Stress-Testing Estate Plan Documents
I built an internal tool that uses language models to stress-test estate planning documents against large sets of client-specific hypothetical scenarios. The idea is to essentially borrow the software concept of unit testing and applying it to a trust or will.
The process works in four steps. First, I sit down with the client's priorities. Some priorities are universal ("avoid unnecessary tax exposure"), some specific to them (in my walkthrough example, "provide for Rose and Leo, the client's children"). Second, I generate a comprehensive set of "what-ifs." Rather than just asking a model for ten scenarios and hoping it covers the right ground, I work from a matrix where the columns are stakeholders (grantor, trustee, beneficiary, etc.) and the rows are categories of life events (family dynamics, asset changes, health situations, and so on). Each cell, such as "beneficiary gets married", becomes a prompt the model expands into client-specific hypotheticals grounded in the actual family. That structure is what gives me confidence the coverage is systematic rather than whatever happens to come to mind. Third, for every scenario, I feed the model three things: the trust document, the scenario, and one client priority, and ask it to assess how exposed the client is if that scenario occurs. Fourth, the results land in a spreadsheet I can scan for elevated-risk items worth a second look.
One important caveat on the scoring: the model returns a number between 0.0 and 1.0, but I don't treat the absolute value as meaningful — run the same scenario twice and you'll get different numbers. What stays consistent is the relative ranking within a set, so the same handful of weak spots keep surfacing across runs. That's what I act on.
Tech stack is intentionally simple: Google's Gemini for the language model work, Google Sheets for the scenario matrix on the input side and the risk-score report on the output side, and Python to orchestrate everything. Documents and client information stay on my local machine for confidentiality; only the output analysis lives in (secured) Sheets.
Full write-up, including a concrete walkthrough using a fictional "Snow White" family and trust is available at the link.
