Harvey AI's Token Burn Hit 12 Trillion a Month. Rescript AI Is Building the Regulatory Layer Harvey Doesn't Cover.

Why Washington DC, Why Now

While San Francisco and New York dominate AI headlines, a cluster of startups building AI for legal and policy teams has been growing in the capital, and Rescript AI sits at the center of it.

Founded in 2023 by Nikhil Ajjarapu and Alexander Pesendorfer, both Princeton graduates with backgrounds spanning Salesforce's Public Sector team and the Air Force's Cloud Computing Group, Rescript set up headquarters at 99 M St SE in Washington, DC. The company is small (two employees listed on LinkedIn), but its investor list is not. Y Combinator, General Catalyst, and Floodgate all backed the company, which has raised $500K in a seed round, according to Dealroom.co and VCBacked. Its customer base includes Bloomberg Government Top-10 Lobbying Firms and Fortune 100 compliance teams, per the company's own site.

Rescript isn't alone. LinkedIn's "similar pages" for the company reads like a directory of the emerging DC regulatory-AI scene: Delve, building AI for public affairs from Washington, DC; Tiga AI operating out of Reston, Virginia; RegScale in Tysons Corner; Ressio Software and Prefect both in DC proper; and State Affairs, which tracks state-level legislation. These aren't general-purpose AI companies chasing consumer chatbots. They're building tools for a specific problem: the volume of regulatory and legislative change in the US is enormous, and legacy monitoring systems can't keep up.

The job market reflects it. Indeed lists 2,080 artificial intelligence roles in Washington, DC, and a separate search for legal AI jobs in the city returns 5,121 postings. LinkedIn tracks 113 roles specifically categorized as AI regulatory legal work in DC. Those numbers map onto a city where the federal regulatory apparatus generates a constant stream of new rules, proposed legislation, public comments, and hearing transcripts that someone, somewhere, needs to read and act on.

Rescript's own hiring signals what this niche demands. A "Founding Engineer" role listed on Y Combinator's Work at a Startup board offers $130K to $190K plus 0.75%–2.75% equity and requires US citizenship or visa eligibility, a constraint that likely reflects the sensitivity of the regulatory data the platform processes. Ajjarapu's LinkedIn post from mid-2024 advertising a software engineering intern position drew genuine interest, with commenters calling it a "rare opportunity" and noting they'd already applied. For a two-person company, that kind of response says something about the talent pool's appetite for this work.

The DC regulatory-AI ecosystem is still early. Most of these companies are pre-product-market fit or just past it. But the direction is clear: the next wave of enterprise AI adoption won't look like a chatbot. It will look like an agent that reads every bill introduced in all 50 state legislatures, flags the three that affect your client, and drafts a summary memo before your competitor's team has finished their morning coffee.

Inside Rescript's Platform

Rescript's core pitch is simple: an AI system that researches, tracks, and organizes laws and public policy across jurisdictions in one workspace. The architecture underneath that pitch is where the engineering challenge lives — and where the signal for hiring demand starts to show.

The platform covers eight source types: web results, statutes, regulations, executive orders, legislation, proposed regulations, public comments, and hearings. That breadth alone isn't the differentiator. Legacy tools like Quorum, FiscalNote, and Bloomberg Government already aggregate legislative data. What Rescript says it changes is the layer on top, moving from keyword-matching and Boolean filters to AI agents that reason about relevance, summarize dense material, and generate cited work products.

The workflow breaks into three stages. Research: users ask natural-language questions ("What regulations govern PBM transparency?") and get answers grounded in documents, citations, and jurisdictional context. Monitor: the system tracks changes across legislation, hearings, rules, and comment windows, then pushes alerts configured by client or issue area. Deliver: the platform produces 50-state surveys, memos, and reports that go straight to stakeholders.

The hearing coverage is the most technically ambitious piece. Rescript transcribes live testimony as it happens, identifies speakers, and turns the record into a post-hearing memo with no human transcript required. Testimonials on the site, attributed to a policy director at a top-30 lobbying firm and a partner at a top-10 lobbying firm, say hearing coverage time dropped by half and that both firms expanded legislative coverage 40% in the first year without adding headcount.

For public comments, the system clusters themes, maps stakeholder positions, and drafts analysis memos. One example on the site shows 675 comments on a single CMS proposed rule, grouped into hospitals, patient advocates, and manufacturers with their respective positions summarized. That's a task that normally eats days of associate-level work.

Rescript's site claims more than 20,000 regulatory changes analyzed per month, responses to regulatory changes at triple previous speed, and over 1,500 hours saved per customer yearly.

The reason this matters for the talent market is what the platform demands from the engineers building it. Regulatory reasoning isn't a retrieval problem. A keyword search can find a bill that mentions "prior authorization." But determining whether that bill's language actually changes a client's compliance obligations in a specific state, citing the relevant statutory authority, and flagging the downstream impact on an existing rule requires an agent that can chain across documents, understand jurisdiction hierarchy, and produce auditable output. That's a fundamentally different engineering challenge than building a chatbot, and it explains why Rescript's hiring looks the way it does.

The Three Engineering Layers Behind Agentic Compliance

Building a production-grade regulatory agent — one that tracks federal rulemaking, parses Federal Register notices, and reasons about compliance obligations — demands a stack that most generic AI roles never touch. Regulatory text is unstructured, jurisdictionally fragmented, and changes constantly. An agent has to do more than retrieve a document. It has to understand that a proposed EPA rule published on Tuesday modifies a compliance deadline that a healthcare client's legal team was tracking under an older OMB circular, and then surface that conflict with context about what it means for the client's operations.

That requires three distinct engineering layers working together.

The first is the data-ingestion and NLP pipeline. Engineers on this layer build systems that scrape, parse, and normalize regulatory text from sources like the Federal Register, congress.gov, state legislative databases, and agency guidance documents. Open-source projects in this space, like the LangGraph- and Firecrawl-based scraping agent on GitHub, show the general architecture: intelligent crawling that handles anti-bot detection, structured-data parsing that converts HTML into machine-readable formats, and entity extraction that identifies agencies, rule citations, and compliance deadlines. Rescript's engineers need to do this at scale across hundreds of federal and state sources, many with inconsistent formatting and no API.

The second layer is the reasoning and retrieval system, where most legal-ai startups fail or succeed. A keyword-matching tool can flag that a document mentions "HIPAA." An agent has to determine whether the specific provision applies to a covered entity versus a business associate, whether the comment period is still open, and whether the proposed change conflicts with existing state-level regulations. That means the engineering team needs experience with retrieval-augmented generation (RAG), vector databases, and prompt engineering tuned for legal and regulatory reasoning, not chatbot fluency but precision and traceability. Every output needs grounding in a specific source document, because a compliance officer acting on a hallucinated summary carries real liability.

The third layer is agent orchestration and product infrastructure. Regulatory monitoring isn't a single query; it's a persistent workflow. An agent has to run on a schedule, compare new documents against a client's tracked topics, escalate when a threshold is met, and integrate with the tools legal teams already use. That demands backend engineers comfortable with event-driven architectures, cloud infrastructure (AWS or GCP), and API design. It also demands enough product sense to build interfaces that compliance officers, not ML researchers, will actually use day to day.

Harvey AI, Rescript's closest legal-ai parallel, illustrates the salary pressure this demand creates.

The broader implication is that DC's AI talent market is bifurcating. Defense and intelligence contractors, Booz Allen Hamilton, Johns Hopkins APL, Leidos, and Accenture Federal Services, are hiring hundreds of AI engineers for classified and government-facing work, as LinkedIn job listings confirm. Rescript and the legal-ai cohort are competing for the same Python-fluent, LLM-experienced engineers, but offering a different pitch: build products at the intersection of law and machine reasoning, inside a commercial startup rather than a government contracting pipeline. The engineers choosing that path tend to want their work to ship as a product, not disappear into a procurement cycle.

Legal AI Crosses Into Production Infrastructure

Harvey AI's monthly token consumption jumped from 1 trillion in January to an estimated 12–13 trillion by May 2026, CEO Winston Weinberg said on the Sourcery podcast, as reported by Business Insider. That 12x surge in under six months isn't a vanity metric. It signals that legal AI has crossed from pilot-phase curiosity into production-grade infrastructure, the kind of embedded, daily-use tool that firms can't operate without.

The numbers behind the tokens are stark. Business Insider reported that Harvey users run more than 700,000 agent-powered tasks a day, with hours spent in the product per user rising 75% over four months. The platform serves more than 100,000 lawyers across 1,500 law firms and enterprises in more than 60 countries, per The Wall Street Journal. Harvey hit $100 million in annual recurring revenue in August 2025 and was on pace for roughly $300 million ARR by mid-2026, with a valuation that climbed from $3 billion in February 2025 to $8 billion in October 2025, and then to $11 billion in a subsequent round, Business Insider reported in March 2026.

This growth trajectory mirrors what enterprise software looked like in the early cloud era, except the adoption curve is compressed. Reid Hoffman, a Greylock partner and Microsoft board member, put it bluntly on The Prof G Pod: private AI companies are reaching top-25 global valuations before IPO. OpenAI's most recent round valued it above $800 billion. Anthropic went from $9 billion to $30 billion in annualized revenue in roughly five months. "Most of the companies in the SaaS universe don't even have $2.5 billion of ARR annually," Hoffman said.

What makes Harvey's position distinct is the shift from chatbot to agent. In May 2026, the company rolled out roughly 500 AI agents for legal work and a redesigned Agent Builder that lets lawyers create custom agents without writing code, per Business Insider. Co-founder Gabe Pereyra, drawing on his background at Google Brain and DeepMind, has been open about the cost implications: a single document draft query runs about $20 in tokens, while a full contract review across 100,000 contracts can cost $20,000, he said on the same podcast. He framed the coming challenge in terms every general counsel will recognize: "I just spent $1 billion on tokens. Where's my ROI?"

That question, how to prove return on investment per token, is the same problem law firms have faced for decades with billable hours. Weinberg has drawn the parallel explicitly, calling it the "billable hours problem" for the AI era. The firms and platforms that solve for transparent, task-level cost attribution will own the next layer of enterprise legal infrastructure.

The competitive dynamics are sharpening. Harvey's primary competitors are not other legal AI startups but the frontier model labs themselves, OpenAI and Anthropic, both building legal-specific products, Weinberg said on the Sourcery podcast. Against European rival Lorraa, Harvey claims a win rate above 70%. Meanwhile, the company open-sourced its Legal Agent Benchmark (LAB), covering 24 practice areas, to position itself as a neutral routing layer across multiple model providers. The strategic logic: law firms face conflict risk if they rely on a single provider, and no single model is best for every legal task.

This is where Rescript AI's niche fits. While Harvey targets the law firm and corporate legal department market, Rescript is building regulatory and legislative tracking agents, a compliance-adjacent layer that sits between government action and enterprise response. The Y Combinator backing signals that investors see regulatory tracking as a distinct infrastructure play, not a feature bolted onto a legal research tool. As the EU AI Act, DORA, and a growing patchwork of US state-level AI regulations force companies to monitor compliance in real time, demand for specialized regulatory agents is tracking the same exponential curve that Harvey's token consumption revealed for legal work.

The enterprise AI compliance market is no longer theoretical. It's where the tokens, and the hiring, are going.

What Rescript's DC Hiring Signals About the Agent-Engineering Talent Market

Rescript's decision to build its engineering team in Washington DC, not Palo Alto or San Francisco, is a data point that says more about where agent-engineering talent is heading than any job-posting survey. It raised that capital in 2024 from the same three investors, and it chose the capital region as the base for building production-grade regulatory agents. That choice reflects a broader shift: the next wave of enterprise AI hiring is pulling engineers toward domain-specific hubs, not just the traditional coastal labs.

The White House Council of Economic Advisers documented the scale of the mismatch. Between 2015 and 2022, job postings requiring AI software skills grew at an average annual rate of 31.7 percent. Over the same period, AI-relevant bachelor's degrees awarded by U.S. institutions grew at just 8.2 percent annually, master's at 8.5 percent, and PhDs at 2.9 percent. Demand is outpacing supply by roughly four to one at the undergraduate level. Rescript's hiring, for engineers who can build agents that reason over regulatory text and not just retrieve keywords, sits squarely in that gap.

What makes the DC angle significant is the talent pool it draws from. The region already houses a dense concentration of policy analysts, compliance professionals, and government-adjacent technical staff. Talent Capital, a regional workforce initiative housed by the Metropolitan Washington Council of Governments, now operates an AI-powered career-matching agent called Celeste that connects DC-MD-VA job seekers with training and roles across the area. The initiative partners with employers, universities, and nonprofits across all three jurisdictions. This kind of infrastructure, a regional pipeline that blends domain expertise with engineering skill, is what agent-building companies need and what pure tech hubs lack.

The skill set Rescript is hiring for also tells you where the market is moving. Building regulatory agents demands more than fine-tuning a large language model. Engineers need to work with structured legal ontologies, handle multi-jurisdictional rule sets, and build systems that can explain their reasoning to compliance officers who will be liable for the output. That's closer to the "Forward-Deployed AI Engineer" role described in recent hiring analyses, client-facing engineers who integrate agents into enterprise data systems, security environments, and custom workflows, than it is to a standard ML engineering position.

Harvey AI's salary bands illustrate the compensation pressure. Senior backend engineering roles at Harvey top out at $290,000, and AI Automation Engineer positions range from $123,600 to $185,400. That's the price point for engineers who can ship agent systems in regulated environments, and it's pulling compensation benchmarks upward across the legal-tech sector.

The international dimension adds pressure. China now produces nearly double the number of science and engineering PhDs the United States does, and the gap is widening. In 2018 China produced 65 percent more; by 2022, 99 percent more, per the CEA report. The U.S. still leads in top AI researcher output and houses most frontier labs, but the report is explicit: that lead "should not be taken for granted." For companies like Rescript, the implication is clear. Hiring domestically in a hub that combines regulatory domain density with engineering talent isn't a preference — it's a strategy for staying ahead of a global talent curve that's accelerating without them.

The takeaway for engineers watching this space: the premium is shifting from general AI fluency to domain-embedded agent-building skill. Rescript's DC hiring, Harvey AI's salary bands, and the regional workforce infrastructure rising around both of them point in the same direction. The next career move isn't toward the lab that trains the biggest model. It's toward the team that deploys the most reliable one in the room where compliance decisions get made.

Working in AI? Zero G Talent tracks the openings: browse AI jobs, openings at Harvey AI, and the people building the field.