← Back
Diligence AI
The Ask-AI functions inside a deal's diligence checklist. Given the deal's registry facts, memory, category summaries, uploaded documents, and internal notes, Claude suggests an answer for every open diligence item — one at a time or in bulk — with structured source references. The banker approves, edits, or rejects.
The Question Behind the Feature
A diligence checklist is 200-400 questions. Each answer is somewhere in the data room, or in notes from prior meetings, or in what the banker remembers about the deal. The tool uses vectorized chunks, thoughtful searching from previously answered questions, and direct context.
Input Sources
+----------------+ +----------------+ +----------------+
| | | | | |
| DEAL CONTEXT | | DOCUMENTS | | TRANSCRIPTS |
| | | | | |
| registry facts | | VDR files | | Teams meetings |
| memory (vers.) | | pgvector | | pulled from |
| per-category | | chunks indexed | | HubSpot; also |
| summaries | | by deal + tag | | vector-indexed |
+----------------+ +----------------+ +----------------+
| | |
+---------------------+---------------------+
|
v
+---------------------+
| |
| CLAUDE ANSWER |
| SUGGESTION |
| |
| per-item suggest |
| bulk sweep ~30 at |
| a time |
| source refs |
+---------------------+
|
v
+---------------------+
| DILIGENCE ITEMS |
| suggested + refs |
| banker reviews |
| approve / edit / |
| reject |
+---------------------+
Data Flow
+--------------+ +--------------+ +--------------+ +--------------+
| | | | | | | |
| CONTEXT |----+| RETRIEVE |----+| ANSWER |----+| REVIEW |
| | | | | | | |
+--------------+ +--------------+ +--------------+ +--------------+
| | | |
v v v v
+--------------+ +--------------+ +--------------+ +--------------+
| build deal | | vector | | Claude | | banker sees |
| profile from | | search over | | drafts | | suggestion + |
| registry + | | doc chunks + | | answer with | | source docs; |
| memory + | | transcript | | source refs; | | approves, |
| category | | search per | | bulk sweeps | | edits, or |
| summaries | | question | | ~30 items | | rejects |
+--------------+ +--------------+ +--------------+ +--------------+
Process Steps
01
Build deal context
The registry has the facts. Memory has the notes. Per-category summaries condense related documents. Together they form a compact context Claude reasons over without re-reading everything.
02
Retrieve per question
Vector search over document chunks and transcript segments returns candidate sources for each diligence item. Each item pulls only what's relevant to it.
03
Suggest an answer
Claude drafts a proposed answer with structured source references — document ID, page number, or transcript timestamp — so the banker can trace every claim.
04
Bulk sweep the list
One click runs suggestions across every open item, batched ~30 at a time. Progress streams to the live activity feed; the banker doesn't watch it happen.
05
Banker reviews and approves
Suggestions are proposals, not answers. Every one is manually accepted, edited, or rejected. The record shows both what Claude suggested and what shipped — so future audits can distinguish.
Stack
Python
FastAPI
Claude API
pgvector
PostgreSQL
HubSpot API
React
Suggestion, not automation. Every answer is ultimately edited and approved — the last mile is always human.