Case study · Client confidential · 2025–2026

AI avatar product concept — research-led north star

Two phases of mixed-methods research reframed an AI avatar concept from "smart speaker with a face" into a sharper expert-coach narrative with two viable beachhead audiences — directly shaping the product brief and supporting an executive demo.

---

Context

A hardware proof-of-concept had generated executive excitement but lacked the research-backed foundation needed for leadership confidence that the product had legs. The core questions: does this concept work with consumers, who is it for, and what does the avatar actually do for users?

I took the methodology lead from kickoff, reworking learning questions across three test plan iterations. The most consequential call came in week 8: I reversed my own initial position on qual-vs-quant sequencing, recommending qual-first based on the early-stage nature of the concept. That decision shaped the entire engagement.

---

Approach

• Phase 1 — Concept viability and audience identification (Nov – Dec 2025) Online survey with concept testing (n=1,200, gen-pop with parent quotas), mobile diary studies, and 1:1 IDIs with smart speaker power users. Low-fidelity sketches conveying form factor; 29 use cases ranked by participants. Goal: establish whether the concept had legs and identify who it was for.
• Phase 2 — Sharpening the value of the avatar (Jan – Apr 2026) Phase 1 had benchmarked the concept against smart speakers. Phase 2 tested the avatar's specific value relative to text chatbots and voice assistants, using refined stimuli and direct comparison tasks across additional IDIs.

---

Key findings

• Two beachhead audiences Tech-centric families and performance-oriented power users — an expansion from Phase 1's family-first framing, giving the client a more flexible go-to-market story.
• Expert use cases win Cooking and exercise were the only scenarios where the avatar clearly outperformed voice-only interaction, scoring 4.7 and 4.5 out of 5. Butler use cases had immediate appeal; companion use cases didn't hold up.
• Avatar as trust signal Mismatched avatars (e.g., childlike voice paired with professional-advice use case) actively undermined user trust — repositioning avatar design from a styling decision to a functional product constraint.
• Portability as must-have Surfaced unprompted in 8 of 16 Phase 2 sessions, anchored to kitchen and exercise spaces. Moved from stretch goal to hardware requirement based on user demand, not assumption.
• Context-dependent privacy Spatial awareness was rejected in the abstract but accepted when tied to specific user-controlled tasks — reshaping onboarding sequencing to never lead with always-on camera features.

---

Impact

Research directly informed hardware placement decisions (bedroom removed from consideration), onboarding sequence design, and product positioning. The engagement grew from a research scope to include concept film support, PRFAQ contribution, north star drafting, and executive demo prep — scope expansions that reflect decision-making trust, not just delivery.

---

Reflections

Reversing my own methodological position mid-engagement — from quant-first to qual-first — was the right call and one I'd make again. The instinct to commit early was wrong; the concept's maturity level should drive sequencing, not habit.

What I'd do differently: identify the single most load-bearing claim in the client's internal narrative and test it directly in Phase 1. The client's pitch leaned on an assertion that never made it into the stimuli — a gap that only became visible late. I'd also name client-side stakeholder misalignment earlier rather than absorbing it into the deliverable.

Synthesis framing matters as much as findings. A mid-engagement feedback session made clear the report read flat for internal selling — findings were right, but buried. Leading with the strategic assertion and using data to defend it, rather than building from data to assertion, is now a standing practice.

Supporting materials available upon request. Client name, concept names, stimuli, and participant quotes abstracted per NDA. Artifact list pending NDA review.