Models hallucinate, misstate, summarize and truncate legal authority, then present the result with total confidence.
Even excellent attorneys struggle to see why generated text is wrong, or how to fix it, because nothing links a claim to its source.
An RFE response is one shot. A single unsupported citation can sink a petition and a person's status in the country.
A Request for Evidence is USCIS's demand for additional proof before it decides a petition. The response is the attorney's one opportunity to save the case.
USCIS issued 61,505 RFEs on Form I-129 alone in FY2025 across 572,582 completions. Standards vary by service center, officer and administration, so what wins cannot be read off the policy manual. It can only be learned from outcomes.
Source: USCIS, Form I-129 Petition for a Nonimmigrant Worker, quarterly Request for Evidence workbooks FY2025 Q4 and FY2026 Q2. DHS/USCIS Office of Performance and Quality.
The beneficiary qualifies as a specialty occupation worker because the position requires theoretical and practical application of a body of highly specialized knowledge, and attainment of a bachelor's degree or higher in the specific specialty is normally the minimum requirement for entry. Petitioner has submitted evidence that parallel positions among similar organizations require a degree in the specific specialty.
Verification scores citation recall × precision, so the system is penalized for over-citing as well as under-citing. In head-to-head testing our verification layer outperformed Westlaw Quick Check.
| Capability | RFE Responder | Harvey · Visalaw · Parley · general LLMs |
|---|---|---|
| Training signal | Adjudicated outcomes — approved vs denied | Work product and expert-written rubrics |
| Source of strategy | 30+ years of a practicing firm's winning filings | Public corpora and synthetic data |
| Claim verification | Split-screen proof, audit-ready package | Citation checking, no proof artifact |
| Evidence sufficiency | USCIS-specific gap identification before filing | Not adjudicator-specific |
| Who built it | Immigration attorney since 1993 + AI engineers | Engineers with legal advisors |
Harvey's own published research trains against partner-written rubrics on synthetic corpora and states it uses no customer data. No competitor is training on adjudication outcomes, because none of them has any.
Harvey research publications, 2026. We compete on data no amount of capital can purchase: consented outcome labels from a working practice.
Small sample, stated plainly. The point is not the percentage, it is that we are the only product in this category measuring approval outcomes at all.
Baseline: 85.97% of H-1B petitions receiving an RFE were approved in FY2025 (USCIS data). Beta filings to date: 25.
Circles are area-scaled to TAM. US immigration legal services $10.6B across 18,417 firms; US law firms $422.4B across 412,000 businesses (IBISWorld 2026). RFE volume from USCIS I-129 RFE workbooks.
We configure the system on the firm's own filings and run alongside their team. High trust, fast payback, and every engagement yields reusable primitives.
Standards drift with each administration, so a static playbook decays. Continuous outcome tracking is a subscription, not a one-time sale.
Attorneys contribute winning strategies and gain pooled access in return. The corpus accretes only through use, so it cannot be bought by a late entrant.
Services are the on-ramp, not the destination. Harvey hand-built playbooks across 300+ customers before shipping Playbook Builder as product. We are following that pattern deliberately, with a stated target of services below 30% of revenue by Series A.
Target customers: solo immigration practitioners, small and mid-size firms, and corporate legal departments with in-house immigration teams.
"We aren't software engineers building a wrapper on an API. We are domain experts encoding 30 years of proprietary legal expertise into a problem engineers alone have failed to solve."
Jon Velie — Founder & CEO · Immigration attorney since 1993 · Member, U.S. Supreme Court Bar
Consented outcome labels from a working practice, not licensed or scraped.
OnlineVisas launched in 2016 and Visas.AI followed. This is not a first attempt at software.
Our team files these petitions and lives on these visas. We build from inside the workflow.
The hardest briefing problem in immigration law, and the one with binary outcomes. Every response filed adds a labeled result to the engine.
The same engine applied to the full petition workflow. Existing platform, existing users, so this is expansion into an installed base rather than a new market.
Outcome-trained drafting with Under the Hood verification across other bodies of law, plus integration services for firms adopting AI.
Variance is the normal condition of legal practice. Because standards differ by adjudicator and shift over time, observed outcomes at volume is the general-purpose answer, and it travels beyond immigration.
To reach $300–400K ARR and 10–20 onboarded firms within four quarters.
Launch RFEResponder.ai. Direct outbound to immigration firms. Absorb design-partner feedback.
Expand engineering. Add enterprise sales capability and deployment staffing.
Deploy drafting into Visas.ai. Begin LegalEngine.ai for litigation and adjacent practice areas.
Onboard 10–20 firms. Surpass $300–400K ARR with services below target as a share of revenue.
Allocation is illustrative — confirm against your model before sending.