My Human–AI Reporting Assistant

Personal project- 2025

Lead designer

I got tired of manual project tracking, so I built an AI assistant to monitor project progress and flag risks early, giving me my time back for the high-value work.


My role

As Lead Designer, I spend hours each week creating progress updates for cross functional partners. I saw an opportinity to reduce manual work, improve clarity, and build a practical AI project to deepen my skills.

I designed and built the a AI-assisted workflow that helps me track project progress faster, flag risks early, and share updates automatically. Through this project, I strengthed my experience with AI workflow design.

#AI UX #System thinking #Workflow & automation design #Prompt design & AI interaction design #Human-in-the-Loop Design

Key impacts

Reduced weekly reporting time by ~60–80%


Boosting communication speed by 90%


Surfaced risks earlier and reducing missed blockers


Overview of my AI Assistant

The solution is a no-code Human–AI reporting assistant that transforms fragmented weekly design updates into a clear, consistent, automated status summary. I designed this as a service workflow that combines structured inputs, AI-generated synthesis, and human review step to eliminate manual reporting, make reporting faster, clearer and more reliable.


Approach

  • I mapped the end-to-end reporting workflow, identified where time and clarity were being lost, and defined the functional and data requirements needed for a reliable Human–AI system.

    Methods: Current-state workflow map, reporting requirements & data needs

  • I designed the service blueprint for the full system, establishing data schemas, touchpoints, and clear responsibility boundaries between human and AI automation to ensure trust, control, and scalability.

    Methods: Service blueprint, human-AI interaction

  • I built the first working prototype in n8n, using real weekly updates to test and refine the prompt logic, interaction flow, and generated outputs. I also validated how AI responds under real conditions.

    Methods: Model in the loop prototype using n8n, iteration with prompt system & validated output

  • I connected the structured inputs, workflow logic, and AI summaries into a seamless reporting assistant, enabling automated weekly updates and maintaining a human review loop for accuracy and clarity.

    Methods: Deployed Human-AI reporting system, workflow refinemt using live usage

Key components

01

A structured input system for reliable AI behavior

A two-sheet data model (Projects + Weekly Updates) that captures updates in a predictable, machine-readable format.

This provides the assistant with clean inputs, reduces ambiguity, and improves the consistency of AI-generated summaries.

02

A Human-AI workflow blueprint

A service blueprint that defines how humans and AI collaborate to ensures clarity, control, and transparency in the system.

✓Humans log updates, blockers, and next steps

✓The AI synthesizes and structures summaries

✓Humans review and adjust the final output.

03

Prompt system for reliable AI behavior

A prompt system that guides the AI to produce reliable and clear summaries and insights

✓Clear sections

✓No hallucinations or invented data

✓How to treat missing data

✓Prioritization based on impact

✓Concise, consistent tone

04

Automated workflow in n8n

A no-code workflow that automates the reporting cycle and connects every part of the Human–AI system

✓ Pulls weekly updates into a structured format
✓ Runs the prompt system for AI summaries & insights
✓ Formats and publishes a report via email
✓ Maintains a human review step for trust and accuracy

Glimpse of outputs

Prompt design

I designed the prompt system as the interaction layer between structured project data and the AI model. The goal was to produce a consistent, executive-ready weekly UX report with zero hallucinations and maintaining human oversight. The prompt sets a clear structure with tables, sections, visual indicators, and fallback text like “No updates this week.”


To keep the assistant accurate and trustworthy, I added guardrails

Guardrails

use only provided data, never infer missing information, surface inconsistencies only when visible, preserve blanks, and return explicit fallbacks for empty sections.

Iteration & Refinement

Refined the prompt through multiple iterations using real weekly updates

Early outputs were uneven, so I tightened tone, prioritization, ordering, and formatting. I also designed rules for edge cases like missing data, mismatched IDs, and conflicting entries—making the system increasingly robust and reliable.

Prompt Evolution: V1 → V2 → Final”


Prompt Version 1

✓Generic role (“business analyst”)

✓ Minimal guardrails → inconsistent outputs

✓ No visualization (difficult for executives to scan)

✓ No Project ID → Name mapping

✓Lacked risk structure or confidence ratings

✓Could infer/guess details if missing

✓Summaries mixed tasks, risks, and accomplishments

✓ Not optimized for email reporting

Final Prompt

✓ Domain-specific persona (“analyzing design team progress”)

✓ Strict non-hallucination rules (IDs, tasks, dates, risks)

✓ Project ID → Project Name validation + data integrity checks

✓ Email-safe visual system (ASCII bars, icons, severity markers)

✓ Tiered logic for Key Wins (Project-Level + Task-Level)

✓ Strongly constrained risk framework + confidence scoring

✓ Structured data ingestion using raw JSON for accuracy

✓ Executive-ready layout optimized for weekly reporting

Report


Behind the work

  • I wanted to show how practical, system-level AI design can turn a manual reporting process into a fast, reliable, executive-ready workflow. By combining structured data, prompt engineering, and human-in-the-loop review, I built an AI assistant that transforms fragmented updates into clear, trustworthy summaries complete with guardrails, visual indicators, and zero hallucinations.

  • I learned that good AI design comes down to being clear, setting the right rules, and working together with the system. By treating the prompt like a real product, leaning up the inputs, adding guardrails, and planning for edge cases the workflow became much more reliable.

    And the biggest lesson was this: AI doesn’t replace human judgment, it supports it. It helps make reporting faster, clearer, and more dependable.