Priori was founded in 2013 on a single question about unbundling, and generative AI has finally made it operative. Two distinct unbundlings are now underway. The work is coming apart into an execution layer and a judgment layer, driven by technology and carrying a feeling of inevitability. Then the institutions built around bundled work come apart and reconstitute into something new. Keeping the two separate is what makes the rest of the industry conversation tractable.

Every demand indicator is up - 97% of general counsel report rising work volume, alongside rises in pro se cases, cases filed, global M&A and law school applicants. At the same time, Anthropic's labor market research estimates AI could theoretically perform 89% of legal tasks, which is not the same as 89% of jobs. These look contradictory but point to the same shift: when the cost of execution falls, demand does not disappear, it grows. Work that could never be justified before can now be done, and AI generates legal demand of its own through governance, verification and regulatory disputes, while also raising the speed of business that legal has to support.

The Jevons paradox. When steam engines got more efficient, Britain did not burn less coal, it burned more.

Legal has always sold execution and judgment as a single product priced by the hour, and the pyramid was how a client bought both at once. AI is pulling them apart. On the execution side - search, extraction, document review, summaries, first drafts - estimates already show 40% to 90% reductions in time for review and diligence, and first-year classes at law firms are starting to shrink. On the judgment side - deal structure, risk calls, difficult interpersonal situations, negotiation, strategy - rates rose 7.3% last year, top partners are reaching $4,000 an hour, AI-skilled lawyers command roughly a 56% salary premium, and postings for AI specialization are up 182%.

Rates, not volume, have carried the pyramid forward - AmLaw 100 revenue has nearly doubled since 2008 on basically flat demand growth. Despite near-universal dissatisfaction, alternative fees still represent barely 6% of outside counsel spend, while 61% of chief legal officers are pushing to change how outside services are priced and 80% say they plan to reallocate work away from traditional providers. The twist is that the pressure may not come from the buyer at all. With fewer hours underneath the pyramid, a firm cannot bill its way to growth, and has to price what is left, which is judgment.

64% of in-house legal teams plan to insource more work on flat headcount, because they can now absorb the execution layer. That makes the in-house function the router of work - the place where decisions get made about what goes where. It also creates the pressure point: the team is being asked to absorb more work without growing, and while demand fluctuates, headcount does not fluctuate as easily. Scaling judgment up and down in step with demand stops being a staffing nicety and becomes a design requirement of the department.

The data cuts directly against treating flexible talent as a commodity bench. 83% of flexible lawyers say demand for their skill set meets or exceeds available supply, only 15% saw any pushback on their rate in the last 12 months, and half never negotiate at all. Looking forward, 41% expect compensation to rise against 9% who expect it to fall. Asked what most strengthens pricing power, the answer was expertise - practice area, industry, or a prior in-house seat - which together account for 88% of pricing. A typical flex talent lawyer grosses around $300,000 and takes home around $270,000. Critically, the variable is hours, not rates, which makes closing the gap between engagements an infrastructure problem the market is actively building.

The first unbundling, the work, is happening whether the profession participates or not. The second, the institutions, is genuinely open. Rubin's warning is that declining to design it is not neutral - it hands the design to litigation, regulation, and the AI companies and their financiers, who are content to let the value of legal work accrue to them. Department leaders, legal ops professionals, flexible lawyers and builders all have a hand in that architecture, and pricing runs through all of it, because when execution and judgment reprice in opposite directions, how you price is the strategy.

Asked to predict five years out, all three leaders said headcount stays broadly flat, but the mix changes substantially. Mark Smolik's team of 91 will ebb and flow with a fast-growing business while adding more data-driven roles; DHL hired its first data scientist three years ago. Rishi Varma expects a shift away from the traditional lawyer and legal professional mix toward data scientists, product builders and AI agents, with Cargill hiring its first legal ops data scientist this year. Mike Haven pushes back on the premise itself: resource needs are unknowable without knowing business demand, and if demand keeps growing, technology absorbs the increment rather than shrinking the team. The through-line is fewer jobs organized around manually moving information and more organized around designing systems, supervising outputs, and translating legal risk into operational logic - including a new class of talent that feels half lawyer, half technologist.

Smolik's AI journey began long before anyone was talking about AI. Joining DHL with a five-person team, Smolik asked where the company's information lived and found most of it sat with external counsel - which meant paying outside counsel not just to handle matters but to report on them. The team built its own library, initially paper, then a document management system, and spent eighteen months deciding how information should be segregated across M&A, corporate and secretarial work. In retrospect they were tagging and structuring data without knowing that was what they were doing. As the database matured across the Western Hemisphere, layering AI tools on top was straightforward, and two of DHL's AI initiatives have now been running a full two years and paying real dividends. Smolik's warning is that plenty of departments are buying tools under peer and board pressure and then discovering they have no knowledge bank to point them at.

Haven identifies a significant gap between what agents can already do and what people are actually doing with them, and breaks the causes down: data foundations, governance concerns, choosing the right tool from a crowded stack, the shortage of legal engineers with capacity to build, and finally self-preservation - real fear that AI will absorb the tasks that constitute someone's job. Smolik adds a market-structure cause: technologists, private equity and venture capital identified a sleepy $1.4 trillion global legal industry running on a business model from the 1980s and 1990s, and built tools faster than buyers could evaluate them. The result is that supply has outrun demand-side literacy, and cautious buyers wait, aware that anything they spend six or seven figures on may be superseded before IT finishes evaluating it.

Imagine walking into a Home Depot technology store for legal. The shelves are full, you are the head of legal operations, and you have no idea what to begin to buy.

Smolik has watched firms extol AI initiatives with no visible evidence behind them, and the test applied is arithmetic: DHL now has cost data on every matter, and those costs are not falling. Commodity work that took ten hours two years ago can take an hour today, yet the bill has not moved - suggesting firms are capturing the efficiency as margin rather than passing it through. Haven's test is behavioral: firms making genuine progress are eager to show it, inviting clients in to brainstorm about solutions they have built, while firms that are dodgy about their AI use tell you what you need to know. Varma sees firms responding differently again, marketing their judgment, regulator relationships and a partner's thirty-plus years of experience as the thing a repeatable platform cannot replace. All three confirm this now factors into outside counsel selection, and increasingly so.

Varma's goal at Cargill was not document search, it was building something that could learn how the company actually tolerates risk - which varies by product, region, and whether the counterparty is a global giant or a small regional customer. Incoming lawyers cannot know that and normally absorb it from colleagues over years. The proof case: a customer's general counsel called about an agreement that had been stuck for twelve months, and Cargill's own team explained the customer was asking them to accept risk the company never accepts. Interrogating the contract intelligence tool showed that across a 161-year history Cargill had in fact accepted comparable risk in comparable circumstances; those particular lawyers had simply never seen it. Once the team confirmed its operations, controls and governance could support that level of risk, it negotiated a deal. Varma is also deliberate about scope, working in bite-sized chunks - litigation profile, M&A profile - rather than trying to ingest everything the company has ever done, and Varma flags that the real operational detail often lives in purchase orders, invoices and statements of work rather than the thirty-page master agreement.

All three departments delegate real routing autonomy, but each has engineered the default. DHL centralized everything when the team was small and external spend was out of control; now senior leaders route within a vetted panel that gets annual performance reviews, most work arrives through an intake tool built over ten years, and only genuinely novel matters requiring a new firm need Smolik's sign-off. Varma pressure-tests the reflex directly - the instinct to go outside because a question is new or complex - and incentivized it out of the direct reports by requiring each of them to find their own "AI dividend" as a quarterly objective. Varma deliberately did not tell them where to look. Cargill's IP team went furthest, layering AI on their patent database to file office actions faster, draft patents and review inventions in the queue, eliminating their need for outside counsel altogether and creating competitive pressure across other teams. Haven instills a sequence: can we handle this internally, and if not, what risk and complexity level is this, and what type of firm actually matches it.

Haven tracks a metric called timekeepers per full-time employee - how many distinct timekeepers a firm burned to deliver the equivalent of one person's year. A high ratio signals inefficiency, because every additional timekeeper has to spin up on the business and the matter. The same anxiety is what makes teams hesitant about new firms and flexible talent. If AI can compress that ramp-up, the metric stops mattering as much and the reluctance goes with it. Smolik describes the foundation that makes this work in practice at DHL, where five flexible professionals handle commercial customer contracts against a playbook of board-approved commercial expectations and an alternative clause library, with AI moving that from something a person has to remember to something a new contract is automatically vetted against. As Smolik frames it, capturing a brain is really just capturing how you do what you do and the four corners within which you operate.

Legal Edge is a technology-driven center of expertise inside the AT&T legal department, dedicated to delivering legal services using advanced technologies and generative AI in partnership with departmental colleagues, the business, and trusted outside counsel. The starting scope was chosen on a clear logic: litigation, investigations and regulatory matters are where an in-house team sits closest to the clients and closest to the documents, which means it can conduct factual development faster and at lower cost than sending the work out. The team is a deliberate mix of attorneys, paralegals, legal operations professionals and e-discovery professionals - the e-discovery group Guo previously led moved across intact.

AT&T did try the conventional route. A few years ago there was a small pilot phase testing a handful of use cases, and it produced some small successes. But as the pace of technology development accelerated, the team concluded it needed to move faster and grow much larger to build its own muscle and keep pace. In the second half of last year they explored how to build an AI-first in-house legal practice, became convinced it could work, and launched at scale in January. Six months in, the team has grown from 24 people to about 35, roughly 50% growth.

McMurrer is deliberate about not adopting the "AI-native" label that many new law firm ventures are using. Legal Edge values the human relationships on its teams - with departmental peers, business colleagues and outside counsel - and does not rely solely on its AI tools. The operating principle is simple and structural: a human asks the question, a human validates the output, and the AI machine sits in the middle. Teams are encouraged to use AI to build more efficient processes, but the people stay in the loop by design rather than by exception.

Guo's counter to the prevailing news coverage is that reading about AI produces a belief that you can sprinkle it over an existing process and get magic. The reality is that the tools still get things wrong or misinterpret - not necessarily the blatant hallucinations of a few years ago, but errors that require human validation nonetheless. The more you actually work with the tools, the more obvious that becomes, and the more apparent the volume of effort required to get a tool to do precisely what you need.

AI pixie dust - the idea that you come in, sprinkle some over everything, and it all suddenly becomes magical.

From the January boot camp onward, Legal Edge told its people they had freedom to experiment: build, test, iterate, test again, and do not be afraid to fail. New hires arriving from law firms find this genuinely refreshing, because at their firms they were expected to bill, and nobody can bill for a failure - so even lawyers curious about AI had only explored it on a personal basis. Legal Edge measures its impact quantitatively on cost savings, time savings and efficiencies per matter as work moves into the model, but explicitly does not measure the experimentation itself, treating it instead as on-the-job training that abstract tool training cannot substitute for.

Contracting looked like the obvious AI win and turned out to be the opposite, because every contract, every clause, every fallback provision and every business interaction is different. What made it work was pairing a Legal Edge person with one of AT&T's contracting experts and sitting down to break the thing apart clause by clause - what is accepted on clause A, B and C, and if not accepted, what fallbacks X, Y and Z apply - then drafting a playbook from scratch and rethinking the whole process. That exercise is what makes the AI behave the way it is supposed to. McMurrer's parallel point is that there is no one-size-fits-all solution either; their best results come from meeting people at their level, finding their individual pain points, and building for those.

Guo's position is not disinterest in technology - quite the opposite, self-described as interested to a nerdy extent. You should understand what sits behind your tools, how they are trained, how they are leveraged and which tool suits which type of work. What you should not do is become wedded to a particular product, because things are changing fast enough that today's favorite may not exist in two years or may have morphed into something else. The higher-value skill is a deep working understanding of what AI is genuinely good at and genuinely bad at, because that is what lets you choose tools at all. Being a master of one thing is worth less when that one thing can be five different things within a year.

Asked what qualities the model demands, McMurrer leads with a curious mind - a willingness to learn, to solve problems, to experiment, to challenge the status quo, and to fail and persist. Above all, believing there is a better way to do things. Guo adds perseverance and something more specific about the change management: shifting people off years of practice is genuinely hard, and what you used to do was not wrong, there are simply now better ways. That means talking to colleagues as colleagues, explaining things so they resonate, and showing the benefit rather than mandating the change. Role diversity helps here too, because lawyers are trained and inclined to be risk averse, while a legal ops professional will look at what a lawyer considers one step and correctly identify it as ten. The team culture is captured in an adage from AT&T's general counsel that Legal Edge has adopted: two is one and one is none.

Boersma's title, Spend Innovation Lead, effectively means that money going out the door to resource the legal team falls in that purview - and the framework starts a step before the sourcing decision. Before you can put the best people in the right place, in any industry, you have to understand the why of what you are buying in the first place. Each resourcing channel exists to do something different, and treating them as interchangeable capacity is what produces incoherent programs.

The first advantage is flexibility in the literal sense - as needs arise, a flexible resource can be deployed very fast. The second matters more and gets discussed less: proximity to the business. Outside counsel engagements are typically discrete, with hyper-discrete subject matter expertise, and those individuals do not sit with stakeholders inside the company. They are not positioned to understand how the business makes money, what actually matters, or what will culturally move the needle. Flexible resources can get embedded with teams far faster, often inside the company's own systems - laptop, email address, chat - and Boersma frames the design question accordingly: where in the business do we need resources that sit close to our stakeholders and genuinely understand the matters? Boersma is also careful that this is not only about lawyers, but legal professionals of all shapes and sizes.

Drawing on prior experience inside law firms on pricing and economics, Boersma explains that firms treat secondments as an investment - a way to deepen themselves in the operations of a business and learn its inner workings, aimed at long-term work and long-term return. But secondees are there for a short period, and the expertise is extracted quickly when they leave. Flexible talent inverts the logic: the company is sourcing a very specific resource for a very specific need it has identified. That difference in who owns the objective is why Boersma sees considerably more room for creativity in a flexible talent ecosystem. Secondments also require a scale of relationship that most engagements do not have, which makes matching the right secondee to the right moment a real operational problem.

In a legal department of Ford's size, the pre-centralization picture was a provider making contact with as many people as there are lawyers in the department, hoping to be top of mind whenever a need arose - inefficient for everyone, including Ford. Centralizing those relationships through legal ops means a provider talks to one person instead of fifteen or twenty. The critical constraint is what stayed decentralized: who gets hired sits with the hiring manager, full stop, because no third party should be telling a lawyer who they can and cannot work with. What centralization contributes instead is a better decision environment - all relevant CVs visible on an equal playing field, easy to access, with the bias of arrival order stripped out. The same logic extends to outside counsel: give a managing counsel enough data that the choice rests on more than "I know them, this is my gut." As Boersma frames it, operations is not there to replace frontline judgment; it exists to make that judgment easier to exercise and better informed. A concrete win is removing administrative load - a frontline lawyer scheduling interviews and combing resumes is time not spent with stakeholders.

Getting involved in whether work goes to flexible talent, an AmLaw 100 firm, or another provider starts with having earned stakeholder trust, and Boersma notes some teams welcome that involvement and some do not, often for good reasons on both sides. Being principled about which resource types work well and why cannot come from pontificating in the abstract - transactions, litigation and appellate work are genuinely different ecosystems with different practices, different decision-making and different people. Boersma's measure of success in serving stakeholders is simply how well those stakeholders and their business are understood, because without that there is nothing to deliver.

Boersma's team operates on an explicit premise: there are things they will be the experts on, and a larger share of things they will not be, and being honest about which is which is the point. Applied to vendors, this produced a genuinely useful discovery - flexible talent providers source talent in materially different ways, which means the information one provider needs to find the right person differs from what the next one needs. Rather than assume a single intake format works, the team asks providers directly what they are missing, which in turn helps those providers find better talent. Boersma credits much of the program's placement success across a range of vendors to that posture, plus a continuous improvement mindset Boersma calls the real game changer.

A marriage. Boersma draws the comparison directly from one, describing what makes a great partner as both people being honest about where things are going well and where they are not. Partnerships without that honesty do not last, in either setting.

Asked what distinguishes successful flexible practitioners, Boersma's answer is knowing what you are excellent at and leaning into it with confidence, rather than presenting as a universal fit. Paired with that, and just as important, is honesty about the edges - saying openly that a particular requirement is not where you are strongest but you are willing to lean in. In Boersma's experience across organizations that honesty is admired far more than a seemingly perfect cookie-cutter match. The same openness applies to logistics: where you are in your own practice, what else you have going on, and what a part-time engagement realistically looks like. Boersma also marks a shift in the market. Seven or eight years ago the industry was still debating where flexible talent could help; that question now sits firmly in the past, and the live question is simply how to find good people.

Unicorns - which, as Boersma notes, is Priori's own mascot. Companies engaging flexible talent are not looking for them. They are looking for people who are really good at what they do and willing to put in the work.

Boersma's theory for the current surge is that AI has introduced uncertainty across a wide range of work streams. Boersma pushes back on the framing first: forty years of AI research has not been about machines replacing humans, whatever the headlines say, but about optimizing machines for tasks they are good at so that humans can concentrate where they are best equipped to advise and counsel. What has changed is that the boundary between the two is now moving constantly. Companies are uncertain what AI will do to their business, their sales velocity, their contracting. Flexible talent is valuable precisely in that gap, because it is a resource type that can be put in place quickly to support the organization through the change - and Boersma expects that dynamic to keep accelerating adoption across the ecosystem. Wheatley adds a supporting data point from a decade in the space: conversations have moved from skepticism that anyone would choose to practice this way to five customers in a single week asking how to do it.

Asked for the most concrete thing a firm has actually shown - a changed way of delivering work rather than a license announcement - Niles points to firms accepting that the firm-client relationship itself has to evolve into something closer to a shared operating system spanning internal and outside counsel teams. The firms doing this are unpacking workflows and engaging on outcomes, on how risk and judgment are being handled, and on the guardrails and oversight that keep AI's risks managed while its benefits accrue to the client. Some are building genuine agentic trust models, and notably the mature ones distribute that mandate rather than confining it to the managing partner or executive committee, letting associates, paralegals and practice group teams engage with it. Niles's evaluation criteria as a buyer are concrete: whether AI and process automation hold up against traditional client expectations on confidentiality, privilege and security; enterprise-scale delivery; jurisdiction-specific and subject matter expertise; measurable client outcomes; and how automation translates into competitive pricing and fee structures. A firm that cannot be transparent about where it sits on that maturity curve, or about its own roadmap, or that will not commit to co-innovating, lands in a worse place than the laggard bucket - it undermines the trust the industry runs on.

Niles deliberately steps away from tech stack questions, because the clarity that unlocked Salesforce's acceleration came from a different direction. The operating principle is that this is not AI for AI's sake, or agentic transformation for its own sake, which means you have to be able to state what the function exists to do. Niles names three things for legal, compliance and trust functions: first, being a genuine enterprise business asset that accelerates growth while protecting trust and license to operate; second, operating leverage, meaning expanded capacity without proportional cost; third, translating that into operating structures and talent capabilities that can actually deliver it. The leverage piece resolves into a resourcing question - right work, right time, right service level, and the right person or function providing it. Having that thesis is what let the team connect AI work to real business value rather than to activity.

Salesforce Legal handles hundreds of thousands of requests a year, and the design question was how much of that volume could be automated or moved into a self-serve model - what many businesses call deflection - with AI embedded at every step including contract drafting, reviews and compliance checks. The detail Niles highlights is cultural rather than technical: because the company runs on Slack, it built a tool called Slackbot, and the legal team is the top user of Slackbot across the entire company. Legal is not merely consuming it either; the team is training and grounding the agents in its own expertise, playbooks and learnings, and calibrating what responsible AI means for the wider enterprise. Niles calls the result a force multiplier, and frames the destination for a technical audience as legal as an API - legal insights and even execution made directly accessible to the functions that need them.

Span of control is a private-sector management concept that Niles observes professional service firms do not naturally reach for. Applied to agents, it inverts what an individual contributor is: the traditional definition is someone who does not manage people, but Niles argues that in the future every individual contributor ought to be orchestrating agents - including, eventually, agents at their law firms across that shared operating system, and supervisory agents that check other agents' work. The appeal is not only capacity; it builds AI fluency and opens genuinely new career and development paths. The necessary counterpart is a new human role. Salesforce was, Niles is told, among the first companies to create agent managers - dedicated people at the intersection of legal expertise and agentic risk, who proactively mitigate legal and reputational risk without becoming a bottleneck, through risk scoring systems, user acceptance testing frameworks, standardized guardrails and governance checkpoints. Niles's argument for why the legal industry specifically should lead here is blunt: adoption will be a function of trust.

Managerial span of control. A leader's is measured in direct reports, and some executives famously carry forty or sixty. An individual contributor's should now be measured in agents.

Salesforce created a joint innovation group that put legal innovation and legal ops under one roof, and the point of it was not to centralize building but to distribute it - to make everyone a builder and surface what the teams already know. Niles describes the ideal as a beginner's mind blended with expertise, on the grounds that the people doing the work know what needs unlocking, what needs simplifying, and where friction actually arises, and those insights should determine both what gets built and what gets scaled. The most striking result is who became the advocates: people who were not technologists, and candidly people who had been fairly oppositional about where these tools could be relevant, are now among the strongest champions and ambassadors. Asked what they would do differently, Niles's answer is to have done exactly this more deeply and faster, with more champions across the teams.

Niles starts with the obvious failure case - firms of all sizes and reputations submitting hallucinations to courts - and notes that trust is earned and can be broken. But the answer is not simply more human review, because humans are fallible, tired and overwhelmed too; it is a hybrid of human accountability and oversight with agents and AI participating in the checking. The deeper gap is structural. In a piece written with Salesforce's chief AI scientist, Niles argues there is not yet a developed trusted agentic architecture, because every existing standard, rule and dynamic is fundamentally built around humans. That also requires clarity about when a workflow needs deterministic approaches and when probabilistic ones are acceptable, and it maps onto the classic distinction between rules and standards, where standards are looser and lean on judgment. Niles holds strong conviction, while expecting dispute, that AI agents are not ready at scale to complete negotiations on matters of importance, nor to adjudicate or arbitrate matters between themselves, and that humans are not ready either. Other named gaps include translation quality, where more progress would bring parties genuinely closer together. Cost discipline belongs here too: alongside client pricing there is token management and optimization, which turns triage into a question of what the right agent for the task is - or in the framing Niles was given, when you actually need the Ferrari.

The old London clearinghouse. Centuries ago, banks had to get in a room together and negotiate standards from scratch, because there were no rules yet governing whether you could trust the other side. Agentic legal work is at the same stage.

Asked the hardest question in the profession right now - if agents absorb the routine work junior lawyers learned on, how does the next generation develop the judgment clients are actually paying for - Niles gives a two-part answer. Part one is that AI literacy, or what Niles calls agentic workflow competence, has to be treated as a core professional skill at every level, from equity and non-equity partners through counsel, associates, paralegals and broader staff, with real programming behind it rather than tone at the top. Because so much legal learning has historically come through apprenticeship and hands-on work, developing this competence means using the tools and building hybrid workflows for yourself and your teams, not watching a video. Niles also anticipates more roles around strategic auditing, where the discipline is not only checking the first pass but improving it in ways that build judgment. Part two is the contrarian half: judgment has building blocks, has context, and is grounded in knowledge, which means it can in part be unpacked - and the goal of unpacking it is either superior judgment or, at minimum, raising the floor of the judgment delivered at scale. Lat offers a personal test case, having drafted a response to a note from a child's teacher, run it past ChatGPT, been told the draft was overreacting and over-disclosing, and found on reflection that the machine was right and that a second human reader preferred the machine's version.

Niles's position is that the billable hour has always been problematized, because even under it the client is assessing the firm on outcomes delivered, relationships deepened, benefits shared, efficiencies found and who is actually doing the work. The only question was whether those criteria were explicit or whether everything got unduly simplified into hours in equals value out. On that reading, Niles does not object to an hourly model that has fixed, measurable elements - provided it embraces the expectation that considerably more gets done inside each of those hours. The prediction is that clients move beyond pure hourly rates en masse, with more alternative fee arrangements and continuing debate about capped structures for particular unpacked workflows. Niles then floats a genuinely different incentive design: rather than a firm billing hours and expecting payment, give the firm a specified budget, ask it to deliver what the client needs within that budget, and let it absorb the gain from doing better - which pushes the firm to decide its own team composition, delivery model and use of AI and agents.

Gogo came to the model wanting greater control over career, over which work to accept, over which people to work with, and over how time gets structured - and stayed for the quality and variety of the work, plus the room to build a separate company alongside practice. Sareen's route was different: after firm environments including a traditional solo practice, the appeal was doing sophisticated, high-level, complex legal work without the business development and without clients new to working with an attorney at all. What both describe as the real deliverable is the break. Sareen can work extremely long hours for a client they value across a three or six month engagement, then take a full month off in which the inbox genuinely is not filling up and there is no accumulating backlog waiting - which is not possible in-house and not possible in traditional firm practice. Wheatley's observation is the one worth holding onto: more control has not meant working less or working on less sophisticated matters.

The fear that stops lawyers from trying flexible practice is whether the work will be there, and both speakers answer flatly that it has been. Their practice areas are the ones under the most pressure - technology transactions, commercial contracts, data privacy, and increasingly AI counseling, which Sareen now treats as roughly half the practice alongside data privacy. Gogo's point is that AI has increased demand rather than reduced it, because AI deals bring their own issues about what data is being transmitted and how it gets protected, and because growing companies, AI companies included, still need commercial transaction lawyers to get deals closed. Sareen adds a sharper version: two or three years ago pure data privacy experience was the differentiator, and now the lawyers who evolve alongside AI are the ones who get to have this experience of work being available whenever they want it. Gogo also names the real constraint honestly - the worry is not finding work, it is managing it all.

Both reject the assumption that law firm training is the critical credential. Gogo is generous about what it does teach - rigor, precision, responsiveness, managing complex matters under pressure - but both identify in-house experience as what makes a flexible lawyer effective, because the core skill is arriving inside an organization and understanding within weeks how it operates, how to navigate its people and internal politics, and how to answer legal questions in a form the business can digest. Sareen puts the distinction in one line: on the law firm side you have some leeway to say no, and on the in-house side that does not exist. Gogo extends it into what good advice looks like - not "here is the risk, full stop," but here is the risk, here is how we get to yes, and if we do these specific things we can get to what you are trying to do.

Asked whether business stakeholders treat a flexible lawyer as inside or outside the team, both describe the same arc. You arrive as an unknown, into a relationship where the business already assumes legal exists to stop it doing things, and you have to prove yourself. Gogo is matter-of-fact that this is universal rather than a slight against flexible practitioners - you get in, you prove yourself, they get comfortable. Once past that point, Sareen describes it as feeling entirely in-house while an engagement runs: no line between you and the business, frequently no line between you and the in-house legal team, and genuine reliance on you as a member of an integrated team. Gogo notes that good clients bring flexible counsel in as a partner and route sophisticated, problem-solving work to them, and that a real engagement always exists because of a real and immediate need - you are never simply parked.

Sareen dismantles the uncertainty worry first - the picture of going months without work has not matched reality, and would not for most lawyers genuinely dedicated to it. The honest version of that concern is narrower: if you need a predictable paycheck every two weeks and want to know exactly what each day holds, this is not for you. The trait that actually matters is the one nobody anticipates. Sareen's formulation is taking nothing seriously except practicing law, which in this context means having no ego about anything else. You are an integrated member of a team while simultaneously not being embedded enough in the corporate environment to hold opinions about how that team is run, how its meetings go, or how someone interacts with you. Something bothering you has to roll off, because the engagement exists for one or two discrete things and will end. Sareen is clear this coexists with taking the legal work itself completely seriously and wanting to make genuine connections.

Gogo draws a distinction that reframes the whole session. Many lawyers want out of the practice of law and are looking for the shortest possible route through it to something else. Both speakers have built full lives outside practice and still actively want to practice - Gogo describes being a lawyer as a privilege, and remembers hoping to get into law school at all, something many people never get to do. That makes flexible practice a conscious choice rather than an exit. The warning attached is specific: work is available everywhere but it will not simply arrive, and it takes the same personality and drive that produce success anywhere else. Engagements can also be genuinely all-consuming, with a deal closing that depends on your availability, because by then you are part of an in-house team and that is how the business sees you.

Sareen rejects the premise of being a natural salesperson and uses that to explain why platforms work. The genuine downside is stated plainly: solo practitioners doing their own business development generally charge more per hour because they work with clients one-off, whereas an engagement guaranteeing twenty, thirty or forty hours pays less than a standalone hourly rate - a real consideration, not a footnote. The upside is that the work never becomes salesy. Sareen is not doing business development, but meeting potential clients and assessing fit, which means the part of the brain that dreads selling never has to switch on - and for a lawyer who believes they would be terrible at sales, that makes this a viable path rather than a compromised one. Gogo's version emphasizes being pre-vetted: because clients trust the platform to send good lawyers, an introduction becomes a conversation rather than a pitch, which is also why sophisticated work arrives. Gogo also names the unglamorous benefit that matters in practice - not having to chase invoices or work out when payment is coming, because somebody else owns that.

Gumbs offers the most concrete signal of the session. Historically, when someone left the legal function the message from the team was that they had to be replaced immediately - there was real urgency attached. Over the last six months, as more AI tools have come into use, Gumbs has observed more delay in bringing in new talent and, more tellingly, more acceptable delay, because teams are using AI and other levers to cover the gap that departure created. Greenberg's evidence points the same direction from a different angle: the department is starting to allocate work differently than before, and Greenberg expects that to keep shifting toward a multiplicity of choices rather than a default channel, evolving slowly rather than all at once.

Edward Jones launched a new preferred counsel program this year, an eight or nine month effort that reduced the number of outside firms used by at least half. The substance was not merely consolidation - the team looked hard at what work was actually going out and thought strategically about what smaller and regional firms could do versus the AmLaw 50 tier. The result is a much smaller cohort of large-firm preferred counsel and considerably more work going to smaller regional and local firms. The pricing terms are the part other departments can copy directly: every firm entering the new arrangement locked its hourly rates for at least two years, and agreed that no further rate increases would happen without pre-approval from the Edward Jones team. Gumbs is careful to credit a shared services team rather than taking the credit, noting that "we" is being used extraordinarily royally.

Greenberg accepts the premise that big law looks fine - rates holding, demand steady, profits strong - and argues that this is simply not predictive. Firms should be managing to the future, because the metrics describe what is already behind them. What changes the picture is structural: AI-native firms, a proliferation of boutiques, and above all collapsed barriers to entry, since AI supplies research and resource capacity that previously required a firm around you. Greenberg expects this to start visibly playing out in 2027, and expects big law to continue existing - but for different reasons and on a different model. The firms that anticipate the shift will be very much in the mix and very successful; the ones that look at excellent numbers and conclude nothing needs to change will not, in Greenberg's view, exist. Greenberg illustrates the gap with an exchange relayed second-hand, in which someone in the legal AI space reviewed a firm's AI strategy and told them they would be out of business in five to six years - and when they expressed shock, offered that maybe it would be seven.

In 2007 the economy looked great. There was a point when Blockbuster looked great and BlackBerry looked unstoppable. Empirically strong performance is not the same as a sustainable position.

Gumbs describes a live comparison across litigation firms. Some deliver an AI-prepared deposition summary within an hour or two of the deposition ending, alongside the lead litigator's own read of how it went. Others take days or a week, which was entirely normal years ago and now reads as slow. The firms delivering near-real-time deposition and hearing summaries are getting more work, and Gumbs is direct that satisfaction as a consumer rises accordingly, out of an unwillingness to wait a week for something that can demonstrably be produced in thirty minutes. Greenberg adds that this is only half the change. A decision relevant to the industry now arrives with an AI-generated summary from a young partner or senior associate within five minutes of issuance - which Greenberg forwards to the CEO a minute later, explicitly labeled as AI-generated, not as a criticism but out of delight at the delivery. The other half is services that did not previously exist at all: the team recently fed a body of material into a database and used AI to produce a strategic take on their negotiation approach.

Edward Jones has been insourcing for a decade, beginning under Gumbs's predecessor. Ten years ago the legal function was literally a handful of lawyers alongside a few hundred compliance staff, with roughly 70 to 80 percent of legal work done outside; today it is a legal team of around 160 within a legal, compliance and risk organization of about a thousand. The cost logic is straightforward - sustainable, predictable work done internally costs materially less, and the discipline is identifying work you know you will need in future so you can bring that talent in rather than paying a premium later. But Gumbs rates the second benefit higher: with outside counsel, the firm becomes the expert in your business, whereas what Gumbs actually wants is a division full of people who know both the law and Edward Jones. Looking forward, the firm's "future of work" initiative runs on four levers applied across every organization, not only legal: AI and automation; outsourcing, split between offshore work and contractors including flexible talent; upskilling and reskilling people into different kinds of work; and process optimization, on the recognition that many processes were designed twenty years ago and may no longer make sense. Greenberg's parallel example is a digital twin - an area of law that lost its dedicated lawyer and was absorbed by a colleague with deep expertise for whom it is not the highest and best use of time. Rather than hand the area to an agent, they are building a twin that learns it from the department's own institutional knowledge, which both reduces outside counsel dependence and frees that lawyer for work where the expertise matters more. Greenberg's read across peer conversations is that departments are largely not shrinking deliberately but letting attrition do it, using the technology to absorb each lost role - a slow decline in headcount while the department's output grows.

Greenberg has written about a tyranny of choice paralyzing in-house teams, now that work no longer simply defaults to big law and the options include AI-native firms, various kinds of alternative providers, new technologies and boutiques. Greenberg's approach is to place a lot of small bets, with a legal ops lead curating the choices so the team is not sitting through endless demos - which is why Greenberg considers the anthropology of the in-house team so important, because a legal ops professional who understands both the technology and how the business and legal team actually work is uniquely positioned to filter. Greenberg sees alternative providers as similarly well placed to guide clients given their exposure across many clients and tools, and identifies an opportunity law firms are largely not taking: coming to a client with what they are seeing elsewhere and offering to help implement it. Gumbs reframes the problem entirely - it feels to Gumbs less like a tyranny of choice than a tyranny of pitches, a constant stream of firms in every bucket. Gumbs is not especially torn about selection, because each provider type has a natural niche: a six-month vacancy suggests flexible talent, a highly specialized need suggests a boutique that does only that. Greenberg adds a category worth naming, off-label AI - a tool marketed for one purpose that solves a different one after adaptation. Greenberg's team took an M&A tool and applied it to a process carrying revenue and legal risk, after a team member recognized the analogy, with the legal ops person acting as liaison to the platform.

Legal ops as the skeleton key - what opens the door to a new opportunity, though Greenberg is clear the door does not open automatically, and implementation does not always go as planned.

Greenberg names two things that change and one that does not. The first change is that knowledge management stops being a noun. It used to be a compilation, a thing that existed; now it is an ongoing process of finding data, adding to it and making it accessible, and it seeps into the architecture of the whole team. The second is that legal ops becomes more important still, partly for the reasons already discussed but also because the legal team is becoming a gateway for enterprise AI - these tools are not like downloading a new version of a word processor, they require understanding how people work, what problem they are solving and where friction sits, which is a vantage point legal uniquely has. The thing that does not change is curiosity. Prompt engineering is real and important, but what makes it effective is being curious, thinking by analogy and spotting gaps - the things lawyers were trained to do. That shifts hiring away from "have you done these filings, have you done these deals" toward lawyers who are ambidextrous, curious and agile with AI, and Greenberg expects the liberal arts capacity of a lawyer to matter more than before. Greenberg also expects emotional intelligence to become more valuable, not less: as more is turned over to machines, the ability to empathize and to be trusted is what big law's moat is actually made of. Gumbs's parallel from Edward Jones is a shared services function across legal, compliance and risk containing a dedicated data team, built because regulatory reporting, discovery and compliance constantly require data that is neither easy to find nor in one place - alongside a target of over 60% average daily AI tool use across the legal division, higher than the rest of the firm.

Greenberg's model of the traditional firm is a cinema. A blockbuster film commands a handsome ticket price, but the theater makes its money on popcorn and soda. A firm charges a lot for a partner - that is the blockbuster - while the economics run on the sixth-year associate, the third-year and the five second-years staffed to the matter. Boutiques, AI-native firms, alternative providers and flexible talent all disrupt that, because once buyers can disaggregate partner-level judgment from commoditizable work and price the latter separately, general counsel start assembling an a la carte menu instead of eating off the prix fixe one. Notably, Greenberg says they may pay more than today for genuinely unlevered access to a specific partner's expertise, paired with a provider handling diligence and document review. The scale gap is still enormous - Rubin cites the AmLaw 200 at roughly $190 billion in revenue growing 13% year on year, against alternative providers at around $30 billion growing 18%, faster but not fast enough to catch up. Asked directly what a firm should do to keep the work, Gumbs starts by stipulating what everyone skips: the quality has to be spectacular. Then, deliver more value through speed, and above all know the client's organization. Generic advice is Gumbs's biggest pet peeve and has nothing to do with money, and that criticism includes even AmLaw 25 firms among those with wonderful lawyers who are poor at customizing advice to the specific business receiving it - a narrowing margin for error, because the ability to get generic advice has grown while the value of specific advice has risen. Greenberg wants a conversation about value rather than price, and more sophistication in pricing than the era when it was an older partner with a legal pad deciding that 10% seemed right. Greenberg's most instructive example is a firm that ran a joint AI experiment with the Cox Media Group team, came back and said they had learned as much as the client had, and took an amount off the rate - which Greenberg could not verify as mathematically correct but which felt fair and felt like it was getting at value. Both leaders end on the same note about why optimism is warranted: Edward Jones is itself an advice business facing the same automation, and what clients want is not the data but a trusted relationship that tells them what to do about it.

Shapiro describes the gap between frontier capability and enterprise adoption as one of the most important facts in the economy right now, and is even-handed about who is at fault. Technology companies see genuinely transformative capability, and skip forward to a world where the transformation has already happened, without fully reckoning with the slow bureaucratic work of getting there - which makes their timelines optimistic. Meanwhile lawyers hear that AI will rewrite their profession, try the tools with a brief prompt about a work product they know well, receive a predictably decent but not-ready-for-primetime response, and conclude the technology is overhyped and can be deferred. Both sides need this resolved: technology companies must eventually show real enterprise efficiency and revenue to justify their valuations, and Shapiro does not yet see numbers indicating that is happening, while firms face client questions about where the AI efficiency has gone and pressure from AI-native upstarts raising very large venture rounds specifically to outflank them. What closes the gap is the input layer - lawyers learning how to talk to the models in order to hand off execution and automate legal process. Crucially, Shapiro argues this has to come largely from legal professionals themselves, and frames the work of the next decade as helping lawyers who already have the judgment and experience translate that into handing off execution.

A genie. What is remarkable about the technology is how elaborate a wish you can cast, and what is dangerous is the classic genie failure mode of literalism - wish for a billion dollars and it falls out of the sky onto your head, wish to be the smartest person on earth and everyone else disappears. Precision in the asking is the whole skill. Rubin called it the best analogy for prompt engineering yet heard.

Pike attributes the split to incentives rather than aptitude. In-house lawyers are told to do more with less, are measured on business outcomes, and are actively trying not to be seen as bottlenecks. Law firms have real incentives to preserve the billable hour and a reputation for conservatism about new technology. The usage evidence is striking: stack-ranked against other knowledge workers, lawyers are the number one power users of the Cowork function - the agentic capability that lets the model operate a desktop computer and work autonomously for hours or days on large projects, which knowledge workers have only had access to for around six months. In-house teams are further ahead partly for a mundane reason: where a company already has the tool deployed, there is no procurement cycle to clear, which a firm starting from scratch has to work through. What Pike is most enthusiastic about is that the ecosystem is open, with builders creating and sharing skills and workflows that others remix.

Shapiro's reframe is the most important economic point in the session. The value big law delivers is perversely different from what its invoices describe. High partner rates matter less to profits per partner than the leverage model does - very large teams of associates doing document review in litigation, due diligence in transactional work, and sometimes checking typos or formatting signature pages. That is work it is hard to imagine a smart Fortune 500 general counsel paying for in five years. The bad news is that the economic engine has to change. The good news is that this was never what clients were trying to buy, which means the value proposition is intact and the margins are simply getting better - something almost any other industry would celebrate. Law's difficulty is that it is one of the last industries running a cost-plus model, which was a poor incentive structure to begin with. The move required is from cost-plus leveraged billable hours to value-based pricing, whether flat fees, success fees, contingency fees or subscriptions. Big law disproportionately holds the best lawyers in the world and the institutions that carry the largest matters, and clients genuinely want that judgment for bet-the-company litigation and front-page M&A. The task is capturing that value as judgment rather than as hours. Rubin notes the irony surfaced earlier in the day: for years in-house teams argued firms should leave the billable hour for the clients' sake, and now firms may need to leave it for their own.

Both speakers reach for the Jevons paradox, and Shapiro applies it precisely: legal supply is severely constrained relative to demand, and there are many things people should use lawyers for and do not, because it is expensive, the cost is unpredictable, and finding a lawyer you trust is a baroque process. Cheaper delivery therefore means more legal services consumed, not fewer - largely good for access to justice, potentially bad in that anyone can now generate a plausible-sounding demand letter, which may simply mean more litigation. Pike's parallel is that personal computers and spreadsheets were expected to end accounting and instead expanded it, possibly moving it down-market, which suggests real mid-market opportunity in legal. The distinction Shapiro insists on is the one most commentary misses: AI automating 50% of legal jobs, which Shapiro thinks is wrong, versus AI automating 50% of each legal job, which is what Shapiro observes among genuine power users. And those power users are not working fewer hours - to a person they are working more, on higher-value work. For junior lawyers the implication is that value-add stops being billable hours of menial work, which is good for the juniors who exist, but it also means law school is a poor choice if the motive is a low-risk upper-middle-class trajectory rather than wanting to practice. Associate development has to change accordingly: the premium is judgment and client service, and the way to build judgment is to be in the room when the partner is making the difficult calls.

Jiro Dreams of Sushi - the apprenticeship model where you make rice for years before you are allowed to touch the fish. Shapiro's point is that armies of associates doing grunt work is not how judgment gets built from here.

Pike connects the surge in law school applications to something other than economics - people are applying because they want to help and to be counselors, and the unmet need is enormous: immigration cases where people cannot find a lawyer or know their rights, landlord-tenant disputes, and a large volume of public service work that goes undone. The structural blocker is not desire but debt, and the career treadmill that law school debt and career services pipelines create. Pike sees AI changing that arithmetic, both because it is easier than ever to start your own practice as costs fall, and because tools can reach people directly. Pike's illustration is a JD-MBA student at UNLV who teaches continuing legal education and runs change management for lawyers in Nevada while still in school, held up as the model for taking charge of a career path. Quoting Sonia Ebron of Courtroom 5 after launching Claude for Legal plugins with them, Pike frames the mission plainly, and is also honest about the counterweight: the explosion in pro se claims includes people who do not understand the technology is hallucinating, which strengthens rather than weakens the case for lawyers being involved.

Rubin raises a genuine tension: these tools are probabilistic, so five identical questions can produce five different answers, and that is a feature rather than a defect - which pulls directly against a decade and a half of legal tech pushing toward standardization through the NVCA forms and the oneNDA project. Pike's answer is candid about limits: Anthropic was founded as a research lab, publishes what it finds on interpretability, and is transparent that the field does not fully understand why models do what they do, which for legal work is a useful disclaimer rather than an embarrassment. What has improved is grounding - giving models the same authoritative sources and tools a lawyer would use makes outputs more repeatable and reduces the confident fabrication of three years ago. Pike also offers a deeper possibility: perhaps there is not always a correct answer in law, the gray area is real, and that ambiguity is precisely where human judgment earns its place. Shapiro agrees the mature system is hybrid, with deterministic code where there is an objective answer - share counts and per-share math in Shapiro's deal tooling, or verifying that case names in a brief correspond to real cases - and generative models where the work is qualitative, taking a lawyer's intent and expressing it in language. On that view the variation does not matter, because different valid phrasings can capture the same logic. Shapiro is impatient with how much of the public conversation fixates on hallucination: submitting fabricated case names to a federal judge is a failure of judgment, not of technology, and it is not too much to ask that an associate read the case names before filing. There are also techniques for having a model check its own work, and for having models from different labs check each other. The deeper hazard is subtler - because these systems produce a plausible response to almost any input, there is real temptation to treat them as permission to stop thinking. Shapiro's own practice is the inverse: do the hard cognitive work first and use it as the input, which now means walking outside and free-associating into a phone for minutes at a time and coming home to a good draft, working harder rather than less.

Giving the model a calculator to do math. Grounding it in the tools and primary sources lawyers actually use is what makes the output repeatable, rather than asking it to produce the answer from memory.

Shapiro is blunt that there is a great deal of theater, and that genuinely skilled practitioners remain a small percentage of the profession - but argues the difference is not subtle once you know what to look for. The first signal is efficiency, which buyers are already watching. Shapiro's example: on a venture deal where the term sheet required the early-stage client to cover the investor's legal fees up to $35,000, Shapiro asked opposing counsel at a large firm to produce the first draft of the industry-standard documents given the client was paying anyway, and was told the $35,000 would not cover drafting them. Weeks later, working with Claude Code over a weekend, Shapiro built a deal engine - a non-AI tool that connects to the model through an MCP server - which takes a pre-round cap table and a term sheet and produces solid drafts of all the documents with the math, per-share prices and preferred series correct. Firms with that capability will price differently. The second signal is subtler and more telling: the things that used to be too expensive in human effort and are now nearly free, delivered as extras. Clients in a priced round always want to know how far along the deal is, how long until close, and what the last round of edits was actually about - and no lawyer previously had time to write that memo, nor would a client pay for it. Working inside Cowork, Shapiro can point the model at revisions it already has and ask for a memo with full detail and nuance, pitched for a non-lawyer. It is one prompt, it costs nothing to produce, no extra charge is passed to the client, and clients love it.

Rubin sets the frame that gives the session its stakes: the technology is arriving whether the profession likes it or not, but the shape of the industry is genuinely within its control, and nobody has to accept the future being sold to them on where accountability lives or what it means to be a lawyer. Just because a technology could stand up in court does not make that the future anyone should want - it has to serve the profession's ends. Shapiro takes up the phrase and applies it to what Shapiro calls the human kernel left at the end, and is deliberately dismissive of the Silicon Valley preference for "taste," because the word blurs together judgment and pattern matching, and pattern matching is something models are very good at. What Shapiro believes is uniquely human, and will increasingly define what practicing law means, is decision-making in the face of uncertainty. That is what lawyers actually do, transactional and litigation alike: helping clients make difficult decisions about legal challenges. Models are strong at execution once you have explained what you want, and genuinely weak at making the decision for you, which is precisely where people get into trouble by expecting otherwise. The practical conclusion is a training prescription: get comfortable making difficult decisions when it is unclear what to do, because building that muscle is the most valuable thing a lawyer can do now.

Asked what stood out, Basha Rubin points to the chasm that recurred across the day's conversations, between the models' capability and what adoption actually looks like at law firms and in-house teams right now. It was less any single session's topic than the current running underneath all of them.

Rather than assume the gap simply narrows over time, Rubin leaves it genuinely open. Law is fundamentally a human profession, humans advising other humans through complex challenges, so part of the distance between capability and adoption may be structural rather than a temporary lag, and worth designing around rather than waiting out.

Mirra Levitt reframes the whole conversation: legal keeps talking about adoption and its challenges, but that is too thin a way to think about the moment. The real question is what human systems need to be built, alongside the technology, for teams to take advantage of it and keep achieving excellence for clients and firms alike.