For years, the norm has been creating multi-step workflows that stitch together APIs and automate tasks.

In enterprise, this is called Robotic Process Automation (RPA). In SMB, this is called workflow automation with popular tools like Zapier, Make, n8n, and Pipedream. All major CRMs have their own internal workflow builders too, such as SFDC, GHL, and Zoho.

But building and maintaining hardcoded workflows is slow, brittle, involving a lot of repetitive code and logic to handle dozens of possibilities and edge case.

AI agents, in contrast, let you hook up tools ONCE, then simply tell your agent what you need in plain English.

💡 The AI agent dynamically plans which actions to take, in what order, and even recovers from errors… WITHOUT manually hardcoding every edge case.

Understanding the leap from the old way of doing things, i.e. rigid workflow automations, to AI agents is key if you want faster, more adaptable, more powerful AI automations.

Here’s the Youtube version of this post:

Real Use Case

In this post, we’ll walk through 3 real use cases, processing customer billing requests:

• cancel a customer’s subscription and refund last payment

• apply promo code to customer’s active subscription

• cancel a customer’s Incomplete/Expired subscription and create new subscription

I use a tool called Intercom for support (btw I HATE its workflow builder), where users can click a button to indicate which use case they need help with. Then, Intercom calls a webhook in n8n/Make to trigger the corresponding automation.

With Workflow Automations…

To cancel a customer’s subscription and refund their last payment, typically you’d build a workflow automation with 5 distinct steps, each corresponding to a required API call:

• fetch customer’s Stripe subscription ID

• cancel subscription

• fetch customer’s last charge ID

• refund charge

• ChatGPT summarize changes made, to be used in reply to customer

To apply a promotion code to a customer’s subscription, you’d need to build another workflow automation or add branching logic to your current workflow. You need the following steps:

• fetch customer’s Stripe subscription ID

• apply promotion code

• ChatGPT summarize changes made, to be used in reply to customer

Let’s say you want to cancel a customer’s Expired or Incomplete subscription, then restart their free trial subscription. Here are the steps required:

• fetch customer’s Stripe subscription ID with status “Incomplete” or “Expired”

• delete subscription that has status “Incomplete” or “Expired”

• create new Stripe subscription

• ChatGPT summarize changes made, to be used in reply to customer

You could build these 3 distinct workflows within the same MEGA workflow with branching logic, or keep them in separate workflows.

In the sample screenshot below, each row corresponds to one of the use cases above.

But the Stripe API is extensive.

This means there are many, many, many COMBINATIONS of actions you can take.

Therein lies a weakness of traditional workflow automations:

You (developer) must hardcode each workflow path and handle edge cases manually.

With AI Agent…

Now let’s compare it with my n8n AI Agent to manage Stripe billing requests.

I only have to hook up the API calls once, then give the AI Agent tasks in plain English.

The AI Agent plans what actions to take to accomplish the task, using the tools at hand.

In our Stripe example, each distinct API call is a “tool” and the AI Agent dynamically:

• analyzes the task

• plans what to do next, deciding which tools to use (i.e. which API calls to make) and HOW to orchestrate them together, in a reasonable sequence

• executes the next action

• reflects on the result of each action (e.g. if something failed, should it retry the same action or try something different?)

• repeat for next action

For example, the AI agent will use “Apply Promo” tool (API call) to apply a discount if that’s relevant to the task. But likely won’t need to use “Refund Charge” for that task.

Here’s a REAL LIVE example:

The only instruction that Intercom sends to my n8n AI agent:

cancel subscription and refund last charge

Here’s my AI Agent’s prompt, and on the left side, you can see all the actions it decided to take, in order to accomplish the task “cancel subscription and refund last charge”:

And, here are the actions the AI Agent executed to accomplish the task “apply promo code to active subscription”:

The key point?

After equipping my AI Agent with tools (i.e. Stripe API calls), I simply tell the agent what I want, WITHOUT manually stitching together multi-step workflows and repetitive API calls.

THAT’S the power and promise of AI agents!!

And, here’s the beautiful summary produced by my AI Agent, explaining the actions it took to accomplish the task. This summary is sent directly to the customer, customized for each use, again WITHOUT having to hardcode message templates:

In this real-world example, I only equipped my AI Agent with Stripe tools.

Now, imagine giving your AI Agent access to multiple services (e.g. CRM, email, support, Google Search, etc.) so that it can perform complex action sequences, across multiple disparate systems

🤯🤯🤯 

Mindblowing, I know!!

Best Practices

You’ll hear the term “guardrails" often in relation to LLMs.

Due to the probabilistic nature of outputs, running the same prompt or agentic flow 100 times may yield different results.

The purpose of guardrails is to say “only do these actions” or “always ask a human before proceeding with high-risk decisions.”

That way, your AI agent won’t run wild or make costly mistakes.

In my customer support billing example, which deals with highly sensitive data and sometimes angry customers, my primary guardrail is controlling what instructions are given to the AI Agent.

Instead of allowing customers to directly chat free-form with my AI Agent and say anything, I have a list of preapproved instructions that can be passed to my AI Agent.

Challenges in Production

Overall, it’s still VERY EARLY days for AI Agents in production deployments.

There aren’t many big companies, outside of FAANG and top startups, that have deployed agents successfully in production, at scale.

Here are some key challenges impeding the broader adoption of AI agents, which I won’t dive into detail in this post but want to at least mention:

• Reliability & Hallucinations: LLMs can invent API calls, make wrong decisions, and misinterpret messages. Even when not hallucinating, LLM outputs are stochastic in nature, so you’re not even guaranteed the same output when running the same prompt(s). If your automation requires high-reliability and poses significant business risk if something goes wrong, a non-agentic traditional workflow is the safer choice.

• Cost & Performance: Every extra API call adds latency and cost. Traditional workflows have more predictable costs and performance, especially at scale.

• Security: Giving AI Agents read/write access to your business’s critical systems? Yeah… NO. Most companies are quite far from that 🤣 

• Interoperability: This is the key challenge that Model Context Protocol (MCP) tackles. Think of it like a universal interface that enables AI Agents to connect securely to many services without writing connectors/calls for every endpoint.

Need More Help? 👋 

1/ If you want to grow on social media and build a business in coaching, consulting, speaking, selling apps, or digital products… check out Blotato

2/ Free AI courses & playbooks here