How We'd Automate 30%-Ruling Packet Prep for a Dutch Expat Tax Firm

The workflow as it typically runs

Note: the description below reflects how this workflow is commonly structured in professional services firms handling 30%-ruling applications at volume. We are describing a genre-level pattern, not asserting knowledge of your specific process.

A typical assignee intake involves:

1. Assignee submits a document package — often via email, sometimes a shared drive, occasionally a form
2. An advisor or office manager checks completeness against an internal checklist
3. Missing items trigger a follow-up request to the assignee or employer
4. Once complete, relevant fields are extracted from the employment contract and other documents
5. The advisor reviews the eligibility criteria against the extracted data
6. A packet summary or draft application is prepared for advisor review and submission

In firms handling 20+ active cases, steps 2–5 are where delays accumulate and where junior staff time is consumed. The work in those steps is structured — it follows a defined logic — but it is currently done manually.

Where the method applies

The Orellis methodology begins with an operations map: every repeating task that takes more than 30 minutes per week, scored by frequency × pain × AI-reducibility.

In this workflow, the tasks that score high on AI-reducibility share three characteristics:

• They follow a defined decision tree. "Is document X present? If yes, proceed. If no, trigger follow-up." This is not judgment — it is a conditional check.
• They extract from structured or semi-structured inputs. Employment contracts follow recognizable formats. Salary, start date, and address fields can be extracted reliably from a well-prompted AI system with a defined output schema.
• They produce a draft, not a decision. The advisor still reviews. The AI produces the first version of the completeness check, the field extraction summary, and the packet draft — the advisor corrects and approves before anything goes to a client or the Belastingdienst.

The automation boundary we would draw

This is the most important decision in the engagement — not what to automate, but where to stop.

AI handles:

• Document completeness check against a defined list
• Field extraction from employment contracts (salary, start date, role, prior address)
• 150km rule flag (comparison against a defined threshold, not a legal opinion)
• Draft packet summary for advisor review
• Draft client status update for advisor review and send

Advisor retains:

• Edge case interpretation (anything that doesn't fit the standard pattern)
• The eligibility judgment call
• Sign-off on every client-facing output
• All communication with the Belastingdienst

This boundary is not arbitrary — it reflects both where AI is reliable and where the EU AI Act requires documented human oversight for high-risk outputs in a regulated context.

What the build looks like

Orellis builds managed agents via Anthropic's Claude API on infrastructure designed for EU data residency (no US data transfer required for the document content). The system runs against your intake data — PDFs submitted by assignees, employment contracts, form outputs. Every prompt and output is logged with a timestamp and operator ID.

The operator — your designated advisor or office manager — runs the system. Orellis builds it and hands it over. The runbook covers: how to run it, how to stop it, what to do when it returns an unexpected output, when to review the system prompt.

Build is scoped during the engagement, not before. We don't quote a fixed build fee until we've walked through your current intake process.

What this teardown is not claiming

• We are not claiming to have delivered this for a client. We are describing how we would approach it.
• We are not quoting time savings or cost reductions. Those depend on your current intake volume, your document submission quality, and your advisor's review speed — none of which we know until we've done the operations map.
• We are not claiming the system removes compliance risk. It adds a documented human-in-the-loop layer; it does not eliminate the advisor's professional responsibility.

If you want numbers, the honest path is: run the audit, map the workflow, and measure. We can scope that conversation for free.