The Workflow That Builds Itself

Building Agentic Workflows through Conversation

Last summer we built an internal AI chatbot that made our organization’s scattered knowledge easily accessible. The system has been running for months now, still restricted to a small group of people within my organization, but it has become something that they find genuinely transformative to their daily work.

As I watched people use it, I could not help but feel like we could do more. The information retrieval aspect was working. People were getting answers faster that previously required sifting through pages of documents and pinging their colleagues. Here we had an AI that understood enough about our work and can figure out how to use the available information to get things done. And yet, why are we still asking people to do grunt work themselves?

Directing vs. Delegating

There’s an important distinction between AI chatbots and AI agents that I think is still underappreciated outside technical circles. When you use a chatbot, you are directing. You ask a question, it gives you an answer, you decide what to do and you ask another question. The human remains in the driver’s seat at every step. The AI is like a very knowledgeable passenger who never touches the wheel.

An AI agent is different. You are delegating. You describe what you want accomplished, and the AI helps you break it down, what steps to take, what to do when something goes wrong. The human moves from a driver to more like a manager overseeing a capable employee.

This sounds like a small shift, but in practice it is enormous. When you direct, you need to know what questions to ask. You need to stay engaged at every step. But when you delegate, you just need to know what outcome you want, and then you can walk away and focus on something else.

The Tooling Problem

I have been thinking about the crop of agentic AI platforms marketed to the enterprise. Most have converged on the same interaction model, but I’m not convinced it’s the right one.

They’ve decided that the way to build agentic workflows is through visual pipeline builders. You start with a canvas, you drag boxes, you draw lines connecting them, you configure each box’s parameters through property panels. It’s “low code”, sure. They claim even non-technical users can build sophisticated workflows easily.

Except…can they really?

I recently sat through an all-day session where senior leaders from my organization worked with a vendor to prototype an automation. We had a specific workflow in mind, one with real labor savings if we could pull it off. The people in the room were genuine experts on this process. They were people who’d been doing the work for years and knew every edge case.

By the end of the day, we hadn’t made much progress. This was not because the platform couldn’t handle it, and it was not because anyone lacked expertise. The problem was that the visual builder required a completely different kind of thinking than the work itself required.

The thing about these low-code platforms is that they don’t eliminate complexity. They just move it somewhere else. Dragging boxes is more approachable than writing code, sure. But you still need to understand the boxes, what should go in each box, what should come out, the branching logic, error handling, and a dozen other concepts that are fundamentally about system design. When it comes down to it, you’ve only traded one kind of expertise for another.

To be clear, I’m not arguing that we throw out visual builders, but there’s a real irony here. We’ve built AI systems capable of remarkable reasoning and planning: large language models that can reason about complex problems, and plan multi-step solutions. And yet, we’re asking humans to do all that design work manually. By hand. By dragging little boxes around a screen. Why?

Just Tell It What You Want

The alternative is hard to ignore once you see it. Instead of making users design the workflow, let the AI design it based on what they actually want to accomplish.

I know, I know. I can almost hear the objection: AI isn’t advanced enough to figure all this out on its own. But hear me out.

Everyone already knows how to use a chatbot. We’ve spent the last few years training billions of people to interact with AI through natural language. They know they can describe what they want in plain English, and the AI will try to help. Why not just extend that? Instead of starting from a graphical paradigm, why not just describe your goal in the same conversational way you’d describe anything else?

“I need to review all the pull requests from last week and check if they have corresponding work items and specific acceptance criteria.”

“Generate user stories from this requirements document and make sure each one has proper acceptance criteria before adding them to our backlog.”

In practice, current AI models are actually quite good at this kind of planning, if you let them. It works with you to figure out what steps might be needed to get it done. And if it doesn’t nail it the first time, you refine together, same as you would with a human colleague.

You might be surprised at how effective AI can be at planning complex work. Not always, of course. But humans have cognitive biases around workflow design. We tend to model processes linearly, the way we’d do them ourselves. The AI doesn’t have the same constraints. It can approach problems in ways we wouldn’t naturally consider.

This flexibility also exposes a fundamental limitation of visual builders. They only work well when you can define the steps in advance, when the workflow is a known quantity. But AI-native workflows open up a new possibility: the task structure can be discovered or negotiated at runtime rather than predefined. You can’t draw a box for a step that doesn’t exist yet.

Try drawing a canvas workflow for drafting a user story until it actually meets your Definition of Ready. How many revision cycles do you put in the diagram? You don’t know. Maybe the first draft nails it. Maybe it takes four passes before the acceptance criteria are solid. The agent figures that out as it goes, checks what’s missing, fixes it, checks again. The number of steps wasn’t decided in advance. It was decided by the work.

Building Through Conversation

So, if the workflow isn’t predefined, how does it come together? Through conversation.

In practice, the best approach I’ve found is iterative. You describe what you want. The AI proposes an approach. You review it, adjust it by prompting, and then you let it execute. You review the results. You refine. The workflow emerges through dialogue, not design.

This shouldn’t be surprising. It’s how humans work together too. No one walks into a meeting with a colleague, silently hands them a flowchart, and walks away. You talk it through. You say “actually, let’s try it this way instead.” You build shared understanding incrementally.

This is especially important because AI can produce a plan that sounds completely reasonable and still be wrong in ways that aren’t obvious. Its output can be so fluent and confident, it’s easier to miss than a misconfigured node in a pipeline. The fix isn’t complicated: you have to actually read what it produces, not just glance at it. Human review is a must.

The execution log, the record of what the AI actually did, becomes a powerful tool in this process. You spot inefficiencies you’d never have anticipated. You learn how the AI interprets your requests. And because it’s all accessible right in the conversation, you can ask questions about it: “Why did this step take so long?” or “Could we have done these two things at the same time?” The AI explains its reasoning, you suggest changes, and the next run is better.

This cycle — execute, review, refine — turns workflow building into something that feels less like engineering and more like coaching. You start with a rough description, let the AI take a first pass, and collaboratively improve it based on what you actually see. Each iteration teaches both of you: the AI learns what you actually meant, and you learn what’s possible.

The Human’s New Role

All of this points to something bigger than a better way to build workflows. It changes what the human’s role in the work actually is. When AI is doing most of grunt work, it doesn’t mean humans become less relevant — quite the opposite.

Returning to the earlier analogy, a good manager doesn’t just hand off tasks and disappear entirely. They set the direction. They define what success looks like. They catch when something is technically done but actually wrong. They bring the judgment, the tacit domain knowledge, and the lived experience that no amount of AI model parameters can substitute for.

That’s what working with AI agents actually asks of you. Not less thinking, but different thinking. Less time on the how, more on the what and the why. In some ways, it raises the bar. When the tedious work is handled, there’s nowhere left to hide. The quality of your judgment becomes the thing.

To be fair, not everyone is equally prepared for that to be the only thing on the table. I don’t know yet how most people will navigate that shift. Some will take to it quickly. But others have built their identity around doing grunt work and will find this disorienting, not because the work got harder, but because the familiar parts disappeared.

Maybe that’s a worthwhile trade. I think it probably is, but I’d be skeptical of anyone who tells you it’s a sure thing.