How to build AI orchestration

AI Agents have seeped into every aspect of our lives whether you like it or not. We literally have Mary Poppins’ endless bag of gadgets available to us. 

‍

How do you sort it out?

‍

We’ve entered the age of “Build Anything” (my newsletter name finally makes sense). We just need a bit of guidance on how to get from point A to point B. 

‍

Do we use Automation? AI? AI in an automation? An AI Agent? A sandwich?

‍

Today, I’ll address the confusion and create some clarity around what to use and when. 

‍

My focus is on internal workflows and tools. If you’re looking to build your own software company, this post is not for you. In other words, we’re Vibe Building not Vibe Coding.

‍

We’ll be focusing on business operations.

‍

There’s a million and one ways to skin a workflow. After this, you’ll know which way is best for whatever problem you’re solving. 

‍

Let’s dive in!

‍

‍

Introduction: Definitions

Everyone has a slightly different take on defining AI Automation and all of its permutations. I’ll ground mine in Zapier so that they’re tied to the tools that make them a reality. Using Zapier should help clarify how I define the following:

Automation (Traditional Zap): a set of determined steps in a workflow that happen automatically in a Zap and use any combination of the nearly 8,000 app integrations in Zapier’s ecosystem. 

AI Automation (AI Zap): an automation but with an AI step using an LLM like ChatGPT, Claude, Gemini, etc.

Agent: a workflow that infers how to accomplish a task and has access to tools like web search, data analysis, and app integrations. 

Agentic Zap: an automation with an AI Agent step.

‍

With the introduction of the Agentic Zap, you can now combine a determined set of automated steps with an AI Agent that infers the steps it needs to get a job done. 

Not to make matters more confusing, but you can even accomplish what you’d want in a traditional automation but by using the Zapier Agents product. An Agent that can handle classic automation? Yeah. 

So this begs the question and brings us to the meat of this post: When would you actually need to use each of the four different types of automation? 

Let’s explore each.

‍

Automation (Deterministic using traditional Zaps)

The most important reason for using a traditional Zap to build an automation is Accuracy. 

You achieve accuracy at the expense of speed-to-build and flexibility. Accounting and finance use cases come to mind. You don’t want to leave anything up for interpretation when it comes to calculating figures. When a customer pays the invoice, you want your accounting system to update systematically every single time. 

The key word here is Deterministic which means stripping the automation of its free will. 

You heard that right. We’re getting philosophical. Buckle up. 

But no, it’s really that simple. A Zap allows you to build a little world where everything that happens is caused by something else, and it couldn't have happened any other way. 

At this point you may feel like a little-”g” god. 

Your work’s not done, yet.

‍

AI Automation (Deterministic Zap with AI step)

The most important reason for using an AI Zap to build an automation is still Accuracy. You just need some AI power.

With an AI Zap you use the power of Large Language Models (LLMs) to do certain things within your automated workflow like:

• Entity extraction (e.g. pull out names, dates, amounts). Keeps your Zap from having to guess which bits matter.
• Data transformation (paraphrase, normalize casing/formatting, translate). Instead of wrestling with Formatter steps.
• Summarization (condense long emails or docs into a bullet-point digest)
• Content generation (draft personalized replies, social posts, subject lines)
• Classification & routing (spam vs. legit, topic tagging, sentiment/intent detection to drive Paths or Filters)
• Data enrichment (append metadata like customer sentiment, intent, or industry from unstructured text)
• Validation & cleanup (grammar/spell-check, compliance flags, profanity filters)

Now, let’s jump to Agents.

Agents (Inferential using Zapier Agents) 

The most important reason for using an Agent to build an automation is increased Speed-to-Build and Data Analysis.

With Agents, you type out what you want it to do and give it the access it needs and that’s it. It can be supersonic to go from zero to one.

Let’s skydive out of the clouds here and make this tangible.

Below are the two features that illustrate the Agent advantage within Zapier.

1. Increased Speed-to-Build: “Let your agent generate/select a value for this field”

Zapier has twelve years of data on how people integrate apps. With that data, Agents has intelligently learned how to map fields. If there’s a field called “Company Name,” Agents is smart enough to know that whatever information it receives or finds that looks like a company name will need to be placed in the company name field.

It’s not rocket science, but it’s definitely some intelligent computer science. 

Imagine you give an Agent a company like “Zapier” and you want it to fill in fields in a table for:

• The rival schools of the alma maters of the founders
• The latest podcast episode the CEO was on and it’s link
• How many days until the CEO’s birthday

An Agent can generate and map the values for these fields simply by you having the fields named in your Table without you needing to bake in the logic and process for finding and evaluating the information.

Agents give you intelligent SPEED.

That brings us to the second Agent advantage: Data.

2. Data Analysis: give Agents large amounts of data to analyze or use in making decisions and they’ll analyze it in seconds. 

For instance, you can plug-in all of your past Zoom recordings and an Agent will have all of the details and transcripts readily available to use at its disposal. 

Can’t remember which customer you discussed a potential AI hackathon with last week? Just ask the Agent.

Want to compare how you pitched your product in each of the five sales calls you had this week?

Have the Agent send you weekly reports breaking it down. 

Even without your own data, Agents are helpful. But if you plug in your own data, Agents are downright powerful, contextual, and fully leveraged.

I’m lifting the veil and showing you something: Agents are best used within larger Agent systems. 

Next, we’ll see how Agentic workflows help build these interconnected systems.

‍

Agentic Workflows

Most people are eager to use “autonomous agents” but are rightfully cautious about the wild beasts that they are. 

So with Agentic workflows you get to put a saddle and bit in the Agent’s mouth. You take an Agent and surround it with the traditional steps in a deterministic Zap and in doing so you increase trust and reliability. 

A powerful combination.

Since Agents are so powerful with data, accessing and analyzing data is how you should think about their role within a Zap.

Every time you get a lead, have the Agent research the lead, summarize any touchpoints you’ve had with people from the lead’s company, and based on your past interactions, or lack thereof, go down one of two paths.

All this AI is great, but the last thing you need is one billion options without a clear direction. It’s like when someone says go buy a bamboo toothbrush on Amazon. God bless. 

Too many options.

So, let me tell you how to choose each approach with AI.

‍

Choosing the right AI approach

Use this framework to make a decision on when to use one of the four types of automation.

Traditional Automation

• Goal: accuracy, repeatability
• Repetitive, rule-based, no interpretation
• Zaps

You’ve got a process and you want the automation to stick to it. No deviations! 

‍

AI Automation

• Goal: accuracy, processing
• Repetitive, rule-basd, includes interpretation/transformation step
• Zaps with an AI step

You’ve got a process that involves some text analysis or transformation and may benefit from knowing things that the entire internet probably knows (AI). But it’s still a set process. 

‍

Agentic Workflows

• Goal: control, data-analysis 
• Repetitive, rule-based, with data analysis step
• Zaps with an Agents step

You’ve got a process that involves a step where large data analysis comes in handy. There may also be some decisions and actions that happen outside of the workflow, but you let the agent make that call. You still have a controlled process to follow after the Agent does its thing.

‍

AI Agents

• Goal: speed, decision-making
• Repetitive, loose instructions, data analysis, decision-making
• Agents

You’ve decided a loose process that could involve data analysis and deciding which action to take and/or you’re prioritizing building speed over accuracy. Likely requires feedback and iterations to fine-tune over time.

‍

Still not clear? Practical examples are coming to the rescue. Let’s take a look.

‍

Practical examples and templates

Let’s run through a few real examples. First up: automation. 

‍

Automation

Your classic Zaps. One of the most used use cases across all customers at Zapier is some version of an onboarding Zap. It could be a new hire or a new client. But usually there are a handful of things that need to happen to get that new person officially onboarded. 

It’s repeatable. The process is more or less the same for every new person.

It’s rule based. Every step follows a set procedure.

There’s no interpretation. You are typically giving someone access or adding them to events which look like a checkbox for the automation to complete.

You can steal the employee onboarding template I made using Tables and Zaps to see this in action.

‍

AI Automation

Take a Zap and throw in an AI by Zapier step. AI by Zapier gives you an incredible prompt wizard with your choice of LLM. So, let’s take a look at that.

A use case we see all the time at Zapier is to use AI to categorize things. This could be new feedback you receive from a customer or a new message intended for a particular team.

Put AI in the mailroom and it will get things sorted.

‍

Agentic Automation

This is the newest form of automation on the block. It’s so new, that I don’t have any fresh templates for you, yet!

But here’s where it’s the most powerful (at this point I should sound like a broken record):

Data. 

So if you want to add data analysis to your Zap, add an Agent. 

Maybe you’re building a Zap to help auto-respond to questions your team ask you in Slack. And then you think, “oh hey, what if I had a huge FAQ list and and an Agent could just go find the right answer?”

Agents, baby. 

‍

Agents

Did I mention how analyzing data is the biggest advantage of Agents? Oh, yeah, I definitely did. Okay, but what else? Is it just data? 

Well, no, it’s more. With Agents, you can bake in some decision making. The control-freaks in the room are squirming. I understand. So, start simple. 

When you build an agent to handle new leads, put a line in there that says, “If you think this lead is particularly important, send me a Slack DM.”

Low risk. Allows you to test the waters.

Or maybe you have an agent that analyzes sales call transcripts from Zoom. It gives you the breakdown and critique. Put a line in there that says, “If you think there was a stand-out moment from this call where the sales rep went above and beyond, give them a quick shout-out in our Slack channel.”

Low risk. Allows you to test the waters.

Ultimately, though, you’ll find that delegating decision making to an agent is risky business unless you give the agent contextual data. Again, data. 

Give your agent context about you, your company, your goals, your decision-making frameworks, etc. and you’ll go further with Agents than you thought possible. 

Need some inspiration? Check out Zapier Agents templates.

‍

Wrapping it up: Orchestrated systems

If you’ve read this far, congratulations. You know more about AI automation and AI Agents than 99% of the world. This stuff is so bleeding edge it’s wild. So hear me when I say you can start simple and you’ll still be ahead of the game.

‍

But. 

‍

If you want to time travel into the future, then you’ll want to start building orchestrated systems. 

‍

Here’s what I mean:

‍

1. The data sources I’ve been raving about. They need to be automated. You need Zaps getting the data and updating it in a Zapier Table. That’s step one. 
2. You might need a form or interface to collect and show data. Build it in Zapier Interfaces. Step two. 
3. You may need to interact with the data via Slack. Build a Zap to handle that. Step three.
4. Then you have an Agent doing the heavy lifting, drafting weekly reports, and helping you analyze data. That’s step four. 
5. You map it all out in Canvas.

‍

Marcus from Remote built the IT Helpdesk and saves $500,000 per year.

‍

That’s AI orchestration. That’s the future. We’re building mini-kingdoms for these Agents to work in. They are digital but they have roofs, roads, draw-bridges, guards, and libraries. 

‍

For decades, only a select minority could build like this.

‍

Now you can, too. 

‍

That’s all for today!

‍

Happy Building,

Bryce