Human-AI collaboration

UX research and design to validate market need, user adoption, and decision workflows in AI-powered seafloor intelligence software

Overview

The company was at an early stage of scaling its AI-powered seafloor intelligence software solution. The software automated seafloor scanning and analysis, combining underwater imagery and sensor data to support environmental monitoring, infrastructure inspection, etc. As their AI capabilities scaled, a critical question emerged,

"Should this product exist in its current form?
And if so, how should humans and AI collaborate within it?"

Responsibility

My work focused on validating both the product’s relevance and its usability: first, assessing whether users genuinely needed and would adopt an AI-driven solution for seafloor examination; and second, understanding how users interpret, trust, and collaborate with AI-generated insights in real-world decision-making.

Process

Mixed-methods research
Think-aloud sessions
Comparative task scenarios
Thematic analysis

Problem

The company was investing in advanced AI capabilities, but two critical uncertainties remained. It was unclear whether users would consistently rely on software-based seafloor scanning rather than existing manual or semi-automated methods. At the same time, early users expressed uncertainty around interpreting AI outputs, raising concerns about trust and correct usage.

The challenge was ensuring the product met real market needs and supported confident user decisions.

Research Goal

The research aimed to validate whether AI-driven seafloor scanning addressed a real and recurring user need, and whether users would integrate the software into their existing workflows. In parallel, it examined how users understood and trusted AI-generated insights, and what UX patterns were required to support safe and effective human–AI collaboration.

Assumptions

Users need faster and more scalable alternatives to manual seafloor inspection.
AI-generated insights can reduce time-to-decision if they are interpretable and trustworthy.
Adoption depends on users understanding when to rely on automation and when to apply expert judgment.

The research goal was to de-risk an early-stage AI product by validating whether users actually needed automated seafloor scanning, and by understanding what it would take for them to trust and use AI outputs in real operational decisions

UX Research

The discovery phase combined mixed-method research with workflow analysis to understand user needs, motivations, and constraints.

Exploring how users currently scan and analyze seafloor data, where existing approaches break down, and whether AI-driven software could realistically replace or augment these workflows. In parallel, the research evaluated early software concepts and outputs to understand how users perceived AI recommendations and whether they felt confident acting on them.

This dual focus allowed validation of both problem–solution fit and usability risks.

Research validated that there was genuine interest in automated seafloor scanning when it clearly reduced operational effort, cost, or risk. However, adoption was conditional. Users were willing to explore AI-driven analysis, but only if outputs were transparent, explainable, and positioned as decision support rather than final authority.

This learning led to a refined understanding of the product’s value proposition and highlighted UX constraints that needed to be addressed before scaling.

Core Success Key Signals & Proxy Metrics

Key signals

Adoption &

Product Value

Adoption intent increased when AI clearly reduced effort or risk, but dropped when outputs lacked clarity or integration with existing workflows—highlighting product–market fit constraints early.

Trust & Explainability

Trust increased when explanations supported user reasoning rather than exposing raw model logic. Over-automation without context led to disengagement or blind skepticism.

Time-to-Decision & Efficiency

AI reduced time-to-decision only when outputs were interpretable at a glance. Poor information hierarchy increased cognitive load and negated efficiency gains.

Proxy Metrics

Preference for AI-assisted scanning → Observed tool choice during comparative task scenarios
Likelihood to adopt → Post-task rating after exposure to AI capabilities
Tasks completed with AI vs. manual methods → Behavioral comparison during usability sessions

Confidence ratings → Pre- and post-explanation confidence checks
Acceptance rates → Decision tracking across AI recommendations
Time spent validating results → Observation of verification behavior
Verbalized hesitation → Coded qualitative signals during think-aloud sessions

Time to decision → Timed task completion with vs. without AI support
Steps to validate results → Workflow mapping during sessions
Interruptions or backtracking → Behavioral friction logging
User-reported workload → Post-task reflection on mental effort

This framework was a design deliverable a structured set of observable signals built to stand in for long-term adoption data at early stage, giving the team a testable definition of what 'working' would look like before live usage data existed.

Limitations & Next Steps

This research was conducted at an early discovery stage, proxy metrics were used to estimate adoption, trust, and efficiency before real-world usage data was available..

Future validation included tracking live adoption rates, reliance on AI recommendations, and time-to-decision in operational settings to confirm long-term impact.

Previous project

Next project