AI Agents, RAG, and N8n: How Companies Can Move from Experimentation to Real Automation
What does an AI agent actually do, beyond the hype?
In this Commit & Push episode, Damien Filiatrault sits down with Micah Johnson, co-founder of Biggest Goal, to unpack a question many teams are still struggling to answer clearly: what is an AI agent, and where does it create real business value?
Their conversation moves from definitions into practical examples. They cover how agents differ from basic chat interfaces, where retrieval-augmented generation (RAG) fits into the picture, why tools like N8n have become so popular, and why so many company AI initiatives stall before they create meaningful results.
What an AI Agent Really Is
Micah’s definition is refreshingly practical. Instead of treating an agent as an abstract AI buzzword, he describes it as an AI-powered system that has instructions, access to tools, and the ability to decide what actions to take and in what order.
That distinction matters.
A normal chat interaction is usually one prompt followed by one response. An agent is different because it can loop. It can examine its instructions, decide which tool to use, take action, evaluate whether the task is complete, and continue until it reaches a stopping point.
In other words, the leap is not just “AI that answers questions.” It is AI that can operate.
That also means an agent does not always need a human to initiate every step. In many business settings, an agent can sit behind the scenes waiting for a trigger, such as an event inside Monday or ClickUp, and then take action automatically. That kind of workflow begins to look less like a chatbot and more like a digital operator embedded inside a company’s systems.
The Best Early Use Cases Are Often Boring
One of the most useful parts of the episode is how grounded Micah stays on use cases.
He does not pitch agents as magical replacements for entire departments. Instead, he points to smaller, high-value gaps inside business processes.
A good example is internal reporting. He describes an agent that reviews time-tracking data each night, analyzes month-to-date hours, and sends a summary to the relevant team every morning. On the surface, that may sound like plain automation. And Damien pushes on exactly that point: does this really require AI?
Micah’s answer is a good litmus test for when AI becomes useful.
If the task is simply retrieving data and calculating a number, traditional automation may be enough. But once the system needs to interpret the data, detect anomalies, identify takeaways, or produce an executive-style summary, AI starts to add real value. The difference is not whether the workflow is automated. The difference is whether the workflow requires judgment, synthesis, or explanation.
That framing is helpful because it keeps teams from overengineering simple tasks while also showing where AI can extend traditional automations in meaningful ways.
Why RAG Has Become Such an Important Pattern
The conversation then shifts into one of the most common modern AI use cases: giving internal teams access to company knowledge through RAG.
Micah uses a straightforward example. Imagine a company has a Google Drive full of SOPs. Instead of forcing employees to manually hunt through folders and documents, you can turn those materials into a searchable AI-friendly knowledge base. Then an agent can query that knowledge base, retrieve relevant information, and respond conversationally when someone asks how to perform a specific task.
This is where RAG becomes powerful. It connects language models to company-specific information without requiring the model to somehow “know” everything in advance.
Just as importantly, Micah points out that this process is becoming far easier than many teams realize. Tools like Raggy and similar services can connect directly to platforms such as Google Drive, SharePoint, and Dropbox, sync documents automatically, and keep the knowledge base updated as files are added, edited, or removed.
That reduction in technical friction matters. A few years ago, implementing this kind of system could mean writing a lot of custom ingestion and sync logic. Today, much of that infrastructure can be set up with far less effort.
A Simple Explanation of Vector Databases
For readers who have heard terms like “vector database” without a clear mental model, this episode offers a useful explanation.
Micah contrasts vector databases with traditional relational databases. In a relational model, you typically have structured tables connected by IDs and relationships. In a vector database, the core unit is often a chunk of content rather than a row in a table.
A document gets broken into chunks. Those chunks are embedded and stored in a way that allows the system to find pieces of text that are semantically related to a query. When someone asks a question, the system retrieves the most relevant chunks rather than forcing the model to consume an entire long document.
That helps solve a practical problem: context limits. Instead of handing an AI a 100-page document, the system can narrow in on the few sections that matter most, while still preserving metadata that links those chunks back to the original source.
It is a useful reminder that RAG is not just “search plus AI.” Done well, it is a structured way of narrowing the model’s attention to the most relevant context.
Narrow Beats Broad When Building Agents
One of the strongest practical lessons in the episode is Micah’s advice on scope.
He compares agents to interns on their first day at a company. If you give them too many tools, too much information, and too many responsibilities at once, they get overwhelmed. In AI systems, that overwhelm often shows up as hallucination, confusion, or poor decision-making.
So the better pattern is usually to keep an agent tightly focused.
A RAG agent, for example, may need only a chat trigger and one connection into a specific partition of a knowledge base. That is enough to perform a useful task well. Expanding scope too early can make the system worse, not better.
This point is especially important because many teams instinctively try to create a single all-knowing AI assistant for the whole company. The conversation suggests the opposite approach: start with narrower agents designed for narrower jobs.
That design discipline can make the difference between a tool that feels reliable and one that feels unpredictable.
Why N8n Has Become a Favorite for AI Workflows
The episode also spends time on N8n, which Micah describes as a major reason these systems are becoming more accessible.
At a high level, N8n is an automation platform in the same broad category as tools like Zapier or Make. But Micah argues that it stands out for a few reasons: it is open source, it can be self-hosted, and it makes both traditional automation and AI workflow construction unusually fast.
Damien highlights another reason it resonates with teams: the visual interface. Instead of burying logic inside code, you can build workflows more like a flowchart, dragging together nodes and visually expressing how a process should run.
That is important not just for speed, but for maintenance. Micah points out that when an AI workflow needs to change, you are often editing prompts and logic in plain language instead of rebuilding a more rigid coded system from scratch.
That ease of iteration is a huge advantage. In a space where teams are still learning what works, the ability to change a workflow quickly may matter as much as the initial build speed.
AI as a Leapfrogging Technology
A particularly interesting theme in the episode is the idea that AI is allowing many people to skip traditional technical learning curves.
Damien describes AI as a leapfrogging technology: something that lets people bypass older stages of technical evolution and move directly into building useful systems. In this framing, a non-developer can now create software-like workflows that would previously have required engineering time, infrastructure setup, deployment logic, and long iteration cycles.
Micah agrees, but adds an important caveat.
Yes, more people can go from idea to execution much faster. But speed alone is not the same thing as good design. Expertise still matters. Security still matters. Workflow design still matters. Institutional knowledge still matters.
That tension is probably one of the defining truths of the current AI moment. The barrier to building has dropped dramatically. The barrier to building something well has not disappeared.
Why So Many AI Projects Fail
Toward the end of the episode, the conversation turns to a question many leaders are now asking: why do so many AI initiatives underperform?
Micah’s answer is less about the models and more about organizational behavior.
Too many companies, he argues, buy subscriptions to tools like ChatGPT, Copilot, or Claude and then declare victory before any actual system has been designed. Employees are told to use the tools, but there is no shared framework, no standardization, no process design, and no real operating model around how AI should support the business.
The result is predictable. Individual people may get small productivity wins, but the organization does not produce consistent, measurable change.
Micah compares this to the way companies often roll out project management software. They buy a platform like Asana or Monday, announce that the team is now using it, and assume the tool itself will fix the operational problems. Of course, it rarely works that way. Without structure and strategy, the software becomes little more than a digital checklist.
His point is that AI rollout suffers from the same mistake. Tools do not create transformation on their own. Systems do.
That may be the clearest takeaway in the entire conversation.
The Real Opportunity Is Operational
Micah notes that Biggest Goal tends to focus less on embedding AI directly into customer-facing software products and more on operational workflows: sales ops, marketing ops, and internal execution.
That is a telling strategic choice.
While many companies are still trying to figure out where to place an AI sparkle icon inside their product UI, there is already a large and immediate opportunity inside the operating layer of the business. Reporting, internal knowledge access, workflow coordination, task routing, and executive summaries are all areas where AI can create practical value right now.
That does not make product AI unimportant. It just suggests that the easier wins may still be happening behind the scenes.
Final Takeaway
This episode is most valuable because it resists the usual extremes.
It is not an overhyped vision of autonomous AI replacing everything. And it is not a skeptical dismissal of the entire category either.
Instead, Micah and Damien outline a middle path: use AI agents where they can make decisions inside bounded workflows, use RAG to connect them to real company knowledge, use platforms like N8n to move faster, and avoid the trap of assuming tool adoption alone equals transformation.
For teams trying to make sense of the AI tooling landscape, that is a useful lens.
The winning organizations may not be the ones with the flashiest demos. They may be the ones that quietly learn how to turn small, well-designed automations into repeatable systems that the business can actually rely on.