What is an AI workflow?

The first thing we do when searching for information, for an answer to this question, is probably ask an LLM. And maybe these very words became part of its answer, and you are not reading this on my blog but on some LLM screen. The answer goes like this:

“An AI workflow (artificial intelligence workflow) is a sequence of steps that uses artificial intelligence to complete a task.

It can involve gathering information, processing it, generating an output, and passing that output to the next step.”

Yes, it’s a definition, and a definition gives us a starting point. It puts the idea into words. But this is the part we skip because we already know those words: AI, work, and flow. We get it. Getting things done by something not human.

But when you asked the LLM this question, what you were really, really looking for is: “How are we actually building one?”

But an LLM will start by giving the definition because that is how it will check whether you’ve stuck around, meaning it succeeded in reading your intent and understanding what you wanted.

The next part of the answer goes something like this…

The practical side: how to build an AI workflow

After searching online, Alex (my ChatGPT assistant) brought back guides from Zapier, Make, and n8n because that is what we are looking for.

Its answers are no different from the ones you will get if you ask online how to build an AI workflow. :)

So, we went ahead and summarised them for you so you wouldn’t have to. In the best possible way. In a table. Because let’s be honest, everything looks better at a glance, and a table is how you do it:

The guides cover these details across their advice on prompting, building, and testing workflows. Zapier also addresses choosing how much control you want during the build.

And because you landed here, Alex (my ChatGPT assistant) made an example for you, so you can see those pieces together in a use case:

“When a customer enquiry arrives in Gmail with the label ‘Product questions,’ use the product information in the Google Drive folder ‘Approved product information’ to draft a reply. Save it as a Gmail draft for me to review. If the information needed to answer is missing, flag the enquiry for me. Start with five enquiries so we can check the results.”

We have a goal, a starting point, tools, information, conditions, and a place for us to review the work.

A shout-out to the original guides. Thank you for giving us a common base of understanding to work with. :)

• Zapier: Best practices for using Copilot
• n8n: AI Workflow Builder best practices
• Make: Maia best practices

Is that all your bot needs to know?

We now have the practical steps for building a workflow. We have a formula, a set of parameters. We add some details, and the workflow should operate. So it seems…

Let’s stay with our customer enquiry example. We gave the bot access to product information. Is that all the information it needs?

An AI workflow enters an operating business environment. There is already a lot of intelligence behind how that business works. Even a small task, like answering a customer enquiry, involves more than technical information.

The people doing that work know things. Some they learned through training, some through experience. They know which questions to ask, what customers tend to misunderstand, and when an answer needs to become a conversation.

They also know how your business handles those situations. What have we promised this customer? What have they already tried? What does taking care of them mean here?

When we give the bot a technical document and ask it to answer enquiries, how much of that knowledge have we actually shared?

And how much are we expecting it to work out on its own?

This is where context engineering and prompt engineering come into the picture.

Context engineering involves recognising the information the bot needs and making it available when it needs it. The product information is one part. Your way of serving customers, the conversation history and the commitments you have made are also part of what helps it understand the situation.

Prompt engineering is how we communicate what we want it to do with that understanding. What should it look for before preparing a reply? What should it ask when something is unclear? When should it bring someone else into the conversation?

And even then, the words we use need attention. We might ask the bot to be “helpful.” But what does helpful mean in your business?

That is where we need to look a little closer. :)

What are we missing?

A workflow is a thought process. When we do the work ourselves, we connect pieces of information to make a decision. But we are not always aware of which information helps us make a smarter decision.

An organisation holds a collective understanding made of pieces and bits of information. When connected, they help people understand what the right answer is beyond the technical information they can find. This is how we should view an AI workflow too.

I know what you are going to say. Most customer questions are repetitive and trivial. Let’s say the workflow covers the basics in 80% of cases, provided your technical documentation is relevant and up to date.

I get that. But we are still looking at a workflow as one task. Once we understand the bigger picture, that a workflow is a thought process, we can use AI to our benefit and implement it in ways that are more effective, less time consuming and able to scale.

Let’s continue with our example to demonstrate what I mean by bigger picture thinking.

We have set up a workflow that, in our example, should work 80% of the time. To keep it working at that level, we will need to continue updating the information it uses.

Now think about what that involves. Even knowing what needs updating is a thought process. Where does that information come from? Who recognises that something has changed? How does that change reach the bot so we can maintain that 80%, without waiting to see a drop?

Someone working on the product knows a feature has changed. Someone serving customers notices a question that keeps coming up. How do we connect what they know to what the bot needs to know? Could identifying and preparing those updates become part of the workflow too?

We started with answering customer enquiries. Looking at the bigger picture helped us recognise another connected piece of work that could help the bot continue doing its job.

This is an example of what I am asking you to think about. Did changing your perspective help you notice something you hadn’t considered?

We are beginning to see that the bot is part of an ecosystem. From this perspective, we can think about what role we want it to play and how its connections to people, information and other processes could benefit the organisation in the long run, with less maintenance and more room to scale. That is part of context engineering too.

What other possibilities become visible when we view a workflow as part of the organisation’s collective understanding? What else could we connect, share or develop together?

If we keep our focus narrow, we risk carrying our existing way of thinking about automation into AI, without nurturing what AI has to offer.

Personal note

When we work with AI, we are collaborating on a thought level with a technology that can think. Using this perspective opens our minds to what it needs. It goes beyond the technology to the bigger picture of co-creating between human and AI resources.

Thinking of a workflow as a thought process, a collaboration we are trying to create, helps us see how people and AI can work together in whatever capacity you and your organisation are ready for.

When we look at an AI workflow this way, we begin to see more of what we can do with it. We notice the information, experience and connections that will help AI do the work.

At AI Thinking, we help you do exactly that through our research and context engineering services, and through prompting, where we help your bots do what they need to do. :) We can create bots for you, or you can just read my book, the one that will help you wire your brain to see the thought process itself. :)

Need help? Let’s talk.

Back to Thought Documentary