Designing for Trust
in Agentic AI

Bosch's accounting team reviews direct-capitalization purchase requests — purchases recorded as fixed assets like machinery and equipment. The approval process was entirely manual: approvers performed up to 15 validation checks per order across 20 orders daily. That's 300 manual validations daily, hunting for data across fragmented, disconnected systems. Incomplete submissions triggered rejection loops, causing rework and delays.

The opportunity was clear: an AI agent could automate and approve these validations. But for enterprise adoption to work, users had to trust the AI's recommendations. Trust doesn't come from speed alone. It comes from clarity, guidelines, and control.

My job was to design an agentic AI experience where users could understand and trust the system's decisions. This meant answering three critical questions simultaneously:

1. 01

   Can the AI reliably do the work across fragmented data sources?

2. 02

   Can users understand what it's doing and why?

3. 03

   Can users trust it when real business decisions and financial liability are at stake?

I led discovery, definition, and design from the start — showing up in daily client meetings to gather knowledge, validate assumptions, and present work in progress. Here is the full arc of what I produced.

Discovery Phase

I mapped the current process across SAP S/4, Bosch Shopping Cart, email, and manual checks to surface friction and align the team around a shared baseline. I defined the core problem: manual cross-checking across disconnected systems was creating delays, rework, and inconsistent decisions.

I built two user personas from daily conversations with the Bosch team: the Accountant Approver (primary) and the Requester (secondary). Personas keep teams focused on real users instead of assumptions — answering questions like: what does this person care about? What tasks do they perform?

I also documented domain language — WBS element, direct capitalization, settlement rule, fixed asset. Unfamiliar terminology makes users distrust the system, even when it works correctly. Speaking the users' language ensured we built something both sound and accessible.

Design Phase

User flow. Mapped the ideal end-to-end flow for the SAP AI Plant Purchase App, defining where the AI acts autonomously, where it surfaces recommendations for review, and where it escalates. Why this matters: without clarity on automation boundaries, teams slip into building auto-approval features, which destroys user trust. This flow became the reference document in every meeting with SAP and Bosch engineers and stakeholders.

Customer journey map. Charted the Approver's experience across the full purchase lifecycle — actions, thoughts, emotions, and pain points at each stage. Why it matters: journey maps reveal where users struggle most, which is exactly where AI can deliver the most relief. This focused the entire team on impactful opportunities rather than guessing.

OOUX model. Applied Object-Oriented UX to define core objects (orders, WBS elements, assets, settlement rules) and their relationships. Why this approach: OOUX mirrors how users mentally organize their work, making the interface feel intuitive rather than imposed. When the design structure matches the user's mental model, everything clicks into place.

Mockups & vibe-coded prototype. I designed high-fidelity screens in Figma with Fiori compliance, reviewing and iterating daily with the client team based on their feedback. To validate the agentic flow with real users and surface edge cases, I built a functional prototype using Claude Code and Figma Make. This tangible model proved essential for stakeholder buy-in and informed our client-based testing. By building within the established design system, I ensured the component could scale seamlessly into SAP's broader product suite.

The core challenge wasn't the UI, it was defining the right relationship between human judgment and machine action. I held key principles throughout:

1. 01

   Machines Handle Routine Work

   The agent runs high-volume mechanical checks; humans focus on judgment calls that require context and experience.

2. 02

   Humans Stay in Control

   No AI auto-approval. A human reviews and approves every critical decision or transaction before it's finalized, keeping accountability clear.

3. 03

   Show AI's Certainty Level

   When the AI flags an issue, it also shows how confident it is. High-confidence issues get priority review; lower-confidence findings need closer scrutiny.

4. 04

   Radical Transparency on Every AI Action

   For each purchase item, users had to see the status, the decision, and the reason. Without visibility into AI behavior, trust collapses.

Designing for a 12K-person event in 2–3 weeks. With SAP Sapphire Conference requiring an interactive prototype and €100M invested in agentic experiences, I had to deliver quality work on an accelerated timeline.

Content design emerged as the foundation for enterprise AI trust. Every word in labels, confidence signals, error states, and AI-generated communications needed precision — it was the core of the design, not a finishing step. Testing with Bosch accountants proved that language alignment with their mental models determined trust, regardless of AI accuracy. Technical or robotic language caused second-guessing, even when the system was correct.

Using AI thoughtfully in design accelerated work through controlled iteration. I used AI for drafting, hypothesis generation, and content variations, but sharp editorial judgment was critical. Treated as a starting point rather than a final answer, AI became powerful without becoming a shortcut.