Case study. Cloud Cruz is a fictional company and every number is synthetic. The program design, the systems and the decisions are my work.Fictional company, synthetic data, my real work.

Tony GilharryTalk to Tony

Case studyBlueprintDecisions

Decision brief: the intent score, Jev, and scaling the automation

A simulated leadership room (fictional role archetypes from Tony's Virtual Leadership Team method). Tony Gilharry is Driver and made the final decisions. Lens for every seat: will this save time, will this save money, is it the simple thing that works, are we moving with intent, should we think again, does it adapt to the market?

Decision

How Cloud Cruz scores partner intent, whether Jev (TypeSafe's System One model) should power it, and whether the automation scales on a flat team.

Context and evidence

  • The fit score is set (decision 29); decision 35 left the intent score to co-design.
  • Jev returns typed judgments (a choice, score or yes/no, each with a probability) from unstructured state; code owns the workflow; TypeSafe's guidance is to send uncertain cases to a person (https://github.com/typesafe-ai/skills, https://github.com/typesafe-ai/typesafe-sdk-js, both MIT). There is no public typesafe-ai/jev repository.
  • Year-1 volume: about 3,500 partner leads (about 70 a week); the team is 9 people, flat (decision 66).

DACI

  • Driver: Tony Gilharry
  • Approver: Tony Gilharry
  • Contributors: VP Global Partnerships, Forward-Deployed AI, GTM Analytics, GTM Operations, Security and Privacy, Enterprise Architecture, Finance, Partner Marketing, Partner Field
  • Informed: the partner team

Round 1: the intent score and Jev

  • Forward-Deployed AI: a points table counts events but cannot read meaning (does this firm have an AI practice, is this a deal in hand, how urgent is the reply). Typed judgments fill those gaps; composite scoring keeps Cloud Cruz's weights in code; low confidence routes to a person.
  • GTM Analytics: keep the points table as the contract; log every judgment with confidence and model version; check calibration against outcomes.
  • GTM Operations: design now, run by hand in the beta, automate in build 3 (decision 34).
  • Security and Privacy: vendor review (retention, training on inputs, processing location, DPA) before any real data; send only needed fields.
  • Enterprise Architecture: store judgments as fields in Salesforce; build a rules-only fallback.
  • Finance: verify per-decision pricing; the payoff is rep time, not cash, unless it avoids a hire.
  • VP Global Partnerships: a strong adapt-to-the-market story, but design a vendor-neutral layer with Jev as the first engine.
  • Partner Marketing: add stated urgency and timeline as a signal.

Round 2, think again: does the automation scale?

  • Forward-Deployed AI: compute is not the constraint; people are.
  • GTM Operations: at ten times Year-1 volume a fixed hot line produces about 140 same-day calls a week; route by capacity instead (top N per rep per day).
  • GTM Analytics: guard quality with a human-review rate at or under 10%, a random 5% weekly audit and weekly calibration.
  • Enterprise Architecture: event-driven, idempotent scoring written to Salesforce; batch work on the scheduler; rules-only fallback.
  • Finance: cost scales with events; set a monthly spend cap on the judgment API.
  • Security and Privacy: minimal fields and retention limits; audit a sample rather than everything.
  • Partner Field: hot leads always get a fast human call; approve.
  • VP Global Partnerships: per-send human approval is the bottleneck at scale; changing that guardrail is Tony's explicit call.

Conflicts and tradeoffs

Automation speed versus human judgment; vendor innovation versus concentration risk; the founding "a person approves every state change" guardrail versus volume at scale.

Decisions (approved by Tony Gilharry, 2026-09-25)

  • 68: the intent score (application 30, meeting 20, stated timeline 20 by typed judgment, events up to 20, visits and content up to 10; deal in hand is urgent; 30-day half-life).
  • 69: a vendor-neutral typed-decision layer with Jev as the first engine; live in build 3; conditions: rules-only fallback, logged confidence, calibration, vendor and privacy review, pricing check.
  • 70: capacity-aware routing (N = 8 per rep per day to start), event-driven scoring, tiered human in the loop, quality guards, monthly spend cap.
  • 71: a person approves every AI send in Year 1 and the demo; policy-level approval only at scale and only by a new explicit decision from Tony.

Actions and next gate

Action Owner Due Evidence
Update 03-partner-profile.md Claude Sep 25 Committed
Add partner-lead judgment fields to the Year-1 data Claude Oct 1 Data dictionary
Vendor and privacy review of TypeSafe Tony Before any real data Review notes
Prototype one typed judgment on synthetic applications Build 3 With Build 3 Demo