Back

Enterprise B2B SaaS / AI Infrastructure · 2026

Designing an Automated AI Agent Evaluation Pipeline

Transformed manual, biased AI experiment testing into a systematic, automated evaluation pipeline. This end-to-end feature allows users to objectively test, compare, and optimize AI agents natively within their existing platform workflows.


Company
Cisco
Platform
Enterprise Web App
Role
Senior Product Designer (UX/UI)
Duration
3 Months
01

Overview

The Challenge:

Managing and evaluating AI experiments manually created immense cognitive load, relying on subjective guesswork without objective benchmarks or regression tracking.

The Solution:

We designed and delivered an end-to-end automated evaluation pipeline that integrates natively into existing enterprise AI workflows. The system allows users to initiate test runs in just 3 clicks, leveraging automated LLM-as-a-judge evaluations while handling long-running asynchronous jobs seamlessly in the background.

Impact:

Replaced manual workflows with an objective, 3-click testing framework, generating sub-10-second reports and providing teams with unprecedented analytical visibility into agent behavior and performance regressions.

02

The Problem Space

Before this initiative, users evaluating AI experiments had to manually test each configuration. This involved sending individual prompts and subjectively comparing the generative responses side-by-side.

Through user research and workflow analysis, we identified several critical friction points:

  • Unscalable Workflows: Testing dozens of prompt variations manually creates immense cognitive load.
  • Human Bias: Without objective benchmarks, comparison relies on subjective interpretation.
  • Missing Metrics: Users had no measurable way to track performance regressions over time or objectively compare latency, token usage, and tool invocation.
03

Systems Thinking in Action: The Core Flows

To ensure the pipeline integrated seamlessly into daily workflows without introducing redundant concepts, I structured the experience around three primary entry points. The overarching goal was to prevent context switching and present actionable data clearly.

1. The Global View: Bulk Evaluation

This is the primary operational hub. Users view their entire repository of experiments in a data table. From here, they can select one or multiple experiments and trigger a batch evaluation.

The global view acts as the operational hub, allowing users to select and batch-evaluate multiple experiments at once.

Contextual Setup

To evaluate, users select a "Golden Dataset" and an Evaluator. We simplified this by providing curated Default Evaluators (e.g., Correctness, Toxicity) alongside a custom LLM Judge flow for advanced users.

Contextual setup: Users select a "Golden Dataset" and choose between default or custom LLM evaluators to begin the test.

Designing for Async

Evaluations are long-running operations. Instead of blocking the UI with loading spinners, the modal closes, and we utilize persistent notifications and inline status updates. Users can navigate away and are notified upon completion, organically surfacing the best-performing configurations in the table.

Asynchronous design: Jobs run in the background with persistent notifications and inline updates, never blocking the user's workflow.

2. The Builder's View: Contextual Evaluation

Targeting the active creation phase (the Prompt Lab), this flow allows users to evaluate while actively tweaking parameters without losing their context.

Zero-Friction Invocation

To minimize cognitive load, the system automatically inherits the active configuration and evaluates experiments using a built-in default dataset. Users don't need to manually configure validation files from scratch; a single toggle ("Evaluate experiment") inside the active workspace is all it takes to trigger an objective run.

Zero-friction setup: The system automatically inherits the active configuration; users can trigger evaluations instantly using the default dataset toggle.

3. The Analytical View: Historical Deep Dive

When a user drills down into a specific experiment, they enter the historical ledger for that configuration. This view turns evaluations into auditable assets.

The historical ledger turns evaluations into auditable assets, allowing teams to track performance regressions over time.

Tracking Regressions

Users can track performance over time, seeing exactly when a change caused a drop in the semantic similarity score. We used progressive disclosure: the UI surfaces the core aggregate score upfront, but users can drill down to view or download the full report, revealing granular data like tool invocation frequency and per-prompt latency. They can also trigger a new evaluation directly from this historical view.

Progressive disclosure allows users to drill down into a specific run to view granular metrics and debug specific evaluation criteria.
04

Outcomes & Acceptance

The delivery of this pipeline fundamentally shifted how teams approach AI experimentation, replacing manual guesswork with an automated, objective framework. By accurately capturing 100% of tool invocations, it provides unprecedented visibility into agent behavior.

3
Clicks to initiate test

Straight from the active context

<10s
Report generation

After job completion

<5%
Score variance

On repeated LLM judge runs

05

Conclusion

Designing for complex enterprise systems is rarely about inventing entirely new interfaces; it is about seamlessly embedding deep technical capabilities into established workflows. By reusing existing platform entities and embracing asynchronous design patterns, we reduced cognitive load and maintained interface consistency. This project reinforced that good UX in technical domains isn't just about making things look clean; it's about making complex systems predictable, measurable, and safe to iterate on.