Treating AI as the interface, not a feature

Most products that “add AI” bolt a chat box onto an existing form-driven app. The form is still the product; the assistant is a novelty in the corner. When I built myAevum, I started from the opposite assumption: the assistant is the product, and the structured data behind it exists to make the assistant trustworthy.

This is a writeup of the reasoning, not the source. The goal is to show how I think about a hard problem — grounding a language model on sensitive legal and financial information — without handing over the implementation.

The problem with a chatbot on top of a vault

Estate planning is a task people avoid because it is open-ended and emotionally heavy. A blank form makes that worse. An assistant that knows what you have already entered, notices what is missing, and asks the next useful question turns an overwhelming task into a guided one.

But the moment an assistant speaks confidently about someone’s beneficiaries or account details, the cost of a hallucination changes completely. A wrong answer here is not an annoyance — it is a liability.

Grounding, then guardrails

Two design decisions carry most of the weight:

First, the assistant is grounded in the user’s actual vault on every query. Their real data is assembled into structured context, so the model answers from what exists rather than from what it can plausibly generate.

Second, the model is constrained from inventing financial or legal specifics. Grounding tells it what is true; the guardrails stop it from filling gaps with confident fiction when the data is absent. The honest answer to “what’s my policy number?” when none is on file is “that isn’t recorded yet” — not a number that looks right.

Two modes, two emotional registers

The same assistant serves two very different people. For the vault owner, it is a proactive coach — surfacing gaps, nudging toward completeness. For a family member at the handoff moment, it is something else entirely: calm, slow, and emotionally calibrated to someone who has just lost a person and now has to navigate paperwork.

Same grounding, same data, deliberately different tone. Designing that second mode was less an engineering problem than a writing one.

What I’d tell another builder

If you are putting an LLM in front of consequential data, spend your time on grounding and refusal behavior before you spend it on anything clever. The magic users feel is not the model’s fluency — it is that it never makes something up about their life.