Who Controls Legal AI: Workflow, Model, or Platform?

Clio acquired Learned Hand in September 2026. Learned Hand sells to courts, not law firms. Its tools handle filing review, legal research, bench memoranda (the research summaries clerks prepare for judges), and draft orders.

Read the acquisition as the latest entry in a sequence of acquisitions and product expansions that goes back to 2018. Together they show a company accumulating workflow surface across every step of legal work, from client intake through court review. Clio is betting that whoever owns the surface where legal work happens will control where AI output lands, whichever model produced it.

Two serious competitors are running different theories of control. Harvey owns the model layer and wrote the evaluation standard. Google owns the enterprise platform and is packaging legal skills inside it. Each position is credible, and each carries a cost the other two avoid. What you want to know is which position is hardest to replicate, and which buyer ends up deciding among them.

Clio: Own the Workflow, Let the Model Be Commodity

Start with the acquisition timeline. Since 2018: client intake, litigation calendaring, document automation, embedded payments, court e-filing, enterprise matter management via ShareDo, $1 billion in legal research and docket intelligence via vLex, Canadian case law via Jurisage, and now judiciary workflow via Learned Hand.

The logic is that if you control enough of the legal process, AI becomes a feature of your platform rather than a product competing with it. Clio doesn't need to win the model race if it controls what happens to the model's output.

The buyer has moved with the ambition. Clio historically served solo practitioners and small firms. ShareDo brought large firms and corporate legal departments, including DLA Piper, Linklaters, and Herbert Smith Freehills. vLex's Vincent AI reportedly reaches eight of the ten largest global law firms. Clio claims 400,000 legal professionals across 130 countries. No public source breaks that number down by firm size or shows how much of the portfolio any single customer actually uses. Treat it as a reach claim, not a depth claim.

The cost of this position is neutrality. Learned Hand puts Clio on both sides of the courtroom: tools that help lawyers advocate, and tools that help judges evaluate that advocacy. Court procurement governance keeps getting more specific. The federal judiciary's interim guidance says courts should not delegate adjudicative authority to AI. California Rule of Court 10.430 restricts entering nonpublic information into public systems and requires verification. The National Center for State Courts recommends contract terms that bar vendors from using court data for training or commercial purposes, and court approval before material model updates.

None of these authorities currently prohibits a court from buying from a company that also serves law firms. But data isolation, auditability, change control, and nondelegation of judicial authority add up to a trust problem a pure law-firm vendor never has to manage. Clio now has to stay credible with institutions whose core concern is independence from the parties appearing before them.

I wrote previously about consequence networks, the human labor that absorbs product failure. In legal AI, that labor belongs to the lawyer who checks a research result before citing it and the judge who reviews a draft order before issuing it. Clio's acquisitions put it closer to both of those verification points than any competitor. That proximity is Clio's main asset, and it's why the neutrality question is structural, not cosmetic.

Harvey: Own the Model and the Scorecard

Harvey works from the other end: own the model layer, and own the standard models are judged by.

The company has built proprietary post-training (additional training on top of foundation models to specialize them for legal tasks) and published the Legal Agent Benchmark (LAB), an open-source evaluation framework with task taxonomies, execution harnesses, and scoring approaches for legal AI. The customer base is substantial: 80% of the Am Law 100 (the hundred largest US law firms by revenue), legal teams at five Fortune 10 companies, and more than 3,500 customers across 85 countries. Those figures carry the same ambiguity as Clio's 400,000. "Customer" is undefined. It could mean a firmwide deployment, one practice group with paid seats, or something in between.

The strongest public evidence of production depth comes from CMS, which reported more than 1,100 daily active users and nearly 3,000 monthly active users before expanding Harvey across its international organization to more than 6,500 users. The firm published those numbers, not Harvey's marketing team, and they predate the wider rollout, so current usage may be higher.

LAB is the more interesting competitive move, and it has real technical uptake outside Harvey. Vals lists it as an industry partner benchmark with results across 76 models. Artificial Analysis built an implementation from Harvey-supplied tasks. Anthropic included LAB results in a model system card. A research preprint reports more than 12,500 agent trajectories on LAB tasks.

What LAB doesn't have yet is a buyer. No law firm, corporate legal department, or procurement body has publicly adopted it as a required vendor-selection framework. It has readers and implementers, but nobody using it to decide what to purchase. Harvey wrote the task taxonomy, and at least some external implementations run on task sets Harvey supplies. That makes it an ecosystem seed, still short of an industry standard. Moderate confidence this holds: nonpublic procurement use is possible, but at meaningful scale it would likely surface in vendor marketing or legal-press coverage.

The cost is the distance between evaluation authority and workflow proof. Harvey can show that its models perform well on its own benchmark. LAB alone can't show that the output survives professional verification: that a lawyer at a 500-person firm trusts it enough to cite without rebuilding the research, or that an in-house team uses it for work product that goes to a regulator. CMS's usage data is the best public evidence that this trust exists in practice, and it comes from the customer, not the benchmark.

If LAB becomes the industry's evaluation standard, Harvey's position is formidable. If it stays Harvey's evaluation standard, it functions mostly as marketing, however technically credible.

Google: Own the Platform, Rent the Vertical

Google's Gemini Enterprise for Legal takes a third approach: build the platform that vertical products plug into.

The preview includes packaged workflows for contract review, regulatory monitoring, legal research, NDA drafting, redaction, and citation verification. The connector list is broad: iManage, NetDocuments, Docusign, Everlaw, RelativityOne, Thomson Reuters HighQ, CourtListener, Courtroom5. Governance controls (identity management, audit logs, data residency, customer-managed encryption) come from the Gemini Enterprise platform.

Look closely at one entry on that list. Harvey appears as a named integration inside Google's legal product. Google is positioning itself as the orchestration layer that connects specialist tools, which means its legal offering is partly built on a company it competes with directly. Harvey has every reason to limit or withdraw that access as it builds its own enterprise surface. Thomson Reuters, Everlaw, and the other named connectors sit in the same position: they participate in Google's ecosystem only as long as it serves them. When a platform's vertical capability depends on partners who may become competitors, its position can degrade without Google doing anything wrong.

The cost is dependency compounded by a vertical depth gap. The product was still in preview at research cutoff, with no disclosed pricing. A September demonstration mixed standard prompting with purpose-built apps and measured nothing about accuracy, exception rates, or production reliability. No public source compares Google's legal workflows against Clio/Vincent or Harvey on the same matters. Google has the distribution and governance infrastructure to be a serious player. The legal-specific evidence doesn't exist yet, so treat its competitive viability in legal AI as speculative until the product exits preview with production data and pricing.

Google's bet works if the buyer is a CIO or legal ops leader who ranks governance and integration breadth above task-specific accuracy. It works less well if the buyer is a practice group leader who needs to know the tool won't hallucinate a citation in a brief due tomorrow morning.

Side by Side

Where This Leaves the Vertical

High confidence: Clio's control position is the hardest to replicate. The neutrality burden is real and will get heavier as court governance matures. But the acquisition sequence took eight years and more than a billion dollars, and it spans workflow positions no single competitor can assemble quickly. Harvey can improve its models and grow LAB adoption without buying anyone. Google can add connectors and legal skills incrementally. Clio had to buy the workflow, the data, the research corpus, and now the court-facing tools, one acquisition at a time.

The rest of the vertical is watching this three-way competition from inside it. Point solutions like Everlaw and Relativity already appear as connectors in Google's product. Thomson Reuters supplies content through more than one of these architectures. The dozens of legal AI startups that raised seed and Series A rounds over the past two years face a narrowing set of options: integrate into one of the three, find a workflow niche none of them covers, or get acquired. Legal AI is consolidating around fewer control points. Whether that ends with three coexisting architectures or one winner depends on which buyer does the deciding.

If practice leaders who judge output quality on specific legal tasks make the call, Harvey's model depth wins. CIOs weighing governance and platform consolidation favor Google. Managing partners and court administrators looking at operational coverage favor Clio. But purchasing authority for AI tools at large firms hasn't settled on any one of those roles. Innovation partners, CIOs, practice group leaders, and executive committees all claim jurisdiction, depending on the firm and the price tag. Which role ends up holding that authority is itself a leading indicator. Watch which buyer each company is cultivating — it tells you which architecture they expect the market to select.

Harvey's LAB moving from marketing to infrastructure. The trigger is a law firm or procurement body publicly adopting LAB as a vendor-selection criterion. Until then, Harvey has a benchmark that technical evaluators find useful and buyers haven't adopted.

Google disclosing pricing and production performance data. Until that happens, the legal product is a preview, and a preview tells you how Google wants to be positioned more than where it actually stands.

Clio publishing, or being required to publish, how it separates court-facing data and workflows from its law-firm business. That is the neutrality test. Given where court governance guidance is heading, expect the disclosure to be forced eventually. The open variable is timing.