Katherine Forrest: Hello, everyone, and welcome to “Paul, Weiss Waking Up With AI.” I’m Katherine Forrest, and I am flying solo again. Anna is going to be back. I promise you, I promise you. I know we all miss Anna. She’s going to be back, but today I’m going to flap my wings and fly solo again. So I am going to talk about, in a moment, what’s going to be— or could be, if I didn’t allow it to be—a boring but important topic, which is AI workflows. Because we hear the word workflows all the time right now in the AI world. And so, you know, Anna and I thought we have got to deal with the term “workflows” and explain to our audience what the term “AI workflows” means so that when they see it, like on every billboard and every advertisement, they see it in the literature that’s coming around at your particular organization, that they have some concept as to what it means, if anything, that’s special. But I’m going to start with sort of a homespun example, which is that workflows is really any kind of task that you’re undertaking that follows an ordered process. And so the ordered process that I’m going to talk about is one that I am currently very involved in, which is my personal workflow with my wife in writing the final, truly final—I mean, I know I’ve been talking about this for a while—but final, final parts of this book that we’re co-authoring called “Of Another Mind.” So I’m going to talk about that workflow first. We’re going to use that sort of kind of analog workflow, if you will, to then talk about AI workflows.
So I’ve got this book under contract that’s supposed to be delivered. I’m on a, maybe we’re on like the second extension. Maybe it’s the third extension. Who’s counting? In any event, you know, books are harder to write than articles because there are so many pieces to them. But the workflow of this book, which is called “Of Another Mind,” and by the way, if you’re wondering what it’s about, I’ll just tell you so you can sort of hear about it. It’s about the evolution of AI capabilities as we move towards AGI and superintelligence and the implications for our social contract that flow from that. So that’s the book, but the workflow goes a little bit like this, and you have to multiply this times two because it’s a co-authored book with my wife and so we do this all together. But, you know, first you have the idea and then you have the outline. So let’s just call that sort of a step one, you know, sort of idea reduced to outline. And then you’ve got step two, and we’re putting aside now the fact that the publisher accepts it on the outline. Well, for us they did because we’ve done some books before. The second step is the research step, and that is extensive and it’s very long. And the research itself, that whole workflow, has sub-workflows because that step in the process actually has multiple steps to be sure that you’ve got the right literature, you’ve got the right academic articles, you have the right version of the right academic articles, you’ve got the right books, etc., etc. And so that takes up every waking minute that you have when you’re not doing the job that actually earns you money to keep the lights on.
And then there’s the next step in the workflow, which is putting pen to paper or your fingers to the keyboard, and that’s its own, you know, little horror show when you hit writer’s block or not. We actually have desks that are in the same room where we talk back and forth constantly. We’re writing different chapters at the same time, but that workflow sort of is ongoing. And then there’s the editing workflow, you know, portion of the workflow, I should say. And then after that, you go back and there’s a second round of editing. But because it’s AI, there’s a constant round of researching because everything is changing. And so it could drive you crazy, but that’s the workflow. And then it goes off to the publisher, which will happen in just a couple of weeks.
So that’s a workflow. It’s an ordered process. And so it’s step by step. And we can take that concept, that sort of analog, you know, homespun concept and turn it into an AI workflow, which is that term that we’re hearing all the time right now. And we can transfer the phrase workflow into the business world. And workflows are obviously everywhere. Basically, for people who are doing your job, you’ll do multiple workflows a day. And let’s just say you’ve got somebody who’s in the bookkeeping department and they’re approving expenses and they’ve now got certain kinds of software programs that help them with approving expenses. And they go through a particular process and they check it against the guidelines, they check it for an approval after it’s been submitted and they, you know, yada, yada, yada, and eventually then the money, if it’s approved, would then get disbursed in any particular format that the company has decided.
But let’s now add on AI and talk about doing a workflow with AI. What does that mean? Well, first of all, it means that you’re taking a task and you’re breaking it down into ordered logical steps. And there are tools now, AI tools, that are being marketed as particularly good for particular workflows. So if you had a bookkeeper who used to do from A to Z to take expenses to get them approved and to get the money disbursed, you might now have an AI tool that actually has taken each of those steps in that process and has now included them in the AI logic and will accomplish that particular workflow. So you’ve got multiple AI tools that can do multiple different workflows within an organization, and some tools can do multiple workflows. And then of course, just like I was talking about with research where my research process has sort of workflows within workflows, you know, circles within circles within circles or whatever the phrase is, you can have a workflow that can actually have multiple sub-steps. And so the AI tool can do that too.
So let’s take an example. Let’s take a human resources leave approval workflow, right? That might just be one workflow that a human resources department has. And let’s assume for the moment that they’ve actually got some AI tools that are now being integrated to do some of these HR workflows. So the AI tool might take in a request, where you enter a request for leave. It might then compare it to the guidelines, let’s say it is slightly adjacent to what the guidelines say. It might analyze whether or not approval is likely, it might then send it for approval. Then it would come back, and then it would go through a series of steps to both notify you and notify anyone else who needs to be notified that your approval has been granted. So it can be repeated, it’s predictable and it’s logical. And so that’s all an AI workflow really is. And so for a regular AI tool, it’s relatively deterministic. And frankly, you didn’t need generative AI to do this. There were a lot of workflows that could occur with narrow AI tools. So they’re much more complex now in the world of GenAI, but they follow a set path and they are smart and they can do things that are different than just a software or a service. They can actually do a variety of smart, learned tasks as they go along.
Now, let’s add another layer to this because a lot of the workflows you see today being advertised have the word “agent workflow” in front of it. There are two ways in which the word “agent” is actually being used with workflow, confusingly. One is that sometimes there are vendors who are calling an AI workflow tool an agent workflow tool, not because it’s agentic AI at all, which we’ll talk about in a moment, but it’s not agentic autonomous AI, but that the agent is taking the place of a person. It’s acting as your agent, as your person who is tasked to do something. So it’s not agentic AI with particular agentic capabilities. It’s just an agent who’s acting at your request. So there are agent AI workflows being marketed that are not agentic. And so this is something to sort of keep in the back of your mind when you see a workflow tool being marketed and it says agent AI workflow. Is it agentic, is it not agentic? And we’ll come back to that.
But let’s take the concept of agentic AI workflow tools because those agentic AI workflow tools are also now increasingly available from vendors. And they can actually take on many tasks, but they can do them not necessarily at all in a deterministic way. The point of agentic AI with the autonomy that an AI agent—the agentic capabilities—is that they can actually be flexible. They can problem solve. They can take a left-hand turn when they need to take a left-hand turn. They can actually take tools from a toolbox that’s internal to the model, to help them actually accomplish a particular task. So if you’ve got an agentic tool, then the agentic tool may be using multiple steps that are not predetermined for that agent to accomplish what it needs to accomplish for you. And that might be with more complicated workflows where the workflow requires problem solving as you go along. It’s not going to be an ordered expense reimbursement or an ordered leave request. It might be a workflow that requires analysis and comparison of multiple documents all at the same time. Also, internal to that, making sure that certain numbers are correct, etc., etc. You can add in a lot to what could be the job of someone, call it an agentic workflow, it would be broken down into multiple tasks. But as people in complicated jobs know, you might have different ways of approaching a problem depending upon the nature and the characteristics of a problem. So the agentic workflow tool can actually then take a non-deterministic approach to these kinds of problems and do them autonomously and do them flexibly.
Now, what are the benefits of these workflow tools? Well, for those of you who are responsible for AI in your organization or are just interested in it or are in compliance, one of the things that a workflow tool can typically do is, number one, it can explain the steps to you, if it’s a non-agentic AI tool. If it’s an agentic AI tool, it can usually describe the types of steps that are available with that tool: that it could take A, it could take B—sort of a decision tree. But the steps are understood. And that can be very, very useful from a compliance point of view. The second thing that happens and flows from that is that the steps can be auditable, that there’s a trail left behind with what steps the actual model will take. So if you’ve got a deterministic AI model, it can leave one kind of audit trail. If it’s an agentic tool and it’s going through the decision tree, it can tell you what it did. And also, you can get certain tools today which can leave also their chain of thought behind, which can tell you why they did what they did. And that’s very important from a compliance—why did certain things happen—point of view and from an audit point of view.
So let’s just take two examples of why this would be useful: finance and healthcare. And you can imagine that in those two areas, there are going to be a number of jobs and tasks where you need to have an auditable trail left behind because you’ve got regulatory reasons why you can’t just be taking steps that you can’t figure out in a black box why they happened. So let’s just take an investment summarization for a financial services company. An AI workflow might use natural language processing to review documents, to summarize them and to compare them all in ordered steps governed by certain kinds of rules, as well as permissions. And in the healthcare area, let’s just say that you’ve got a triage workflow; well, you may then have a whole series of ordered steps that occur that are then audited and understood and go one by one by one where you could actually then look and see on the patient’s record exactly what happened. So a clear workflow can be a compliance professional’s best friend because the steps can be documented. And if it’s not agentic, they can be predictable. If it’s agentic, then hopefully what you’ve got is a decision tree where they can be predictable within certain boundaries.
But let’s just sort of pause for a moment on what a person who is involved in compliance with AI would want to know. Number one, you want to know with your vendor, if you’re taking in a tool that purports to do workflows, what the series of steps are that that tool is going to be undertaking. You want to understand what the record is that’s going to be produced, particularly if you need an audit trail. And you need, if it’s got agentic capabilities—not that it’s just your agent and the word agent being used in just sort of an easy way as a replacement for a person, but agentic capabilities. You want to know how that decision tree works, if you actually need to have a clear audit trail because you may have things that that agent can do and things you do not want that agent to do.
And so you want to understand what that decision tree is, and then you want to understand the trail of breadcrumbs it’s going to leave behind for you in terms of the chain of thought and also its footprint of how it’s made its decisions, its agentic decision-making footprint. So the takeaway there is that workflow is just doing a job, it’s doing a series of tasks. You do want to know if an AI workflow is agentic or non-agentic because it actually makes a difference in terms of whether or not the steps are determined in advance or not. And that’s really all there is to it. So the word “workflow” is used a lot right now to just put a term to doing the job of somebody else or a series of tasks that somebody as a human may have done previously. And you may now have an adjacent job that’s actually monitoring it, testing it, assisting it and learning certain things. But that’s really all there is to it.
So there we are for today. I hope I did not bore you to tears, but I think it’s an important concept that we all sort of demystify. And I want you to know that I am going to finish this book. And then I’m going to announce victory on this podcast, and my own personal workflow will have been completed. So we’re really close to it. It won’t be done by the next episode that we record, but hopefully not too long after that. All right, folks, have a terrific week. And I am Katherine Forrest, and I’ll hope that you tune in, that you tell every single one of your friends to like and subscribe.
