Product Case · Performance Marketing · 0–1 AI Product Strategy
Marketers had AI that writes ads and AI that manages bids — nothing that connected the two, or knew when to ask first.
Creative fatigue burns out a winning ad in 1–3 weeks; budget drift wastes spend for days before a human notices. The point tools on the market generate creative or automate bids, but neither watches performance and proposes the next move — and marketers already distrust anything that spends money without a visible paper trail. I designed the connective layer: a copilot that routes every AI action through a confidence-based trust boundary — auto-apply, suggest, or escalate — instead of shipping another black box.
Self-directed portfolio case study — a simulated 0–1 AI product build in a performance-marketing context. Company, users, and figures below are illustrative, built to demonstrate product judgment.
WhoIn-house performance marketerTeam of 1–4 managing $50K–$500K/month, no dedicated data scientist.
WhatMake the trust boundary explicitDefine what the AI can do alone, what it must escalate, and how trust is earned over time.
WhyAccountability, not capabilityThe adoption barrier was never model quality — it was proving every dollar spent has a paper trail.
Problems
Creative fatigue outpaces production
Budget drift caught days late
Disconnected point tools
Black-box distrust from platform automation
Unmanaged LLM spend at scale
Two point tools, neither connected
Generators that don't watch
AI creative tools produce copy and images on demand, but have no view of live performance — they can't tell you a variant is fatiguing until a human goes looking.
Automation that earns no trust
Platform-native and third-party bid automation acts on budget with no visible reasoning, which is exactly what taught marketers to keep automation on a short leash.
What I designed
Every action gets one laneCreative variants and budget/bid changes are classified into exactly one of three lanes before a reviewer ever sees them.
Everything defaults to Suggest in v1Even high-confidence recommendations wait for a human click — autonomy is granted after trust is earned, not assumed.
Cost controls are v1, not v2Tiered model routing, feed-hash caching, and hard per-account generation caps ship on day one, before spend can outrun the ad budget it's managing.
Attribution stays platform-nativeNo proprietary multi-touch model in v1 — recommendations sit on top of Meta/Google's own numbers, so they reconcile with what marketers already trust.
The core mechanic — review lanes
Auto-applyHigh confidence, low blast radius. New/low-data accounts are excluded regardless of model confidence.
SuggestNeeds a human click to approve, reject, or edit before anything executes.
EscalateFlagged as high-risk — new market, thin data, or a budget change above threshold.
Every AI action — creative or budget — is classified into exactly one lane. This is what separates a copilot from "call the API and ship whatever it says."
The product isn't valuable because it uses AI — it's valuable because it makes the trust boundary explicit.
What it produced
15
Functional requirements defined
creative, monitoring & review lanes
3
Trust-gated release phases
Suggest-only MVP → multi-platform → Auto-apply
5 min
Rollback window on any auto-applied action
hard NFR, not a v2 add-on
One north-star metric — incremental ROAS attributable to accepted suggestions — instead of a dashboard of everything, so the team can't optimize for engagement-with-the-tool over outcomes-for-the-customer
Immutable audit log across every lane, with per-account monthly LLM caps that alert at 80% before they're ever hit