Designing AI Behavioral Systems

Designing a system to align teams, define AI behavior, and enable better decisions at scale

Overview

Inside a regulated healthcare data company, AI was accelerating - but teams were misaligned on how to define, govern, and ship it responsibly.

Sector:

Regulated healthcare data
Enterprise SaaS
Clinical research

The Challenge

The organization was moving fast on AI. The problem wasn't momentum - it was coordination. Product, UX, engineering, and domain experts were each answering the same foundational questions differently: What should this AI do? What does good output look like? When do we know it's ready to ship?

Decisions were being made implicitly. AI behavior was under-specified, leading to inconsistent outputs, costly rework late in development, and eroding trust in AI-generated results. In a regulated healthcare context, the stakes of that erosion were high.

This was not primarily a technical problem. It was an alignment and governance problem that needed a system.

The Work

Tools & Craft:

Figma
AI prototyping
Behavioral specification frameworks
Workshop facilitation
Playbook design

Designing how work moves: the Innovation Flow

I designed a five-stage model - Ideate > Prototype > Discovery > Build > Deploy - that gave teams a shared path from idea to production. Each stage clarified when and how decisions were made, which stakeholders needed to be involved, and what readiness looked like. A Stage Gate Rubric standardized evaluation across four dimensions: problem clarity, data readiness, risk assessment, and cost of failure.

Defining AI behavior: the Prompt Triad

The deeper challenge was specification - teams were prompting AI systems without a shared language for defining what they actually wanted. I developed a Behavioral Specification framework and a co-authoring model (the Prompt Triad) that brought Product, UX, and Domain expertise together to define: what the AI should do, what good output looks like, where users might be misled or confused, and how failure is recognized and handled. This shifted teams from prompting and hoping to defining and validating intentionally.

Translate alignment into a unified system

An end-to-end AI Innovation Playbook

The frameworks were packaged into a practical playbook covering the full lifecycle from idea through deployment - with defined roles, engagement points, and decision criteria across Product, Engineering, Architecture, and Governance.

Outcomes

Cross-functional alignment established before build began, reducing late-stage surprises and rework across Product, UX, Engineering, and domain expert teams
AI behavior defined with shared specificity for the first time - consistent outputs, clearer expectations, measurably improved trust in AI-generated results
A Stage Gate Rubric adopted as the standard readiness model across multiple AI initiatives, replacing implicit decision-making with a shared evaluation framework
A scalable foundation - the AI Innovation Playbook and frameworks are now used across teams, not just the original initiative that prompted their creation

Key Insight

The hardest part of AI development isn't building the model - it's getting the humans to agree on what the model should do.

Behavioral specification is a design problem. When you treat it like one - with structure, shared language, and cross-functional ownership - AI becomes a system for how teams think and decide, not just a tool for generating outputs.