rather than anything substantive, and became cautious. She referenced a Substack piece on the "absence of the mother figure" in AI development — the cautious, caring presence that's missing from most adoption frameworks.

Heidi Brown offered the empirical counter to the "falling behind" narrative. In her legal writing and generative AI courses at New York Law School, women went from 23% of enrollment in 2024 to 57%, to 66% in the most recent semester. Every single professional development director at the law firms that contacted her about AI programming has been a woman. The framing issue, she argued, is that the metrics being used count minutes-of-use, not quality-of-use.

Sateesh Nori opened Day 2 with a personal speech that set the emotional register for everything that followed. He spent years as a legal aid lawyer in New York City, representing tenants in Brooklyn Housing Court facing eviction. He described sitting across from people whose entire lives — their children, their schoolwork, their dignity — hinged on whether someone in a robe knew their name. Many of them came to court alone, holding paperwork they didn't understand, in a language that wasn't theirs, against a landlord with an attorney handling two hundred cases a week.

"There are not enough of us," he said. "There will never be enough of us, not at the rate we're going."

Every reform of the last fifty years — more legal aid funding, more pro bono, plain language forms, self-help centers, court navigators — stacked on top of each other, and the gap hasn't closed. He called it not a gap, but a chasm. A canyon. The question for Day 2 was not whether AI could help. It was whether the legal community would choose to use it for that purpose.

Nicola Shaver opened Day 2 from Munich — she had been in London the day before, a detail she mentioned as a point of emphasis: "This is a global phenomenon." Her perspective is built on that breadth: a lawyer by training who practiced for a decade before moving into legal innovation, she led global innovation at Paul Hastings before founding Legal Tech Hub (now Your Legal Intelligence), a subscription intelligence platform tracking the global legal technology market in real time.

Her keynote moved through four interconnected arguments: the market has changed more dramatically than most practitioners realize; there are genuine signs of structural fabric change in the legal industry for the first time in generations; the Anthropic "SaaSpocalypse" moment has further destabilized an already unsettled landscape; and law firms that do not look beyond efficiency gains to genuine differentiation are in the dilemma zone.

The Market Has Moved Faster Than Anyone Predicted

Shaver's firm has been tracking generative AI solutions in the legal market since early 2023. The first map, published in early 2025, astonished the industry. One year later, the map had more than doubled. "If you look at startup activity as a signal in and of itself, it is not suggesting we are anywhere close to seeing a bubble bursting or a slowdown of the hype or the growth." Funding numbers: $4.28 billion raised in 2025 across the legal tech market; $1.42 billion already raised in the first quarter of 2026 alone — 33% of the prior year's total in a single quarter.

On agentic AI specifically: her firm's first Agentic AI map, published in July 2025, showed roughly 100 solutions. By March 2026, that number had exploded. "Already, by now, in April 2026, I can say with certainty that we're finding a lot of use cases for agentic AI in legal, and generally speaking, the dialogue around AI and legal has shifted to agent." In-house adoption went from 23% to 52% in a single year — one of the most consequential trends in the market.

Signs of Fabric Change

Shaver's most consequential argument was that the market data point to structural changes in the legal business model that have never before occurred in the industry's history. Four signs of fabric change: (1) AI-native law firms — approximately 20 emerged in 2025 alone, built on AI as their operational foundation; (2) accelerating firm mergers — a 21% jump in law firm merger activity in H1 2025; (3) private equity investment, through managed services organizations that insulate PE investors from non-lawyer ownership restrictions; (4) the client relationship inverting — clients who in 2023 were prohibiting AI use are now explicitly demanding it, moving from 63% encouraging to 78% encouraging between 2023 and 2024.

She invoked Clayton Christensen's The Innovator's Dilemma directly: incumbents focus innovation energy on enhancing what already makes them successful; nimble disruptors enter from the margins, appear unthreatening, and then grow fast enough to overtake. "I think we are there. We're in that dilemma zone."

The SaaSpocalypse

When Anthropic launched its Claude Legal Skills and legal plugins — intended simply to demonstrate that Claude was a strong building block for legal workflows — the market reaction was immediate and disproportionate. Stocks of publicly traded legal tech companies collapsed. What followed was a wave of vibe coding enthusiasm and Anthropic rapidly partnering with Harvey, iManage, and NetDocuments through MCP connectors. Shaver was clear-eyed about both the excitement and the risk. "It's not a slam dunk." But the episode was significant as a signal that the market is not settled.

The New Horizons Model

Shaver closed with a framework adapted from McKinsey's Three Horizons, compressed dramatically in timescale. She invoked Jack Welch: "When the rate of change on the outside exceeds the rate of change on the inside, the end is near."

Her three horizons: Horizon 1 is time-and-cost efficiency — table stakes as of last year. Horizon 2 is better output and augmented judgment: higher quality work, human-machine collaboration producing better results for clients, not just faster ones. Horizon 3 is fully transformed systems — new service models, new client relationships, legal expertise abstracted into repeatable scalable products.

The critical observation was structural: most law firms cannot achieve Horizon 3 plays because the partnership model prevents it. Consensus-based governance kills transformative initiatives. "If you do not change these things structurally at your firms — or move outside of the firm into a subsidiary structure that allows for a completely different setup of work — any innovation play that you make at that Horizon 3 level will be killed."

In March 2026, Artificial Lawyer launched the first directory of AI-native law firms, with 27 firms listed at launch. The directory's existence signals something more significant than a marketing category — it marks the moment when "AI-native" became a recognized organizational type rather than an aspiration or a marketing claim.

An AI-native law firm is not a firm that has deployed Harvey or Legora or any other tool. It is a firm where AI is the operational foundation — where the workflows, staffing model, pricing structure, and client delivery mechanism were designed around AI from inception, not retrofitted to accommodate it.

Several distinct models have emerged. The lateral-departure model: senior lawyers leaving established firms to build new practices on AI-native platforms like Legora, Clio, or EVE from day one. The product-to-services model: legal technology companies expanding into legal services. Udia launched Udia Council; Norm AI (backed by Blackstone) launched Norm Law as "the first AI-native full-service law firm for global institutional clients." In the UK, Orbital — a real estate technology platform — launched a fully automated conveyancing law firm where the end-to-end workflow runs without human drafting. Manifest raised $60 million at a $750 million valuation; Carta acquired AI-native funds law firm Avantia.

The regulatory environment is beginning to respond. In May 2025, the UK's Solicitors Regulation Authority approved Garfield.Law Ltd as the first law firm authorized to deliver legal services entirely through AI technology — a precedent with significant implications for jurisdictions still wrestling with non-lawyer ownership rules.

The structural implications are serious. The traditional law firm pyramid is collapsing into what analysts call an "obelisk" model: fewer human layers, more direct leverage from AI systems, and pricing that naturally gravitates toward fixed fees over billable hours. Y Combinator's 2025 Request for Startups challenged founders to "start your own law firm, staff it with AI agents, and compete with existing law firms." Some founders took that literally.

Damien Riehl, who leads solutions work at Clio across a billion-document dataset, opened with Hemingway's bankruptcy line: slowly, then quickly. That is where legal practice is. On the corporate client side, he described speaking with a Fortune 20 company that had already built 55 different AI workflows to reduce external legal spend, was building 56 more, and planned to have approximately 160 workflows operational by end of Q2. That is the speed at which major clients are moving.

Megan Ma, Executive Director of Stanford Law's LIFT Lab, identified the biggest impact she's seeing as cultural rather than technical: for the first time, attorneys are genuinely being touched by technology. What's still missing is figuring out the right model for human-machine collaboration. Most current tools are hyper-user-driven — single-direction, reductive, extracting summaries from documents. The emerging paradigm, embodied in tools like Claude Code and agentic systems, is fundamentally different: collaborative, constructive, expansive rather than reductive. "We're just slowly transitioning into that space."

Joanne Sprague brought an eDiscovery lens. Document review is one of the most AI-ripe workflows in legal. Everlaw's new "Deep Dive" tool — essentially a ChatGPT for a closed document corpus — is already changing what's possible in investigations and litigation. Entire categories of legal work are not just being made faster; they are being structurally redefined.

On May 12, 2026 — two weeks after the LawDroid conference — Anthropic made its most decisive move into legal to date, formally launching Claude for Legal. The announcement was not a product launch in the traditional sense; it was a structural repositioning of the most widely used underlying AI model in legal technology into an active participant at the front of the stack.

The offering shipped on four tracks simultaneously: twelve new practice-area plugins covering commercial, employment, privacy, product, corporate, and AI governance work; more than twenty new MCP connectors linking Claude to DocuSign, Ironclad, iManage, NetDocuments, LexisNexis, Thomson Reuters, Box, Everlaw, and others; a cross-application integration with Microsoft 365 embedding Claude across Word, Outlook, Excel, and PowerPoint as a single context-carrying agent; and open-source ecosystem participation from Harvey, Legora, and others.

The numbers behind the launch: legal had become the number one power-user job function across Claude's Cowork platform, with over three times the usage of any other professional function. When Anthropic held a How Legal Teams Put Claude to Work webinar in April, more than 20,000 legal professionals registered. At Freshfields, Claude usage grew approximately 500% within the first six weeks of firm-wide deployment. Claude Opus 4.7 scored 90.9% on Harvey's BigLaw Bench — the highest score any Claude model had achieved.

The announcement drew candid responses from incumbents. Harvey CEO Winston Weinberg was blunt: "Gabe and I have said for years that long term we would end up competing with the model companies. This is validation of both our initial strategy and how we believed the landscape would change." Thomson Reuters CTO Joel Hron reframed the dynamic: "The control point isn't where work starts. It's whether the output is accurate, grounded in authoritative sources, and defensible." Artificial Lawyer's Richard Tromans articulated the structural shift: "Claude becomes the legal AI fabric, upon which the other legal tech participants embroider their additional workflows and their curated data."

Mark Pike illustrated the access-to-justice implications directly: a paralegal on a four-person pro bono team, opposing an AmLaw 200 firm in an elder abuse trial, built a tool on the Claude API that sat at counsel's table pulling in live cross-examination lines as witnesses testified. The small team walked out with a large jury verdict. "This shows what Claude is making possible — giving a small team capabilities that would typically require more funding and a much larger team."

Westermann, an assistant professor of law and AI at Maastricht's CyberJust Lab, built this workshop on the foundation of the prior year's prompt engineering session — and the upgrade was significant. "A lot has happened since then — so I'm very happy to be talking about skills today."

His setup: LLMs are general models, but legal work is specific. Every firm, every practice area, every client type has particular knowledge, standards, and judgment that cannot live in a general model. Skills are the mechanism for closing that gap. He walked through how the field has evolved — predictive models → LLMs → multimodal LLMs → reasoning models → agentic AI — and grounded the room in what's happening now: models can spawn sub-agents, operate with external tools, and execute tasks across long sessions.

Skills, he explained, are something between prompts and plugins — specific guides containing expertise that help models figure out how to do particular tasks. They are reusable, shareable, composable, and editable. The model does not need to be told to use a skill; it identifies the appropriate one from context. The Matrix analogy he offered was well-received: "Skills are kind of like... in The Matrix, Neo gets given this diskette, it loads onto his brain, and suddenly he has these new skills." Structurally, a skill is a .skill file (a zip archive) containing a SKILL.md file — the name, trigger description, and instructions — plus optional supplementary files like templates, examples, or legal source documents.

The live demo illustrated all three phases. Westermann dragged a sample Quebec residential lease into Claude and asked for the five biggest legal risks. Claude identified three real violations: an illegal security deposit clause, an illegal repair clause, and an illegal immediate termination clause, all running contrary to the Quebec Civil Code. He then asked Claude to edit the Word document with tracked changes and comments, and draft a client reply. The final step: he asked Claude to package the entire workflow into a reusable skill. Claude built a "LegalDocReview" skill from the workflow it had just demonstrated.

On the difference between skills and prompt libraries: skills eliminate the copy-paste problem and the self-selection problem. "The models can use them themselves — you don't need to provide them every time. The model will figure out for itself when it needs a skill and when to use it... You could have tens or more than that skills just sitting there, and the model will go in there and look for what it needs." On cross-model compatibility: skills are becoming a standard — ChatGPT and Gemini have both begun integrating them.

When Sateesh Nori asked whether this leads to the "one person at the top of the pyramid with a bunch of agents with skills" law firm: Westermann acknowledged that some practitioners are already deploying AI agents on computers with access to all files and all email, with instructions to "be the lawyer." He was measured: "There's a lot of risk for that, of course, so not something I'd necessarily recommend doing today. But on the other hand, it's happening." His closing call: identify the tasks you do often, the workflows that repeat, the knowledge you bring to recurring categories of work. "Once you've realized that, you can package this in a skill."

Conor Malloy runs a legal aid hotline in Chicago, and his workshop was a practical manifesto for non-coders who have ideas. Vibe coding — using AI to build software without a traditional CS background — has moved from curiosity to genuine capability. Malloy walked the audience through building a functioning lawyer website in near-real time, using AI as a collaborator throughout. The message: the barrier between having an idea and manifesting it as working software has collapsed. Legal aid organizations, small firms, and independent practitioners no longer need a vendor relationship or a dev team to prototype a tool.

By 2026, 83% of lawyers have "broad AI access," up from 61% in 2025, and the tools enabling non-programmer development had become accessible enough that the barrier is motivation, not skill. Solo practitioners are building their own contract and matter management systems. In-house legal ops teams are open-sourcing outside counsel management tools on GitHub. A Quinn Emanuel partner built his firm's litigation AI platform on Claude with "virtually no coding background," treating Claude like a case team member — onboarding it with chronology, key excerpts, and themes the way you'd onboard a partner joining mid-case.

The governance concerns are real and growing. Research published in December 2025 found that AI co-authored code contained 1.7 times more major issues than human-written code, with logic errors 75% more common and security vulnerabilities 2.74 times higher. More immediately, lawyers are building tools outside firm infrastructure — hosted on personal AWS accounts, operating beyond IT visibility — with firms often unaware they exist until something breaks.

The Legal Quant is something distinct from the vibe coder, though the communities overlap. The concept was articulated by Jamie Tso, a funds lawyer formerly at Clifford Chance in Hong Kong, in a January 2026 Substack post titled "The Jane Street of Law: The Rise of the Legal Quant."

Tso's thesis: law is bifurcating between lawyers who use AI at the margin to do existing work faster, and a smaller group — the Legal Quants — who treat AI as infrastructure. The Legal Quant is comfortable in a terminal, thinks in terms of prompts plus retrieval plus evaluation, has opinions on RAG architecture and failure modes, can build tools, and can collaborate tightly with engineers. "It's about control — the difference between merely using a tool and expanding the frontier of what's possible."

Tso and Raymond Sun, a Sydney-based tech lawyer, co-founded LegalQuants.com — an invitation-only community for lawyers who have built tools for the industry. The Jane Street analogy is deliberate: Jane Street identifies technical talent early and pays for it accordingly. BigLaw rarely does. The shared frustration among the Legal Quants Tso catalogued is that the skill set is invisible to traditional compensation structures — which he argues creates a genuine arbitrage opportunity for whoever moves first to underwrite it.

The Lawcadia 2026 trends report framed the profession-level version of this split directly: legal teams will bifurcate into judgment experts who apply legal strategy and bear professional responsibility, and hybrid operators who build, govern, and optimize legal technology systems.

The most unexpected legal AI story of early 2026 did not come from a well-capitalized startup or a BigLaw innovation initiative. It came from an Austrian developer named Peter Steinberger, who published a self-hosted AI agent framework in November 2025 under the name Clawdbot. After a trademark dispute (the name "Moltbot" lasted three days), the project was renamed OpenClaw on January 30, 2026.

By March 2026, OpenClaw had surpassed Linux on GitHub's all-time star leaderboard with 247,000 stars and 47,700 forks — the most-starred non-aggregator software project in GitHub's history. OpenClaw is not primarily a legal product. It is an open-source AI agent framework that runs on any infrastructure, including entirely self-hosted deployments — and lawyers found it.

The appeal is specific. OpenClaw runs on your own server. Client communications and matter details never leave your deployment environment. There is no third-party data processing agreement to negotiate, no vendor privacy policy to audit, no bar ethics opinion to obtain about uploading client files to cloud infrastructure. At roughly $45 per month in infrastructure costs, it automates client intake, deadline tracking, billing reminders, document drafting, and email — an AI paralegal available around the clock, without the data sovereignty concerns that have slowed enterprise AI adoption in legal.

Bloomberg Law identified the OpenClaw phenomenon as raising novel questions about AI agents acting in fiduciary capacities — specifically whether an agent with sufficient autonomy over client matter management could be acting as a trustee. The Steptoe law firm published analysis noting that agentic AI of the OpenClaw variety represents "a glimpse of a future that has arrived before relevant guardrails and standards are firmly established."

Kezia Hill, Legal Content Manager at Bay Area Legal Services and former family law trial attorney, built knowledge bases for legal AI systems — and her workshop may have been the most practically useful session of the conference for anyone building legal information tools for non-lawyers.

Hill's thesis: most legal AI assistants fail not because the technology is wrong but because the knowledge bases are poor. "A universal problem that we are facing with our legal AI assistants is that a lot of them fail because they're trained on using knowledge bases that are just purely structured on information. There's no care or attention paid to the actual reasoning or the thinking behind — okay, why is it structured this way instead of another way?"

She described the common failure modes: FAQ-based systems, static content no one updates, no decision logic, no procedural intelligence. The AI answers without knowing where the client is in a process — and the result is "vague, incomplete, or just wrong."

Hill also addressed the fear of hallucinations in closed systems: when a legal AI assistant is built on a closed knowledge base, it cannot reach outside it to invent case citations. "It's only using the information that you put in." This distinction — closed system versus open web — is foundational to how legal organizations should be thinking about deployment risk.

Her five-step methodology: (1) Understand the User — customer discovery, listening to client intake calls, analyzing Google Analytics (found 89% of users were seeking landlord-tenant law, not the mix of issues staff assumed); (2) Deep Research — "revisit these subjects as if you are learning them the first time," primary source research, identifying gaps practitioners take for granted; (3) Structure the Knowledge — mind mapping, procedural timelines, test by explaining it to a non-lawyer ("can I explain this to my sister, who works in healthcare?"); (4) Translate to AI Logic — decision trees, conditional responses, modular content blocks, action-oriented outputs, terminology consistency throughout; (5) Ontology — structured metadata at the end of each document: issue type, applicable legal rules, conditions and exceptions, required evidence, user action steps, procedural posture. "This allows the system to recognize what phase of a matter the client is in."

The product built with this methodology is Bailey B — Bay Area Legal Services' closed AI chatbot for housing and landlord-tenant law. On whether context engineering raises UPL concerns: "There are no inferences being made. There's no reasoning that it's doing. It is simply translating information that is already made available." Hill was also clear: "AI could never completely replace what we do as attorneys... there's this human component that AI does not have."

Steenhuis opened with a provocation: in the age of generative AI, document automation is not obsolete. It is more relevant than ever, and agentic AI has now made it possible to build these tools dramatically faster than before.

Document automation is, at its core, TurboTax for legal forms. Think guided interviews that produce completed documents — eliminating search-and-replace errors, automating condition logic, and integrating with filing systems. Steenhuis builds on Docassemble, a free, open-source platform. The three primary use cases: internal workflow efficiency for recurring documents, intake and triage guided interviews, and do-it-yourself tools for family law, estate planning, and corporate formation that can reach people who will never interact with an attorney.

His argument for code-based over no-code was counterintuitive: "You can use AI to write code much easier than you can use AI to click buttons in an interface... In some ways, that kind of intermediate, helpful step of a really nice graphical interface gets in the way of using AI."

The live demo built toward "Weaver" — the lab's new agentic AI tool for constructing Docassemble guided interviews from scratch. The demo form was California's Delegation of Parental Authority — a caregiver's authorization affidavit urgently needed by families facing immigration enforcement. Target time: thirty minutes from blank to working draft. The demo hit a bug — alpha software, Steenhuis noted without embarrassment.

The 50-state Delegation of Parental Rights project was explicit about its humanitarian rationale: "Imagine you suddenly need to leave the country. For some reason, you might be detained or at risk of deportation. What can we do to make sure you have a way to take care of your kids?" The lab's Court Forms Online platform has been visited more than 37,000 times in the past year, with Michigan partner sites handling 100 simultaneous users per hour.

The distinction from vibe coding was pointed: "When we're doing a document automation solution that's gonna be facing the world... we need guardrails. The skeleton for the workflow, the rules for creating the workflow, and a checklist to review the work — all of those parts are really important when we're doing any kind of work with AI."

Robinson described Upsolve as TurboTax for bankruptcy — helping low-income Americans file Chapter 7 pro se, with no attorney and no cost. To date: 23,000 families served and more than one billion dollars in debt eliminated. In early 2025, the team was asking a harder question: why do users drop off?

The assumption was that drop-off was a knowledge problem — users encountering legal concepts they didn't understand and giving up. The data said otherwise. "Most of the drop-off was a design problem, rather than just a knowledge transfer problem. Most of the drop-off was happening in legally simple, but very complex design sections."

This reframing led Upsolve to build not a legal knowledge assistant but a navigation companion. The AI side panel is embedded directly in the filing flow. Three design decisions distinguish it: the AI knows which page the user is currently on; it has access to the UI elements visible on that screen and any error messages; and the prompt-writing is based on research into what actually gets users stuck on each specific page.

Robinson offered a deceptively simple example: a user asks "why can't I move forward?" and the last name field isn't filled in. The AI gives the concrete answer that unblocks them. "Obviously with this example it's fairly straightforward, but on a complex page, this is really, really exciting for the user."

The results: approximately 40% of Upsolve users regularly engage with the AI tool, and monthly filings have increased fourfold since launch — without a proportional increase in human staff.

Robinson's three design principles for anyone building AI tools for vulnerable populations: build for action, not just understanding; reduce friction wherever it exists, based on actual user behavior rather than assumptions; and design AI as a companion that lives inside the product, not a separate resource the user has to go find.

Sartor began as a law school volunteer advocate for teenagers in foster care, fielding questions that should have had easy answers: "Do I have the right to visit my siblings? What happens if I'm prescribed psychotropic medications that I don't want to take?" She searched for a comprehensive guide to foster care rights in Florida. Found nothing. Nationally — found a little, none of it good.

Her solution was labor-intensive: a team of 15 law students, 300 hours of pro bono research, and a 66-page Know Your Rights Guide. Youth loved the content. They lost the booklets almost as fast as they received them — which, for teenagers in foster care who often lack stable housing and any reliable way to hold onto possessions, was not a design failure. It was a design specification.

The consistent feedback from youth: get it on our phones. So Sartor built Foster Power. The co-design process was the feature, not the constraint. Youth were involved at every step — name, logo, color scheme, line-by-line content review — and were paid for their time and expertise. The name was chosen by youth, and one young person's reasoning stays with the project: "People like to act like we don't have any power, but we do."

The critical design decision: most youth in foster care do not have continuous cell phone service, and group homes frequently shut off Wi-Fi at night. "Knowing this, we made the app completely accessible without Wi-Fi access or cell phone service." The app works offline. In a system that routinely withholds connectivity as a control mechanism, that is not a convenience feature. It is a rights protection.

Foster Power launched in May 2023 as the first app and companion website in the country dedicated to explaining the legal rights, benefits, and protections of youth in foster care. Results: 5,800 app downloads, 58,000 website views, 220,000 video views across social media. Features on NPR and CNN. Interest from 15 states. National expansion planned beginning in 2027.

Sartor's design principles, distilled for practitioners building for vulnerable populations: center the people you aim to help, and know their constraints well before you build; value their feedback and compensate them for their expertise; and do not be deterred by people who tell you it cannot be done. "The reality is, it just hasn't been done yet."

Lauritsen used his ten minutes to do something unusual at an AI conference: he refused to pretend the current moment is unprecedented.

He opened in 1976. He was a summer associate at a Manhattan law firm, recording time on paper timesheets in triplicate. No email, no web, no desktop computers. The cutting-edge legal technology of the day was the Lexis UBIC terminal — a dedicated connection to a mainframe in Ohio. If you knew Boolean connectors and proximity operators, you were "kind of like the prompt engineering wizards of the day." He went on to join Harvard's Project Pericles in the mid-1980s — working with a VAX minicomputer to build document automation systems, expert systems, and early computer-aided instruction tools. He attended the first International Conference on AI and Law at Northeastern in 1987.

Then he read the room a quote. It was about feeling overwhelmed by the volume of commentary on AI and law — the breathless publication pieces, the stacks of material, the daily webinars. The audience recognized it. He told them he wrote it in 1994. "Some things really haven't changed very much yet."

The irony was the point. The sense of standing at the hinge of history in this field has been recurring for fifty years. Some of those who felt it were right. All of them were working with incomplete information about where their moment would actually lead.

On 2026: Lauritsen named Claude Cowork as the poster child for the current moment. He described the explosion of "incredibly highly valued ventures that are wrapping around some of these foundational technologies" as "wrappers on steroids."

His 2076 projection came through three parallel long-term trends. Hardware: mainframe to minicomputer to personal computer to smartphone — to what? Software: procedural coding to rules engines to LLMs to vibe coding — to what? Lawyers: ignorance and avoidance to reluctant use to enthusiastic adoption to over-dependency — to what? He offered a word: "enslavement might be on the horizon."

"If you take all these three trends and you triangulate them, it's pretty easy to come up with a vision of what's going to be like in 2076. But unfortunately, I'm out of time, so I'll have to save that for next year." The humor was earned. But beneath it was a genuine provocation: the practitioners in this room are in the middle of a long arc, not at its end.

Patrick Pallis spoke about the rule of law. Not AI specifically — the rule of law, its current fragility, and the obligation of lawyers to hold it. Tom Martin described this as a speech that felt more necessary than expected: the legal profession's commitment to first principles — who we serve, what we stand for — is not a given. It needs to be actively defended, and that defense is more than technical.

Tom Martin and Sateesh Nori closed the conference together with a presentation on AI agents — where they are, how they work, and what they mean for legal practice.

Martin framed the progression in three steps: traditional AI (ChatGPT-style sequential chat), workflows (predefined execution paths, like custom GPTs or Gemini Gems), and agents (where the model can take autonomous action using tools, memory, and skills). Workflows are predictable and scalable; agents are more powerful but probabilistically unpredictable. For legal work, where the cost of error can be high and the documentation of process matters, the choice between them should be deliberate.

Nori offered a conceptual reframe that clarified the whole arc: most AI use to date has been reductive — summarize this, distill that, extract the holding. The shift to agents and skills is the move to constructive use. "You reduce it, and then you build with that reduction — new things. The entire cooking process is managed autonomously. That's the unlock."

The agent architecture walkthrough covered memory, knowledge bases, tools, and sub-agents — smaller, focused agents spawned by an orchestrating agent to handle discrete tasks in parallel, reducing cost and improving accuracy by keeping context windows focused. Martin showed a METR chart tracking AI agent attention span and noted that Claude 4.6 can now operate with coherence for up to twelve hours — a figure that was measured in minutes just a year ago.

The demo portion introduced two LawDroid products. Ally is an AI-native legal platform combining case management with AI: deadlines, document analysis, conflict checking, time and billing, and a parallel task board that simultaneously drafts a demand letter, researches opposing counsel, and prepares a court answer. The second demo introduced an AI workforce framework: a managing agent (Casey) with sub-agents handling podcast management, social media, action items, and research briefings. Monthly cost per agent: roughly $25.

In early 2026, every major legal technology vendor shipped some version of agentic AI. The transition was fast enough that the industry barely had time to define its terms before deployment outpaced the definitions.

The practical distinction that matters: AI assistants respond. AI agents plan, reason, and execute. An agent receives a task, decomposes it into subtasks, executes them sequentially or in parallel, evaluates its own output, and returns a result. Contract review agents read the whole agreement, check defined terms against their uses, cross-reference schedules and exhibits, compare positions against a playbook, and produce marked-up output with deviation summaries — without a human prompting each step.

The shift from task to workflow is more than a capability upgrade — it changes the economics of legal work fundamentally. AI assistants compressed the cost of individual tasks. Agents compress the cost of entire workflows.

The most vivid illustration appeared on May 20, 2026, when Finnish legal tech expert Antti Innanen open-sourced Lavern — a full multi-agent legal system he built over six months, alone, running on a Mac Mini. Lavern deploys 67 specialist AI agents across eight end-to-end legal workflows. A Partner agent handles client intake, decomposes the matter, routes it to specialists, runs three-layer internal verification, and halts at human decision gates before anything consequential proceeds. Total cost: negligible. Nothing leaves the local machine. License: Apache 2.0.

Innanen's explanation for open-sourcing: the prevailing "AI as junior associate" framing felt like the wrong analogy to start from. A system designed as a system from inception behaves differently than a human workflow with AI inserted.