How I Sold a $10K AI System to Automate 8 Hours of Daily Work in Half

I recently met with an Australian migration visa company. Their team was drowning in paperwork, spending a full eight-hour day processing the documents for a single client. It was a bottleneck that capped their growth and burned out their best people.

They were using AI, or so they thought. The team was sharing a single ChatGPT account to help process complex documents related to labor market testing, salary justifications, and position descriptions. On the surface, it seemed like a step towards innovation. In reality, it was a classic example of a Level 1 automation problem: using a consumer-grade tool for a critical business function, leading to more chaos, not less.

The shared account was a mess. Every time a new document came in, an employee had to start from scratch, re-writing prompts and trying to give the AI the right context. Chats from different employees and clients were jumbled together in a long, confusing thread. Worse, the process still required extensive manual work. Team members had to constantly leave the chat, search for specific immigration rules or salary benchmarks online, then copy and paste that information back into the prompt.

They were stuck in a loop of manual effort, paying for a tool that was adding friction instead of removing it. This wasn’t a system; it was a symptom of a broken process. They needed a real automation asset, not a shared chat window.

From a Single AI Agent to a Specialist Team

When I first approached the problem, our initial instinct was to build one large, sophisticated AI agent to handle the entire document from start to finish. It seemed logical. One document, one AI.

I was wrong.

The performance of this single “monolithic” agent was poor. It struggled with the complexity and nuance of the different document sections. It would hallucinate details, misinterpret instructions, and produce inconsistent output. The quality was simply not good enough for a professional setting where accuracy is paramount. A single error in a visa application can have serious consequences.

This failure forced me to rethink our entire strategy. The problem wasn’t that the AI was incapable; it was that I was asking one generalist to do the job of an entire team of specialists.

So, I broke the problem down. Instead of one agent, we decided to build a team of highly specialized AI agents or nano agents, each responsible for one, and only one, section of the document. The title page had its own agent. The “Labor Market Testing” section had its own expert agent. The “Salary Justification” analysis was handled by another. Each agent was designed and trained for a single, specific task.

This multi-agent approach was the breakthrough. I saw an immediate and significant increase in the quality and accuracy of the output. By narrowing the focus of each agent, we eliminated the confusion and inconsistency we saw with the monolithic model.

Building the Automation Engine

To bring this specialist team to life, I built a system that was both powerful and easy for the client to use.

First, I created a simple graphical user interface (GUI) where the team could upload the raw client documents. No more copy-pasting into a chat window. Just a clean, simple drag-and-drop interface.

Under the hood, I used the low-code platform n8n to build the workflow that managed the entire process. When a document is uploaded, n8n acts as the project manager, routing each section of the document to the correct specialist agent.

The real refinement came when I trained each agent. I didn’t just give them instructions; I gave them examples. I fed each specialist agent a collection of previously processed, high-quality documents. I showed the “Position Description” agent exactly how a perfect description should be structured — whether to use tables, bullet points, or formal paragraphs. This provided a clear template, ensuring the output was not only accurate but also perfectly formatted every time.

The final piece was to eliminate the manual research loop. Instead of having employees search the web for information, I gave our AI agents their own research tools. I integrated Perplexity into the workflow, allowing an agent to perform its own deep research. If the “Salary Justification” agent needed to find the current market rate for a specific role in Sydney, it could now find that information on its own and pull it directly into the document.

The setup for a multi-agent system like this can be complex. For those who are not developers, I will be adding the full n8n workflow and detailed instructions to our Corporate Automation Library (CAL). The CAL is a growing resource of over 50 high-impact automations designed for businesses ready to move beyond Level 1.

Click Here to gain access to CAL. I upload 2–4 new corporate automations weekly.

Cutting Eight Hours of Work in Half

The results were clear. Based on our benchmark tests, the new system cut their document processing time by 50%. A task that once took an entire eight-hour day now took half that time.

This wasn’t just a small efficiency gain. It fundamentally changed their business capacity. They could now process double the number of clients with the same number of staff. They avoided the significant cost and time of hiring and training new employees, directly impacting their profitability.

Furthermore, the system created a structured, searchable database of all processed documents. No more lost chats or forgotten prompts. They now had a historical record they could reference, providing a level of consistency and quality control they never had before. The prompts are set from the get-go, but the team can easily access and tweak them within n8n to continuously refine the output. The visual, low-code nature of the platform makes it accessible for them to make adjustments without needing a developer.

An Honest Look at the Hurdles

The project wasn’t without its challenges. One of the client’s requirements was to include screenshots of certain websites as evidence in the documents. I built functionality to automate this, but I ran into roadblocks with sites protected by services like Cloudflare, which blocked our automated tools.

This is a common reality in automation. You will hit technical walls. The solution is not to give up, but to iterate. I’m already working on a more robust solution to tackle this, but it’s a reminder that building these systems is a process of continuous improvement. Being transparent about these limitations with the client is crucial for building trust.

This multi-agent approach is not limited to migration visa companies. The same principles apply to any industry that relies on complex document processing. Think of legal firms drafting contracts, insurance companies processing claims, or medical professionals handling underwriting forms. Any process that can be broken down into a series of specialized tasks is a prime candidate for this type of automation. It’s a scalable model for turning unstructured information into structured, high-quality output.

By moving from a single, shared tool to a dedicated system of specialist agents, the company moved from the chaos of Level 1 to the structured efficiency of Level 2 and 3 automation. They didn’t just buy an AI tool; they invested in an intelligent system that will serve as a foundation for future growth and a lasting competitive advantage.

Ritesh Kanjee | Augmented AI