Designing Trust: An AI Coaching Product from Research to Prototype

Default every visibility setting to private

Options considered: Default-open visibility (faster manager buy-in) vs. default-private visibility (deliberate opt-in only).
Decision: Default all manager-visibility settings to off; make visibility a deliberate, informed opt-in labeled "Recommended" rather than a passive default.
Why: Across the research, users described the same fear in nearly identical language: "if this can reach my manager, I'll only say safe things." Data collected for development being repurposed for evaluation was the single most consistent trust-killer in both internal studies and peer-reviewed literature. Default privacy was the strongest available signal of genuine safety.

Give the AI a non-directive persona, not an advisory one

Options considered: An AI that summarizes and gives direct advice vs. an AI that reflects, reframes, and asks questions without prescribing an answer.
Decision: Built the AI's behavior around one rule — every response adds something new (a reframe, an observation, a question) and never just repeats back what the user said or tells them what to do.
Why: Research on AI trust was consistent: people disengage the moment an AI tool feels like it's directing them. Preserving the user's sense of control over their own thinking predicted sustained engagement more reliably than the perceived intelligence of the advice itself.

Make every AI interaction user-called, never system-surfaced

Options considered: Ambient/proactive AI that surfaces coaching prompts contextually vs. an AI the user has to actively open.
Decision: The AI coach is available on demand but never interrupts — the user always initiates.
Why: Internal usability testing found AI-surfaced interactions were universally disliked, while user-called interactions significantly outperformed them on trust and engagement. Being "helped" by AI you didn't ask for reads as surveillance, not support.

Built the leading concept as a structured, scaffolded program — not an ambient assistant

Options considered: Three entry-point concepts were prototyped in parallel — an always-available ambient assistant, a structured onboarding program with a defined journey, and a coaching moment triggered by feedback events.
Decision: The structured-program concept became the strongest candidate with stakeholders.
Why: It extended a multi-modal, incremental learning-path pattern I had designed earlier in an unrelated product area at the same company, so it built on a product pattern users already trusted rather than asking them to adopt an entirely new interaction paradigm.

Research readout

Problem solved: Made a 17-page synthesis of internal and academic research legible to leadership as concrete, testable design conditions rather than an abstract literature review.

Self-Determination Theory, validated across hundreds of workplace studies, identifies three psychological needs whose satisfaction predicts authentic engagement and whose frustration predicts surface compliance: autonomy ("I'm in control of this"), competence ("I'm actually growing"), and relatedness ("this understands me"). Every prototype direction was measured against these three conditions.

Concept 1 — AI Coach across the Talent Suite

Problem solved: Tests whether an always-available coach earns real use when nothing ever pushes a leader into it.

Enrollment arrives as ordinary mail. No program, no schedule: the coach is simply there when the coachee wants it.

The coach sits in the corner of the normal dashboard — user-called and never AI-surfaced, so it's present everywhere without ever popping up uninvited.

Compass narrates sessions back inside an overlay panel, then goals and saved insights in the coachee's own words — growth as a felt sense of accumulation, not a completion bar.

Concept 2 — AI Coach creates a program (the one that won)

Problem solved: Tests whether a structured six-month program can still feel personal rather than assigned — extending a multi-modal, incremental learning-path pattern already trusted elsewhere in the platform.

A named six-month program from a real person at the company, not a system notification.

The hero card narrates how the coachee's framing of the problem has shifted across sessions — "where you started, where you are" — not a completion metric.

The next session is live; later ones stay locked, and completed work collapses into an accordion.

Expanded, the full arc of agenda-building sessions is visible — relationship, leadership, personal, business.

Concept 3 — AI Coach embedded in feedback

Problem solved: Coaching enters at the moment 360 results land — when the need for interpretation is highest — testing whether first contact works better in-product than as an email.

Compass reads the peer data as a narrative first; each theme then pairs a plain-language synthesis with the assessment scores behind it.

Configuration as coaching input, not app settings

Problem solved: Generic coaching language is the fastest way to lose a leader's trust. Rather than a typical settings panel, each setup question was framed as something Compass would visibly use in a real session — with a live "What Compass will now know" preview after every answer.

An admin's exact internal phrases — "earning the room," "above the line" — become words Compass recognizes and uses natively in sessions, instead of paraphrasing them into generic business-speak.

Trust architecture, configured explicitly

Problem solved: The same "private by default" principle from the coachee-facing product had to be enforced here too — as a deliberate admin decision, not a buried toggle, since this is the one screen that determines whether an entire organization's employees will trust the product at all.

Privacy defaults to on, manager visibility defaults to "sees nothing," and the copy names the risk directly: "function creep is the most consistent trust-destroyer in workplace AI."