Flow Intelligence
Flow Intelligence
AI that acts before you have to ask.
I designed and built an ambient AI layer for complex insurance workflows. It uses live case context, client needs, carrier requirements, compliance rules, and historical patterns to anticipate problems and recommend the next-best action before the user has to ask.
Inside the case, not beside a chat.
Most AI assistants wait for the user to notice a problem, open an assistant, provide context, and ask the right question. Flow Intelligence reverses that model. It evaluates the case already in progress and, when multiple signals point to a preventable delay or risk, surfaces one focused recommendation directly inside the workflow.
No prompt. No context switching. No separate AI destination.
Observe. Detect. Recommend.
Flow Intelligence continuously evaluates available context, but intervention is intentionally rare. The system only surfaces something when the signals converge on a specific, useful action.
Observe
Case stage, elapsed time, open requirements, client deadlines, carrier patterns, compliance context, and relevant history.
Detect
Look for combinations of signals that indicate meaningful friction or preventable risk—not every anomaly or unusual event.
Recommend
Surface one next-best action, with enough evidence for the user to understand why it matters and decide what to do.
The human stays in control.
Flow Intelligence advises; it does not take over the workflow. Every recommendation can be inspected, acted on, or dismissed. The system exposes the evidence behind its recommendation while leaving judgment and accountability with the person doing the work.
Inspect
See the signals that caused the recommendation to appear.
Act
Turn the recommendation into a concrete next step without leaving the case.
Dismiss
Ignore the suggestion and continue working. No forced automation and no penalty for saying no.
See it happen.
Open a fictional underwriting case and watch Flow Intelligence respond when a new signal changes the risk profile. Inspect the evidence, take the recommended action, or dismiss it yourself.
Case NS-18472
Daniel Brooks
20-Year Term Life · Harbor Life · Day 9 · APS pending
What I was testing
Generating a recommendation was the easy part. The harder product problem was deciding when an AI system deserves the user's attention at all.
I designed the prototype around restraint: one recommendation, shown only when multiple signals justify an intervention, embedded directly in the existing workflow, with evidence and human control built into the interaction.
Intervention
When is the expected value of an interruption high enough to justify it?
Restraint
When should the system remain silent?
Evidence
How much context should appear immediately, and what belongs behind “Why this?”
Control
How does the human retain clear judgment and accountability?
Behavior
Does the recommendation actually change what the user does next?
My role
Concept · Product strategy · Interaction model · UX/UI · Functional prototype
Presented internally and selected for implementation by the company's Innovation Lab, with sponsorship from the CPO and CTO.