# Study spec: AI-agent supply-chain matchmaking - wedge selection

Simulated multi-persona Mom Test study. Up to 5 cohort-sequential rounds.
Goal: decide which wedge (if any) survives, for whom the pain is real, and
what real humans must confirm. Simulated results are hypothesis generators,
NOT validation.

## The bet (riskiest assumption)

US SMB manufacturing participants (buyers, shops, intermediaries, acquirers)
lose enough quantifiable money to bad/slow supplier-customer matching that
they will pay for software-agent-run, evidence-backed, relationship-preserving
matching, despite a 125-year graveyard of directories and a hated
broker-marketplace incumbent.

## Wedge candidates under test (from SYNTHESIS.md, do not reveal to personas)

- W1 Verified Shortlist: buyer pays $200-500 per evidence-backed supplier
  shortlist dossier (48h turnaround).
- W2 Agent-amplified trading/rep desk: agents 10x a trusted human book;
  commission on close.
- W3 Castings/forgings reshoring desk: vertical verified matching, both sides
  pay per qualified engagement.
- W4 Succession demand book: verified customer-relationship map sold to
  acquirers at diligence ($5k-25k/deal).
- W5 Commission rep in software: outcome-priced demand gen for job shops
  (5-15% of first-year revenue, zero upfront).

## Rubric (every interview, every round)

Fit strong/medium/weak/no; Pain X/5; WTP specific $ + own-pocket vs firm or
"none"+why; Channel-fit; Dollar loss (concrete $, or "none surfaced"); Owns
channel; Existing tool spend (+use% +trust); Incumbent risk; Life events
matter. Template: ~/.claude/skills/customer-discovery/assets/interview-template.md

## Mom Test non-negotiables

Past behavior only, no pitch until the final labeled "*vibe check (biased)*"
turn. Personas push back with real objections, real economics, real "no"s.
Chase a dollar figure on every pain. A no-fit is a valid result. If the
roster skews positive, the study failed.

## Personas (fixed across rounds; 3 designed weak/no-fit as control)

| # | Slug | Who | Primary wedge | Expected |
|---|---|---|---|---|
| 01 | dana-hardware-buyer | Dana Whitfield, 41, sourcing/ops lead at a 40-person consumer-hardware brand, Pacifica-adjacent DTC + retail, imports ~70% BOM from Shenzhen, tariff-squeezed since 2025, actively re-sourcing to Vietnam/Mexico. Tools: Alibaba, one bad sourcing-agent experience, spreadsheets. | W1 | in-market |
| 02 | marcus-oem-procurement | Marcus Oyelaran, 48, procurement manager at a 120-person industrial-equipment OEM (Ohio), buys castings, machined parts, weldments. ERP: Epicor. Uses ThomasNet out of habit, trusts his rep network. | W1/W3 | medium |
| 03 | rita-dtc-alibaba | Rita Falcone, 33, owner of an 8-person Amazon/DTC kitchenware brand. Sources happily on Alibaba + Accio free tier; margin-obsessed. | W1 | CONTROL weak: free tools suffice |
| 04 | tom-cnc-owner-63 | Tom Brzezinski, 63, owner of a 12-person CNC job shop (Michigan), no successor, ThomasNet-burned ($2,400/yr zero leads), Xometry-burned (jobs below material cost), 80% revenue from 3 customers he's had 20 years. Phone and trade shows only. | W5 supply / W4 subject | skeptical, real pain |
| 05 | kelly-fab-successor | Kelly Nguyen, 34, second-generation successor who took over her dad's 15-person fab/weld shop last year (Texas), digital-native, needs new customers badly, ran $500 of Google Ads that produced $35k of work. | W5 | in-market |
| 06 | hank-foundry-sales | Hank DeWitt, 55, sales manager at a 90-person gray/ductile iron foundry (Wisconsin). Reshoring inquiries up sharply but conversion is poor: wrong-fit RFQs, tire-kickers, tooling sticker shock. One regional rep covers him. | W3 supply | medium |
| 07 | priya-casting-sourcing | Priya Raghavan, 38, senior sourcing engineer at a mid-size ag-equipment OEM re-qualifying Chinese castings domestically under tariff pressure. Last search: 11 weeks, 4 foundries reached, 1 quote. | W3 buy | in-market |
| 08 | sal-manufacturers-rep | Sal Marino, 61, independent manufacturer's rep, 30 years, 6 principals, 5-15% commission, phone-native, proud Rolodex, thinks software leads are garbage. | W2 analogue / incumbent | CONTROL hostile: the human trust layer itself |
| 09 | jenny-trading-co | Jenny Park, 45, owner of a 10-person US-Asia trading company (electronics + housewares), like Phil: undisclosed markup 12-25%, drowning in qualification grunt work, passes on small deals. | W2 | in-market |
| 10 | derek-search-fund | Derek Sloan, 36, search-fund principal in diligence on a $2.1M-EBITDA machine shop; terrified the customer book walks with the 66-yo owner. | W4 | in-market |
| 11 | aisha-pe-rollup | Aisha Bello, 31, associate at a precision-machining PE roll-up platform (6 shops acquired), runs customer-retention diligence by hand. | W4 | medium |
| 12 | greg-ems-purchasing | Greg Voss, 52, purchasing manager at a 400-person contract electronics manufacturer. Approved-vendor-list + AS9100 + ERP-gated procurement; personally buys nothing without a 6-month qualification. | none cleanly | CONTROL no-fit: process-gated |

## Round arc (adapt from round 2 on via the round-designer)

R1 Open Mom Test (broad, zero pitch; light generic vibe check at end).
R2 Deepen the strongest pains; quantify dollar losses; frequency.
R3 Cadence and behavior: how they work the pain today; where tools fail.
R4 Concept and assist: reveal the RELEVANT wedge mechanics; find what breaks;
   trust/autonomy limits.
R5 Segmentation and WTP: firm numbers, buy path, hard objections.

## File layout

- discovery/round-N/questions.md (round question set; R1 committed, later
  rounds written by the round designer)
- discovery/round-N/NN-slug.md (one per interview, template format)
- discovery/round-N/SYNTHESIS.md (round synthesis, written by designer)
- discovery/CROSS-ROUND.md (final movement synthesis)