Sebastián Balderas — Home
STRATA

Strata — AI-native Action Systemfor Enterprises

An AI platform for contract office furniture that coordinates work among manufacturers, dealers, and installers. I designed the operating layer that unifies execution.

The Strata dashboard open on a laptop at a desk at night, city lights behind it — a personalised greeting, a sidebar of workflow modules, a recent-activity feed and a monthly performance panel with progress bars.

My Role

Sebastián Balderas waving hello
Senior UX/UI Engineer

Leading discovery, workflow and system modeling, interaction design, design systems, and AI-assisted delivery in a live production environment.

This case study covers my work at Agentic Dream, embedded with their client Avanto since Aug 2024. Avanto has two products: Orderbahn, the earlier document-centred platform, and Strata, its evolution from documents to actions. Orderbahn already had a design system and modules when I arrived — I improved the existing modules, built new ones for new functionality, and carried the platform through that evolution.

Context

Strata is an AI-native operating system for contract office furniture. A few manufacturers sell through many dealers who do not know each other, so no participant sees the whole order — the market spans manufacturers, dealers, rep groups, and property management. It coordinates execution among manufacturers, dealers, reps, and installers, and serves enterprise accounts including CBRE and Bank of America. I designed the operating layer that defines workflows, decisions, roles, and permissions for safe, scalable execution. Strata is the evolution of Avanto's earlier platform, Orderbahn: the same operations layer moved from documents to actions.

What I delivered

AI-Native Role & Permissions Layer

An AI-native layer that models roles and permissions for safe orchestration across teams.

Action-First Action Center

An Action Center that replaces dashboards with action-first workflows.

Design System for Code Delivery

A Figma design system aligned with code to support incremental delivery.

Human+Agent Decision Boundaries

L2 shipped to production for bounded prioritization and routing; L3 was designed and deliberately held back until audit trails and permission scopes are safe enough to remove human review.

Before

Teams relied on disconnected workflows, manual coordination, and unclear ownership.

Data was visible, but not actionable.

Users struggled with exceptions, permissions, and next-step ambiguity.

After

Teams work through a unified, guided system of action.

Ownership, urgency, and next steps are surfaced clearly.

Managers and operators act faster with fewer errors and less coordination overhead.

Product Foresight & Constraints

Because Strata was already live and embedded in operations, I focused on evolving the system incrementally, reusing components, tightening roles and states, and designing changes engineering could ship without disrupting daily work. I also partnered with Product Managers and Product Owners to translate prioritized workflow states into Jira tickets, keeping execution aligned with the state model.

Selected Impact

Improved SLA compliance by +40%.

Increased throughput by +40%.

-30% operational delays through state-based execution flows.

Execution is traceable end to end, with dashboards driven by workflow state rather than static reporting.

How these were measured

The figures came from the platform's own reporting, compared before and after the Action Center rollout and benchmarked against Orderbahn, the platform Strata evolved from. The supporting evidence was client SLA dashboards, product analytics on drop-off points and time spent per workflow, and moderated and unmoderated user interviews.

The attribution matters as much as the method: I owned the workflow mapping and the research. The measurement instrument was the client's.

📌 Three core problems

Lack of operational clarity

Static dashboards and disconnected tools led to manual coordination, missed context, and inefficiencies at scale.

Data, not decisions

Dashboards showed metrics but didn't guide action or next steps.

Complexity without guidance

Overlapping workflows, exceptions, and permission layers led to errors and slow execution.

🗺️ Process — AI-augmented from research to delivery

Discovery & Research

Stakeholder interviews, workflow mapping, information architecture audits, decision-point analysis, and operational shadowing, anchored in real workflows rather than UI assumptions.

Synthesis & Validation

Extracted behavioral patterns, bottlenecks, ownership gaps, and exception scenarios; validated flows through stakeholder reviews, role-based simulations, and scenario walkthroughs.

Prototyping with AI

Used Claude, Claude Code, and Figma AI to generate early interaction models and wireframes, iterate workflows quickly, and validate feasibility faster.

Design System & Deliverables

Consolidated UI and system logic in Figma with a scalable component library, workflow-adaptive modules, role-aware patterns, and engineering-ready specs.

See the B2B2C research and workflow mapping in full

⚙️ Constraints & Design Principles

Strata was already deployed in production across multiple dealer networks. Every design change had to be backward-compatible, shippable by a lean engineering team, and non-disruptive to active workflows. I established four principles that filtered every decision throughout the project.

Action-first

Surface the next action, not just data. Every screen guides the user toward a specific step.

Role-aware

Each role sees only what moves their work forward — operators, managers, and admins see different surfaces.

State-driven

UI reflects real-time workflow state. A task in-progress looks different from one blocked or pending.

Incrementally shippable

Every change ships independently and doesn't break the system currently in use.

🤖 AI-native workflow & design-to-code

AI embedded in the design process

I used Claude and Claude Code, orchestrated with n8n and MCP, to generate initial workflow state models from operator interviews, compress research synthesis cycles, and draft edge-case scenarios for exception flows. This cut discovery-to-definition time and surfaced states that traditional user research would have missed.

AI didn't replace judgment — it compressed the distance between raw data and structured decisions.

Design-to-code handoff

The Figma design system was structured for direct engineering consumption: semantic variables for color and spacing, annotated component states, and named interaction patterns. Every component had a documented state model aligned with the backend workflow states.

This meant engineering could implement directly from the design file without translation — reducing back-and-forth and keeping design intent intact through shipping.

💡 Key design decisions

01

Unified state model

One source of truth for all workflow states. Every UI element maps to a named state, eliminating ambiguity between design, product, and engineering.

02

Role-based dashboards

Operators see task queues. Managers see exceptions and bottlenecks. Admins see configuration. No role navigates through another's context.

03

Action surfaces over data views

Screens are structured around what the user does next, not what data is available. Passive monitoring screens were replaced with guided action surfaces.

04

Progressive disclosure

Complexity is revealed on demand. Default views show essentials; details, history, and advanced options are one tap away — not in the way.

05

14 identified users, scoped down to 3

Research surfaced 14 distinct users and 25+ pain points. I scoped the model down to the 3 driving adoption — manufacturers, dealers, and rep groups — and designed for their control points instead of trying to solve every pain point at once.

06

A system of action, not a system of record

The foundational architecture treats the platform as somewhere work gets done, not somewhere work gets logged. I presented it to leadership and the wider Strata team — aligning the organization on one architecture before the interface existed.

07

Taste & Critique: alternatives, not instinct

When Strata needed to decide the direction for the dealer/rep/installer roles, I put together an A/B comparison of the options — including the UI I'd already designed under Orderbahn's document-management model — and presented it to the CEO and the project manager. They chose to pivot to Strata's AI-first, action-execution model, which meant that UI no longer fit. Rather than force it into the new architecture, I retired it and reused the underlying research and components as the foundation for the new UI Kit — so the work wasn't lost, it was redirected.

B2B2C Expansion for Strata

📡 B2B2C

Strata was originally designed for enterprise back-office operations — a B2B system. As the platform matured, the opportunity emerged to extend the same operational logic outward: from dealer networks (B2B) through to the end customer (B2C), creating a B2B2C loop. I designed this layer as wireframes and a front-end prototype; it was never integrated to production.

B2B — Enterprise layer

Fleet operators, procurement managers, and admins manage the operational backbone: inventory, SLAs, task routing, and exceptions. Strata's core surface.

B2B — Dealer network

Dealers receive, process, and close service requests. They operate within Strata's workflow but with scoped visibility — only their assignments and queues.

B2C — End customer

The layer was designed so customers could receive real-time updates, raise exceptions, and track service progress through a simplified mobile-first surface on the same state model.

How I validated the decisions

I validated these choices through workflow mapping, stakeholder reviews, and scenario walkthroughs, focusing on how people actually coordinated work in a live environment.

That process helped confirm which interactions needed to be standardized, which needed role-specific treatment, and where AI could safely support execution without taking control.

✨ Solution

Strata is an action-first operating model that gives humans and AI agents shared context to coordinate work safely and efficiently.

Core components include:

An adaptive Action Center.Workflow states.Task surfacing.Guided paths.

Users land in a workspace that shows only relevant work, guides decisions step by step, and adapts to role and permissions.

📈 Outcome

Improved cross-role operational clarity and reduced error-prone handoffs.Faster task prioritization and lower dependency on manual coordination.A stronger foundation for AI-driven guidance on top of real, well-defined workflows.

💡 Learnings & positioning

Systems design matters more than individual screens in enterprise environments.Enterprise UX is about orchestrating work and decisions, not decorating interfaces.AI only adds value when UX, states, and governance are structured so agents can act safely.

AI Automation Maturity Framework

I evaluate AI automation across three levels to decide where automation should stop and agentic judgment should begin:

L1

Workflow Automation

Shipped

Point A → B automation, fixed rules, no reasoning.

L2

Agentic Integration

Shipped in production (action prioritization, routing)

LLM agents interpret context and make bounded decisions within a step.

L3

Adaptive Agent Loops

Designed and deliberately held back — never shipped or active — until audit trails and permission scopes are safe enough to remove today's human review.

Agents would evaluate options in real time against historical context, SLA data, and permission scopes, closing the loop with minimal human intervention on low-risk decisions.

This staged approach kept Strata's live operations stable: L2 shipped into production where the decision surface was narrow enough to trust, while L3 was designed but deliberately held back until audit trails and permission scopes were mature enough to remove human review safely.

Where AI could act wasn't settled as a policy — it followed a chain. The end-to-end map of a single order, across dealer, rep, manufacturer, quoting, contract and cash order, was the research; the exception moments and ownership gaps it exposed were the insight; the rule that the AI prioritizes and proposes while a person reviews that prioritization and the impact decisions was the strategy; and L2 reaching production for bounded prioritization and routing, with L3 designed and deliberately held back, was the outcome.

See where that boundary sits inside a single order

Agent structure, and which model runs each step

I authored the agent-structure document for Strata: an end-to-end workflow map in Figma, and on top of it the agent architecture derived from the researched workflow — including which model each step would use and which it would not. The model choices were very specific at that point, and they were made to control cost.

That is a different axis from the three levels above. Those decide where an agent is allowed to act; this decides what it costs to run once it is allowed. A step can be safe to automate and still be the wrong one to hand to an expensive model.

Closing

Strata shows that enterprise platforms should execute work, not just store it. It positions my work as designing AI-native systems of action that align human and agent workflows with real-world constraints and delivery.

What this shows about my Skillset

I can design the information architecture of an orchestration platform in a live environment.I create AI workflows that agents can act on safely.I align Action Centers and design systems across cross-functional teams to accelerate adoption.
MEDIUM · CONTRA

Why holding L3 back was the decision:Read the full article on Medium

I write about these decisions in more depth.

AI workflow audit, fixed scope:Check the price on Contra