Building AI Agents That Replaced 30+ Hours of Manual Work

The problem was time. Running multi-channel marketing operations across paid media, ecommerce, content, and creator programs generates enormous amounts of data, and that data needs to be collected, reconciled, and turned into decisions on a regular basis. At both Innova Electronics and FITH, this work fell to one person: me. Hours each week spent moving between spreadsheets, pulling data from separate platforms, formatting reports, and compiling them into something a stakeholder could actually read and act on. That was hours not spent on forecasting, strategy, purchase order planning, or anything that actually required human judgment.

I had no development background. No computer science degree, no prior coding experience. What I had was a clear understanding of the problem, a willingness to learn through doing, and access to tools that made building possible without traditional engineering skills. Claude Code, Replit, Lovable, and ElevenLabs became the foundation of a self-taught automation practice built entirely around solving real operational problems. The approach was simple: identify what was consuming time, plan the solution thoroughly before touching any code, and build it through prompt-based development and iteration.

The result was a growing library of AI agents and automated workflows deployed across both companies that reclaimed hours of daily manual work and redirected that capacity toward higher-level strategy. These were not experiments or proof-of-concept demos. They were production tools, used every day, continuously improved, and directly tied to business outcomes.

The first major workflow was a data reporting agent built to eliminate the most time-consuming recurring task: weekly performance reporting. The agent exports raw data from source platforms into CSV files, parses and reconciles those files, generates slide decks with structured visualizations, and produces a high-level written analysis of the data. This workflow was deployed across Meta Ads performance, Shopify sales data, product sales forecasting, purchase order forecasting, and content creation analytics for both Innova and FITH. What previously took hours of manual spreadsheet work per reporting cycle now runs with minimal human input, producing outputs that go directly to stakeholders.

The most technically complex workflow built was a content production pipeline that connected competitor intelligence directly to video editor task queues. The agent ingests transcripts from competitor video content, analyzes the messaging and structure, adapts the script to our own products and brand voice, and outputs the adapted scripts to a tracking sheet. From there, it integrates with ElevenLabs to generate AI voiceovers from the finalized scripts, then connects to JIRA to automatically create project tickets and assign them to the appropriate video editors. A process that previously required manual research, writing, briefing, and administrative follow-up across multiple tools became a single automated workflow from input to assigned task.

For FITH specifically, the most operationally critical tool was a custom inventory management system built to solve a problem that could directly kill revenue: stockouts. The system tracks real-time sell-through velocity per SKU, applies weighted demand calculations across recent sales windows, detects suppressed demand from out-of-stock periods, calculates dynamic reorder points based on lead time and safety stock logic, and generates suggested purchase order quantities automatically. When stock hits a reorder threshold, the system sends an alert and surfaces a replenishment recommendation in the dashboard. What had been a reactive, gut-feel ordering process running weekly off manual spreadsheet checks became a proactive, data-driven system that flagged issues before they became lost sales.

Every workflow was built with the same principle: automate the process, but keep a human in the loop before anything consequential happens. Reports are reviewed before they go to stakeholders. Content scripts are approved before voiceovers are generated. Purchase order recommendations are vetted before orders are placed. The automation handles data collection, calculation, formatting, and routing. The decisions remain with people. This distinction matters because automated systems will produce errors, and they did during development and after launch. Because the tools were accessible to the team, bugs were documented, reported back, and fixed iteratively through the same prompt-based development process used to build them in the first place.