Early concept demo showing system feedback

Emergence AI Craft Platform

Designing Trust for Agentic Enterprise Workflows

Define interaction models for a fundamentally new kind of product — one where the system acts on behalf of users, not just responds to them.

Context

As the AI market evolved, former use cases such as near-field voice and physical conversational interfaces were being adapted to model- and intent-based systems.

Operating in stealth mode and leveraging years of learning from IBM Watson and Merlyn Mind voice-enabled classroom assistants for K-12 EdTech, the company was now looking to expand into cross-vertical enterprise agentic use cases.

Opportunity

This entailed leveraging the existing tech stack to minimize R&D spend while applying UXR methodologies to market-traction research for use-case identification.

Through close partnerships with investors, leadership, engineering, AI research teams, and revenue teams, use cases were built out and experiments launched to validate model reasoning and demonstrate concepts to investors.

Challenge

It was a burgeoning landscape with little precedent for how user expectations would affect adoption and usage. The nature of the emerging landscape meant that there was no playbook, no traction data, no competitor comps, and no limits on how we handled interactions with the system.

Given the range of use cases, assumptions were made about the amount of system feedback and transparency needed. This led to explorations into trust in the agentic systems being built.

Differentiating on Trust

Having spent years in the edtech space as Merlyn Mind, and building the tech stack and model infrastructure for teachers, students, schools, and districts. AI Safety was a central part of how the company was already operating before the cross-sector pivot. The strategy was twofold:

  1. The model layer
    Create small, single-purpose, self-contained models focused on education work-flows, rather than large LLMs trained on internet data writ large without safety protections specific for schools.

  2. The user layer
    Build the orchestrator in a way that makes reasoning legible and transparent at each step, that shows sources and assumptions, that allows for human intervention and redirection while maintaining human judgment.

CIO Priorities Driving Spending

Cross vertical expansion revenue drivers

Rapid prototyping, agentic workflow, and model testing from engineering teams.

Task

Build an interface that makes AI reasoning legible, actionable, and trustworthy at each step — without hiding the system's assumptions or removing human judgment.

Moments when automation must pause, surface its reasoning, and give control to the user.

Trust Breakpoints

I mapped the trust breakpoints- the moments when users needed to understand what the agent did, why, and what they could override. From there, I experimented with a pattern library of explainability states: confirmation flows, fallback surfaces, and progressive disclosure for multi-step agent tasks.

Intent Capture

Proprietary concepts were evolved rapidly. Much of it can’t be shared. This is an example from the mobile consumer demo on the Intent Capture breakpoint.

Results

We used the demos for internal and funder validation and evaluation. As a team, we launched Emergence AI out of stealth mode, designed the Emergence Orchestrator platform and playground environment, and helped secure $97M in Series B funding.