Here's the strange fate of the American abroad: the eagle follows you. U.S. citizens are taxed on worldwide income no matter where they live, and the filing stack that lands on a teacher in Lisbon or a founder in Singapore — Form 1040 plus 2555, 1116, 8938, 5471, 8992, FBAR — is dense, interdependent, and rewritten every season.
And here's the deeper problem underneath the paperwork: people speak in stories, and the IRS accepts only boxes. A client tells their lawyer, "I moved to Portugal in March, kept my Chicago condo, my employer's in Delaware, and I've got two bank accounts in Lisbon." The IRS wants that same life as discrete typed fields across five interacting forms. For Willco Law — a firm serving exactly this population — the translation from story to boxes was consuming the practice: attorney and paralegal hours spent on repetitive intake, repeated threshold explanations, and hand-transcription onto IRS PDFs, all before any judgment-based legal work began.
Eagle Abroad is the translator Rocket Now built between the story and the boxes.
The Build: From Conversation to Filing-Ready PDF, No Re-Keying
The assistant that knows the terrain. A streaming conversational tax assistant grounded in a purpose-built expat system prompt: Forms 1040, 2555, 1116, 4868, 5471, 8858, 8938, 8992, and FinCEN 114, the Physical Presence and Bona Fide Residence tests, treaty interactions, and the current year's reference data — FEIE maximum, FBAR thresholds, expat deadlines. With guardrails a law firm requires: the assistant states plainly that it provides guidance, not tax advice, and routes complex cases to a CPA or the firm.
Live IRS grounding — because the rules move. Static tax content decays, and stale guidance is a liability issue for a law firm. So the assistant classifies which of 8 form/topic classes a question touches — matching form numbers and natural aliases like "FEIE," "FATCA," "GILTI" — and retrieves the current instructions from IRS.gov at request time, injecting them into the model's context, capped at the two most relevant forms per turn to control cost and latency. When retrieval fails, then the turn degrades gracefully to the base prompt instead of erroring. Guidance tracks the IRS itself, not a snapshot of it.
The translation layer — story becomes boxes. A separate structured-extraction service reads the full conversation transcript and returns typed tax data through a strict tool-call schema: identification, spouse details, addresses, employment, foreign earned income, foreign taxes paid, residence and travel history, foreign accounts and assets. That structured record pre-populates the form wizard — the client never re-states in fields what they already said in conversation.
The boxes themselves — real IRS PDFs. Seven forms fully modeled — 1040, 2555, 1116, 4868, 5471, 8938, 8992 — as step-by-step wizards across 28 guided sections with plain-language explanations, runtime discovery of the official PDFs' AcroForm fields, live side-by-side preview, and a download that is a filled, filing-ready IRS PDF. The same session that answered the client's questions ships the artifact. No re-keying step between advice and filing — that gap, the one that ate the firm's hours, is simply gone.
And the confidentiality a law firm's clients assume: row-level security on every table, ownership-scoped policies with message access enforced transitively through conversation ownership, server-side — with sensitive identifiers like SSNs handled client-side for PDF fill rather than expanding stored PII.
Stated Plainly: The Model Layer Today and Next
In the house style of this series, the provider position without embellishment: the AI layer runs on another provider today, behind an OpenAI-compatible abstraction — and Willco Law and Rocket Now have scoped a migration of the reasoning-critical paths to Claude, for reasons specific to this workload. Retrieved IRS instruction text is long and legally precise, and paraphrase drift is a compliance risk — long-context fidelity to supplied source material is the job. Extraction correctness is pre-fill correctness, so schema-conformant tool use is the job. And in regulated guidance, correctly declining to guess is a feature — calibrated refusal is the job. The migration plan is the disciplined kind: a fixed evaluation set dual-run across providers on fidelity, schema conformance, appropriate-refusal rate, latency, and cost; the lowest-risk surface (extraction, human-reviewed in the wizard) cut over first behind a feature flag; the assistant staged 10% → 50% → 100% with attorney spot-review at every gate; and then the net-new Claude-dependent capability — a pre-filing review pass that audits a completed form against the retrieved IRS instructions and flags threshold conflicts (FBAR vs. 8938 overlap, 2555/1116 double-benefit, 5471→8992 consistency) before the PDF ever downloads. When every cutover is flagged, gated, and attorney-reviewed, then a law firm can adopt frontier AI at the speed of evidence.
The Results
- Intake inverted: from attorney-led interviews and hand-transcription to self-service conversation whose transcript converts to form data automatically
- 7 IRS forms end-to-end, 28 guided sections, filled filing-ready PDFs — advice and artifact in one session — verified from the delivered codebase (~8,000 lines, 3 AI services in production)
- Live-grounded guidance across 8 form/topic classes, materially reducing the stale-content risk a law firm carries on public tax content
- Per-client isolation enforced at the database layer, with deliberate restraint on stored PII
- A scoped, evaluation-gated Claude migration with a genuinely new capability — the pre-filing audit — waiting on the other side
"[Quote pending Willco Law sign-off.]" — [Name, Title], Willco Law
Why It Worked: The Right Step Was the Translation Layer
The Rocket Now principle, notarized: a single step in the right direction is worth exponentially more than many steps in the wrong one. The wrong steps were the legal-tech defaults — a chatbot that answers questions and leaves the transcription problem untouched, or a forms tool that assumes the client already knows their answers. The one right step was building the translation layer — conversation to typed data to official PDF as one continuous motion — because that's where the firm's hours actually went. When the story becomes boxes automatically, then attorneys spend their time where judgment lives, and the eagle, for once, is easy to feed.
Practice drowning in intake instead of judgment? Build the translator. → RocketNow.com
Editor's note — before publishing:
- Confirm authorization to name Willco Law; replace the placeholder quote
- Usage metrics, client counts, contract value, and projected token volume are all [TBD — client] in the source docs — only codebase-verified figures appear here; keep it that way
- The current-provider-today / Claude-scoped-and-pending distinction is accurate per the docs — don't let edits promote the migration to the present tense
- "Guidance, not tax advice" framing mirrors the product's own guardrails — preserve it in any edits