Legal AI · Due Diligence · In development

MyAdvisor.ai Due Diligence

The due-diligence module — built to produce evidence a lawyer can check, not answers they have to trust.

Role: Founder · Architecture · Engineering
Category: Legal AI · Due Diligence
Status: In development

The problem

Due diligence is a discovery exercise under time pressure: hundreds of documents, a small number of things that actually matter, and a deadline. It is exactly the shape of problem AI looks good at — and exactly the shape where a confident wrong answer does the most damage, because nobody has time to re-check the ones that looked fine.

Approach

Build the module to surface evidence rather than conclusions. Every finding points back to the document and passage it came from, carries the date context it was true as of, and stays inside the tenant that owns the documents. The reviewer's job stays the same; the search cost drops.

What it does

Findings tied to documents: Every output traces to the source passage, so a reviewer verifies rather than trusts.
As-of-date awareness: A document that was accurate two years ago is not a current fact. The module carries that distinction rather than flattening everything into the present tense.
Tenant isolation: Due-diligence material is among the most sensitive a firm handles. Isolation is an architectural property, not a filter applied at query time.
Structured over narrative: Output is organised so it can be checked item by item, rather than delivered as prose that reads well and hides its gaps.

The hard parts

Absence is a finding: In due diligence, what is missing often matters more than what is present. That is much harder to detect than to summarise, and easy to report falsely.
Evidence beats fluency: The temptation is to produce a readable summary. The useful output is a traceable one, even when it reads less impressively.

Detail

Why this module is the hard one

The rest of a legal AI product can lean on retrieval quality. Due diligence adds a second problem: the reviewer is looking for absences, inconsistencies and things that should have been in the data room and are not. A system optimised to answer questions is poorly shaped for finding what nobody asked about.

What it does not claim

I make no legal conclusions and no assessment of risk. The module organises evidence and shortens the search. The judgement stays with the professional, which is both the correct division of labour and the only defensible one.

Stack

Analysis: Document ingestionStructured extractionSource attribution
Controls: As-of-date awarenessTenant isolationAudit trail

Case studies

The decisions behind the build

CS-01Building AI due diligence around evidence instead of answers CS-02Building legal AI that refuses unsupported claims

OPINION"No source, no claim" should be the default for high-stakes AI AI NOTEAI agents are easy to demo. Production agents are different.SYSTEM NOTEThe architecture behind a source-grounded legal AI assistant

All work