AI Systems Design · Multi-Agent Architecture

AI-Assisted M&A Origination System

A multi-agent AI workflow that turns each private-equity client's investment thesis into a tailored, scored list of acquisition targets — and remembers every company it touches.

Role AI Systems Designer / Architect
Timeline Ongoing (2026)
Domain M&A / Private Equity

Overview

I'm designing and building an AI-assisted origination system for an M&A advisory firm — a boutique that finds privately-held companies open to being sold and matches them to the right private-equity buyer. Origination is the top of the deal funnel, and it has traditionally been slow, memory-dependent, and hard to repeat: every campaign started from a blank page.

The system is a coordinated team of specialized AI agents built on top of a durable data layer. Agents do the heavy lifting — research, list-building, scoring, formatting — but the design philosophy is deliberate: the AI produces drafts, humans make the decisions, and the compounding value lives in the data the firm accumulates, not in any single generation.

"AI as a tool, not a crutch — every agent output is a draft a human reviews, and the firm's durable asset is the memory it builds, not the model that runs it."

The Problem

Challenge

Boutique origination runs on human memory and one-off effort. A client explains what they want on a call; an analyst holds that thesis in their head and builds a list keyed to a single theme — usually just "industry." Companies already owned or already contacted slip back into lists. When a company finally replies weeks later, someone has to remember which client it was ever meant for. Nothing compounds — the next campaign begins from zero.

The specific failure points I set out to solve:

The Solution: A Data Layer With an AI Workforce

The spine of the system is not the agents — it's two living data stores that give the firm memory. Everything else reads from and writes back into them:

The agents are the workforce that operate on those stores. Tooling can change; the theses and the tracker come with the firm. On top of that spine, I designed three core AI workflows.

Workflow 1 — Deep Research Pipeline

Company & Industry Intelligence

Research → Verify → Edit → Publish

Before the firm approaches a company or briefs a client on a sector, it needs research it can trust. I built this as a four-stage relay of specialized agents rather than a single "go research this" prompt — because a lone model produces confident, plausible, and sometimes wrong output. Separating generation from verification is what makes the result defensible.

Ultra-Researcher The Verifier The Editor The Publisher
  • Ultra-Researcher runs deep, multi-source investigation across the web — financials, M&A history, market landscape — into a structured draft.
  • The Verifier is an adversarial pass: it re-checks claims against sources and flags anything unsupported, so findings aren't just plausible, they're substantiated.
  • The Editor tightens the verified draft into a clean, decision-ready narrative.
  • The Publisher formats the final report for the client or internal use.

The output feeds both sides of the system: sector intelligence sharpens a client's thesis, and company-level research informs how a target is approached.

Workflow 2 — Living Client Thesis & List Generation

Source of Truth · Client Repository

Thesis Architect → List Builder

This is the source-of-truth workflow — everything else feeds from it. Instead of a thesis living in an analyst's head, the Thesis Architect turns raw client material — intake-call notes, their portfolio, their website, quarterly updates — into a structured, living thesis document. Two design choices matter here:

  • Inferences are tagged apart from stated facts. The agent never lets an assumption masquerade as something the client actually said, and it flags open questions to raise on the next call.
  • It maintains, not just creates. Each quarterly catch-up updates the same living document with a dated change log — so the thesis is always current and its evolution is auditable.

The List Builder then reads that thesis and translates its firmographic criteria — business model, revenue band, geography, ownership type, seller-intent signals — into a ready-to-run search against a private-company intelligence platform. Two deliberate behaviors:

  • It scopes multi-sector theses. Rather than a broad, unfocused blast, it recommends one focused slice (starting with the client's active platforms) so effort lands where it converts.
  • It excludes already-owned companies at search time — pointing the query at the firm's "already in CRM / already contacted" lists — so a company the firm already has is never re-surfaced.

A key architectural nuance: the List Builder keeps metrics the data platform can't filter (EBITDA, margin, growth) out of the search entirely, deferring them to the scoring stage — matching each tool to what it's actually good at.

Workflow 3 — Company-to-Client Matching

Scoring & Attribution

The Matcher — bidirectional fit scoring

The Matcher scores a company against a client thesis and returns tiers — Strong / Possible / Weak / Review / Excluded — each with a plain-English rationale. It's the judgment layer that turns a raw candidate list into a ranked, defensible one. Two things make it more than a similarity score:

  • It runs in both directions. One thesis → a whole list (build a campaign), and one company → every client thesis (a new company appears, or a reply lands — who is this actually for?). That second direction is what powers the firm's attribution memory.
  • Fit and confidence are separate axes. Two-stage scoring applies confident-only gates first, then an ordinal weighted score. Thin data lowers confidence without faking a low fit, and an uncertain signal is flagged for human review — never silently dropped.

Scored companies flow to a CSV export agent that maps each tier to a 0–100 match score, attaches the verified owner contact, and produces an import-ready file for the firm's CRM — with every score and every blank field explained, and nothing fabricated.

How It Fits Together

End to end, a client thesis becomes an import-ready, scored target list — with a human review gate at every step, so no expensive action happens without approval.

Client Thesis

Thesis Architect turns client material into a structured living thesis.

AI Draft

Draft the Search

List Builder translates the thesis into a firmographic query, excluding already-owned companies.

AI Draft

Pull Candidates

The human orchestrator runs the search and pulls firmographics + verified owner contacts.

Human-Run

Score the Fit

The Matcher tiers each company against the thesis with rationale.

AI Draft

Export

0–100 match scores + owner contacts become an import-ready CRM file.

AI Draft

Record & Close the Loop

Every company lands in the Target Tracker and the exclusion list, so the next build skips it.

Human-Approved

Architecture & Design Decisions

The interesting problems in this build were less about prompting and more about systems design:

Impact

In one worked campaign, the system turned a single client's add-on thesis into 25 fully-scored acquisition targets — each tiered, rationale-attached, paired with a verified owner contact, and formatted for direct CRM import. What used to be a multi-day analyst grind became a reviewed, repeatable pipeline.

More importantly, the firm now compounds. Every company it touches is remembered, attributed, and automatically excluded from future searches — so each campaign starts smarter than the last.

Old Process This System
List basis One theme — usually just "industry" The client's full thesis, scored and tiered
Already-owned Caught by memory, ad hoc Excluded automatically at search time
Attribution Recalled from memory when a reply lands Recorded up front, instantly reverse-lookupable
Contacts Manual hunt, deal by deal Owner name + verified email pulled in-flow
Memory Every campaign starts from zero A compounding, versioned data asset

Tech & Approach

Claude Code Multi-Agent Orchestration Custom Sub-Agents MCP Connectors Markdown + Git Data Layer Human-in-the-Loop Design Prompt / Context Engineering Schema-Driven Contracts

What This Project Sharpened

Roadmap

The core origination pipeline is built and producing real deliverables. What's next:

Let's build something awesome!

JamesJTeeling@gmail.com
Dallas, Texas, USA