The Intake — Weekly briefing

This week in legal AI

Harvey publishes the first open-source benchmark for legal AI agents — 1,200 tasks across 24 practice areas, backed by OpenAI, Anthropic, Mistral and DeepMind. Enter, a São Paulo startup, triples its valuation to $1.2 billion in eight months on a $100M raise from Founders Fund, Sequoia and Ribbit — becoming Latin America's first AI unicorn. Slaughter and May completes the Magic Circle sweep by going firmwide with Harvey. Linklaters launches Applied Intelligence, a team of lawyers and data scientists selling bespoke AI solutions directly to clients. And the US federal government intervenes in the lawsuit challenging Colorado's AI Act, taking sides in the AI regulation fight for the first time.

Week of 1 May – 8 May 2026
Category Market intelligence
Reading time 8 minutes
01 — The week at a glance

Five stories that matter

April 30 — Adoption
The last holdout among London's Magic Circle firms announced a firmwide deal with Harvey covering all practice groups — M&A, due diligence, regulatory research and document analysis. Every Magic Circle firm has now committed to a named AI platform. The last time an industry-defining technology moved this quickly through the entire top tier of a professional services market, it took a decade rather than three years.
May 4 — Regulation
The Department of Justice intervened in xAI's lawsuit against the Colorado AI Act, filed the same day. Separately, SB 189 — a rewrite of the original law — moved through committee and proposes pushing the effective date back to January 2027. The collision between federal preemption ambitions and state-level AI regulation is no longer theoretical: it is now in a courtroom.
May 5 — Funding
Enter, founded in São Paulo by Mateus Costa-Ribeiro, raised a Series B led by Founders Fund, alongside Sequoia Capital and Ribbit Capital, tripling a $350M valuation in under eight months. The company builds AI agents that handle mass litigation — reading each lawsuit, pulling case-specific data, and drafting a defence — for clients including Itaú, Santander, Nubank, Mercado Livre, and Airbnb. It processes more than one million cases a year inside a court system that generates more new lawsuits annually than any country on earth.
May 6 — Product
Harvey open-sourced its Legal Agent Benchmark (LAB) — the first industry-wide framework for evaluating whether an AI agent can do real legal work. Each task mirrors how work is assigned in a law firm: an instruction, a client matter with relevant materials, a required work product. The benchmark launches with support from Nvidia, OpenAI, Anthropic, Mistral and DeepMind. Harvey simultaneously announced 500 agents live in the platform, with an Agent Builder in early access.
May 7 — Strategy
Linklaters described Applied Intelligence as a "first of its kind" team that co-designs and delivers bespoke AI-enabled workflows for clients' most complex challenges — working across frontier models and large data sets to build custom tooling that an off-the-shelf product cannot deliver. The unit is client-facing and revenue-generating. It is a law firm using AI to sell what a law firm has never sold before: the software layer underneath the legal advice.
02 — The benchmark question

Harvey just defined what "good" looks like for legal AI agents

The release of the Legal Agent Benchmark is, on one reading, a product announcement. On another reading, it is a bid to set the standard before anyone else does. That distinction matters enormously for how this market develops.

Every mature software category eventually gets a shared benchmark that buyers and vendors both reference. Security has CVE and CVSS. Machine learning has MMLU, HELM, and a dozen others. Legal AI has had nothing that measures what actually matters: whether an AI agent can take a real legal matter and produce a work product a lawyer would sign off on.

Harvey's LAB is an attempt to fill that gap. The benchmark is open-source on GitHub, built around 1,200 agent tasks across 24 practice areas and evaluated against 75,000 expert-written rubric criteria. The structure is explicit: an instruction, a client matter containing relevant materials, and a work product. Not a question-answer pair. Not a research retrieval test. A task that looks like the actual work.

The backing is significant. Nvidia, OpenAI, Anthropic, Mistral and DeepMind are all named launch partners — which means the model providers are aligning to a legal-specific capability measure at the same time as the legal-AI vendors. Harvey has declined to publish a leaderboard on day one, saying it expects the dataset to evolve. That is a deliberate signal: the benchmark is designed to be the long-term frame, not a one-week news story.

What it means for buyers

For the first time, an enterprise legal team evaluating a legal AI agent has a shared vocabulary to work from. The question is no longer "does this tool make lawyers faster?" — which is unmeasurable and unanswerable in a trial. The question is "where on the LAB task set does this agent pass, and where does it fail?" That is a procurement-grade question. It will change the conversation in every RFP that a legal AI vendor answers from this week forward.

1,200+
Agent tasks in the Harvey LAB benchmark at launch
24
Legal practice areas covered by the benchmark
75,000+
Expert-written rubric criteria used for evaluation
500+
Agents live in Harvey's platform as of May 5, 2026

Why the benchmark is also a moat move

The legal-AI market has so far competed on a mix of credentials (Magic Circle clients), fundraise headline ($11B valuation), and narrative ("fiduciary-grade", "agent-native"). The LAB changes the competition surface in Harvey's favour. By publishing the benchmark — and by being the company whose products will inevitably score well against it because Harvey built it — Harvey has done something sophisticated: it has created a public measure of capability that its own training data is optimised for, while inviting every competitor and model provider to be evaluated on it.

The right comparison is not to a software benchmark. It is to what AWS did by publishing the Well-Architected Framework, or what Salesforce did by publishing the enterprise SaaS evaluation criteria it knew its own products would pass. The entity that sets the standard usually wins the market that forms around it.

03 — The global moment

Latin America just produced a legal AI unicorn — and the thesis is the same one

Enter's $1.2B valuation is not an outlier. It is the clearest evidence yet that the structural problem legal AI is solving is not an Anglo-American problem. It is a feature of every legal system built on high-volume, rule-based work done by expensive human labour.

Brazil's court system receives more new lawsuits per year than any other country in the world. The practical consequence is that large Brazilian companies — banks, fintechs, platforms — each face hundreds of thousands of active cases at any given moment. The traditional response was to maintain enormous internal legal teams or outsource to law firms at per-case rates that added up to material operating costs. Enter builds AI agents that read each incoming lawsuit, pull the case-specific facts, and draft a defence brief. The company processes more than one million cases a year.

The founder's background is worth noting. Mateus Costa-Ribeiro became Brazil's youngest practising lawyer at 18, finished Harvard Law, passed the New York Bar at 20, and left a fully funded Stanford MBA seat to start the company. The $100M Series B was led by Founders Fund and joined by Sequoia Capital and Ribbit Capital — three of the most structurally rigorous investors in technology. All three diligenced the mass-litigation use case and concluded it was venture-scale.

What Enter is
An agentic legal service. The client sends a lawsuit. The agent reads it, drafts a defence, and routes it for a lawyer to approve. The unit of price is the outcome, not the lawyer-hour. Clients include Itaú, Santander, Nubank, Mercado Livre, and Airbnb. This is not a productivity tool sitting inside a law firm. It is a replacement for the law firm on a defined category of work.
What the moat is
Not the model. The model is available to everyone. The moat is the Brazilian legal corpus — case law, court filing formats, judge-specific patterns, procedural quirks across 27 state courts — and the training data accumulated across one million real cases. A general-purpose AI model cannot build that corpus in a year. Enter has a six-year head start on it.
The Flank read

Enter is doing in São Paulo what Flank does in London and New York: taking the category of legal work defined by high volume, rule-based repetition, and a fixed relationship between input and outcome, and moving it from expensive human time to supervised agents. The Brazilian court system is an extreme version of the problem — more volume, less predictability, harder to automate — but it makes the proof stronger, not weaker. If the thesis works at one million Brazilian lawsuits per year, it works at ten thousand UK NDAs per year. The cost structure is the same. The inexpensive work is being done by expensive people. That is the structural inefficiency agents remove.

04 — BigLaw's pivot

The Magic Circle is complete — and one firm just started selling AI as a service

Slaughter and May's firmwide Harvey rollout is a milestone, but the more interesting story of the week is what Linklaters did next. One firm completed the tool deployment. Another firm started selling the deployment as a product.

The Magic Circle is now fully committed

Every Magic Circle firm has now publicly committed to a named AI platform. The spread is instructive:

Harvey
A&O Shearman
First major firm. Adopted Harvey early, pre-merger, and expanded post-merger across the combined entity.
Harvey
Slaughter and May
Firmwide from April 30, 2026. All practice groups: M&A, due diligence, regulatory research, document analysis.
Legora
Linklaters
Linklaters has largely gone with Legora for its primary AI platform. Also now operating Applied Intelligence as a client-facing unit.
Microsoft
Clifford Chance
Doubled down on Microsoft Copilot and the firm's own Assist AI tool. A bet on horizontal infrastructure over vertical legal AI.
Anthropic
Freshfields
Struck a firmwide agreement with Anthropic — the only Magic Circle firm betting on the model provider directly rather than a vertical layer built on top.

The variation in choice is itself a signal. The market has not converged on a single platform. Harvey has the most Magic Circle logos, but Legora, Microsoft and Anthropic are each present. The competition is real, and the firms with the most sophisticated procurement processes are still reaching different conclusions.

Linklaters and the new revenue line

The more structurally interesting event of the week is Linklaters' Applied Intelligence launch. The unit is not an internal AI programme. It is a client-facing team of lawyers and data scientists that co-designs and delivers bespoke AI workflows for client use — analysing large data sets, building custom tooling, and selecting from frontier models on a project-by-project basis.

The framing in Linklaters' own announcement is careful: it is described as an AI-enabled legal service, not a product, keeping it within the firm's existing regulatory structure. But the economic model is different from any legal service that came before it. Clients are paying Linklaters to build AI infrastructure for them. The deliverable is software. The price is not a per-hour rate; it is a project fee. The work is not being done by a partner billing at £1,200 per hour. It is being done by a team designing something that, once built, runs at near-zero marginal cost.

The structural question

What does a law firm do when its most profitable work — the high-complexity, high-judgment advisory — is still protected by expertise and relationships, but its volume work can be delivered by agents at a fraction of the cost? Applied Intelligence is Linklaters' answer: sell the agent layer as a premium service, not just the advice it produces. Whether that model survives long-term depends on whether clients conclude that a law firm is the right vendor for AI infrastructure. Many will not. But it is a rational first move.

05 — The regulatory battlefield

The federal government picks a side in the AI regulation fight

The DOJ's intervention in the Colorado AI Act lawsuit is the most significant US AI regulatory event in months — not because of Colorado specifically, but because of what it signals about how the federal preemption fight will actually be conducted.

The Colorado AI Act (SB 24-205) was the first comprehensive US state AI law — passed in 2024, originally effective June 30, 2026. It imposes duties on developers and deployers of high-risk AI systems to avoid algorithmic discrimination in consequential decisions affecting Coloradans: education, employment, financial services, healthcare, housing, insurance, and — directly relevant here — legal services.

xAI filed a lawsuit seeking to invalidate the Act in April 2026. The Department of Justice intervened on April 28, supporting the challenge. The White House's December 2025 executive order had already established an "AI Litigation Task Force" specifically to challenge state AI laws inconsistent with federal AI policy. This is that task force acting.

At the same time, inside Colorado's own legislature, SB 189 moved through committee proposing to drop the requirement that companies explain how their AI works and push the effective date to January 2027. As of this week, the Colorado AI Act is simultaneously under attack in a federal courthouse and being rewritten by the legislators who passed it.

Jurisdiction Measure Status as of 8 May 2026
Colorado SB 24-205 (Colorado AI Act) Original June 30, 2026 effective date. SB 189 proposes January 2027. DOJ litigation in parallel.
European Union AI Act Annex III high-risk systems August 2, 2026 deadline remains in force. Follow-up omnibus trilogue pending (~13 May).
Connecticut Comprehensive AI regulation Failed. Governor veto threat ended the session without passage.
New York RAISE Act (as amended March 27) In force. Transparency-led, not risk-classification. Reporting framework only.
Federal (US) AI Litigation Task Force Active. DOJ intervention in Colorado Act challenge is first direct action.

What this means for enterprise legal teams

The Troutman state AI law tracker for May 4 catalogued another cluster of new proposals across multiple states. The net result of this week is that the US AI regulatory landscape is simultaneously more active and less clear than it was a month ago. The federal preemption play — if successful in Colorado — would simplify the landscape significantly, but it will take years to resolve in the courts. Enterprise legal teams deploying AI are planning against a target that is moving in three directions at once: federal preemption, state legislation, and international obligations.

The compliance burden is, in itself, a kind of market signal. Every hour a legal team spends mapping AI obligations across 50 state regimes and the EU AI Act is an hour not spent on work that creates value. That cost — borne by expensive in-house lawyers doing work that is, structurally, research and mapping rather than legal judgment — is exactly the kind of work agents should be doing.

06 — So what

What this week tells us

This week's stories describe a market that has moved past the question of whether AI agents can do legal work and is now engaged in three consequential follow-on questions: how do we measure it, who owns the infrastructure, and what happens when governments start fighting each other about the rules?

The benchmark sets the competition surface
Harvey's LAB is not just a product announcement. It is a bid to define what legal AI capability means before anyone else does. Buyers now have a shared vocabulary. Vendors will be evaluated against it. The entity that sets the standard usually wins the market that forms around it. Procurement conversations will change from next week.
The thesis is now provably global
Enter at $1.2B is the cleanest proof that the core structural problem — inexpensive work done by expensive resources — exists in every legal system with high-volume, rule-based litigation. Founders Fund, Sequoia and Ribbit diligenced a Brazilian mass-litigation agent and wrote the cheque. The addressable market for supervised agentic legal services is not London and New York. It is every jurisdiction where companies face repetitive legal work at scale.
The Magic Circle is in; the question is what comes next
The firmwide deployment story is largely written. Every Magic Circle firm is committed. The interesting action is now in how those firms use AI to change what they sell — not just how they work. Linklaters Applied Intelligence is the first example of a BigLaw firm turning its AI capability into a client-facing revenue line. Others will follow, or lose the premium clients who prefer to buy the infrastructure directly.
The regulatory fight just escalated
The DOJ's intervention in the Colorado AI Act challenge is the opening move in the federal preemption play the White House's December 2025 executive order signalled. If the federal government wins, the state-by-state compliance patchwork simplifies — eventually. If it loses, the compliance burden on enterprise AI deployers (including legal teams) intensifies. Neither outcome arrives quickly. The planning assumption for 2026 remains: design for the most complex regulatory environment you expect to operate in, because it will be with you for several years.
The hallucination problem is being measured, not solved
Damien Charlotin's database crossed 1,350+ globally documented AI hallucination cases this week. Courts fined lawyers $145K in Q1 alone. The benchmark Harvey published measures task completion, not hallucination prevention — those are different questions. The open problem for every legal AI vendor, and every buyer evaluating one, is the relationship between capability on structured tasks and reliability on unstructured ones. That question is not answered by a leaderboard.
The Flank view

Two of this week's stories — Harvey's LAB and Enter's unicorn round — define the same moment from different angles. The LAB defines what "good" looks like for a legal AI agent on a discrete, structured task. Enter shows what the commercial model looks like when you build a business around agents that can do those tasks at scale, on a defined category of work, for clients who need a million of them done per year.

The common thread is the work itself. Mass litigation in Brazil. Contract review in London. Procurement processing in New York. These are all the same structural problem: a large volume of inexpensive work being done by expensive resources inside a billing model that prices the resource, not the outcome. The agent fixes that by decoupling the output from the input cost. Enter proved it on lawsuits. Flank does it on contracts. The market is now validating both at the same time, with the same investors, for the same structural reason.

The benchmark matters for Flank's customers specifically because it makes the conversation easier. The question is no longer whether agents can do legal work — Enter's million cases a year, Harvey's 500+ live agents, and the LAB's 1,200-task evaluation set all answer that in the affirmative. The question is whether your legal team is outsourcing to agents yet, or still paying a law firm to do the same work at ten times the cost. That is the question we are having with every GC we talk to. This week made it easier to ask.

Subscribe

The Intake

Weekly briefings on what's actually changing in legal AI — the market shifts, regulatory moves, and structural questions that matter for enterprise legal teams. Written by the Flank team.

Subscribe on Substack
Flank

Insource legal work to supervised agents

Enterprise legal teams use Flank to handle high-volume contracting end-to-end — NDAs, MSA redlines, procurement, triage. Agents that know your templates, terms, and escalation rules.

Learn more at flank.ai