Who Owns the Workflow? The Organizational Problem AI Actually Exposes

Paul Tyler (00:02.686) Hi, this is Paul Tyler with another episode of the L and A Hub. And this is gonna be a fun episode. You know, great guests, but even I'm gonna say better topic. I have with me Professor Wei Chen from the University of Connecticut. Wei, welcome.

Wei Chen (00:20.782) Hi Paul. Thank you for having me here.

Paul Tyler (00:23.522) Yeah, can you tell us a little bit about what do you do? You know, who are you and what do you do? And give us a little bit of your backstory as well.

Wei Chen (00:32.09) Sure, so I'm an associate professor from the University of Connecticut in the School of Business. So I've been a researcher for more than fifteen years in the School of Business doing research on different topics.

But since three years ago when I moved to University of Connecticut, I started to interact with a lot of industry folks. I was co-leading the Digital Frontiers Initiative from the University of Connecticut and want to make my research more relevant. And so I made a mission of myself to make Gen AI accessible for the last three years, taught graduate courses, wrote a textbook, and now

I've written a business novel about how Gen AI will change your work.

Paul Tyler (01:23.394) Yeah, this is great. And you know, by the way, you're a great lecturer, great professor. I think I'm trying to think of maybe Harvard. I think I maybe sat in on one of your lectures around—it was a Gen AI workshop, a lot of code, a lot of—yeah, so how does somebody move from writing Jupyter notebooks to writing a novel? How?

Wei Chen (01:39.268) Yeah.

Paul Tyler (01:51.03) What did you do with it? I notice it says that you used AI to help you actually write the book.

Wei Chen (01:56.601) Yes, and so the idea is simple. So like we want to make Gen AI accessible to everyone and I've been doing those workshops for executives, for staff members, but like what I felt is it's like the scale of influence is quite small. People don't have big picture of how Gen AI works and how would that

have an organizational consequence. So when I talked to my co-author Fassen about the idea of writing this novel, he immediately got it. Because in operations management, there has been a very famous book called The Goal. It's about process management in manufacturing, on the manufacturing floor. So they started with this one, then in the IT industry there is the DevOps book, the Phoenix Project.

So we modeled after these two books and we thought there has to be something like that for the Gen AI era. And it's one possibility of the future that we try to imagine through.

Paul Tyler (03:09.462) Yeah, no, this is great. And spoiler alert for everybody, the book actually involves an insurance company and a manufacturing company. So there's actually insurance in this, which is great for our topic. I think in your framework you mentioned—and I think before we jump into this, which I think is boy, apropos today—AI changes tasks.

Okay, workflow needs owners. So talk to us just a little bit about what does that mean and how did you get there?

Wei Chen (03:44.282) So that actually comes from my teaching. So I was teaching a graduate class on Gen AI for business. And the more I taught about this, I started to think about what our students are facing.

In industry when they graduate and get into their jobs, what they are facing, what the future organizations have to be like so that we can prepare them better for the future. So we started to think about the organization part. Then I realized like there is one very

like important missing link in this whole picture. Like AI automates a lot of things, but it's more on the task level. Even the research is on the task level, decomposition of jobs into tasks and what AI can automate. That's from all the like leading publications and Anthropic reports, right? It's about specific tasks. But

In organizations, we also have the other extreme. Board has a directive to executives. Hey, you need to have an AI strategy. And people start to allocate resources, running pilots. That doesn't really transform the organization. So I think what's missing in the middle is this workflow that connects the tasks into the organization so that we can really change the organization in the AI era.

Paul Tyler (05:10.796) Yeah. And I'm sure we didn't talk about this before, but you know, Anthropic put out a paper, I think it was maybe a study a couple of months ago on industries and which ones they thought it would impact and which ones it already had. You know, I'm assuming you've looked at that—we didn't ask before. Was that a task-oriented look at AI or do you think that was a workflow?

Wei Chen (05:28.195) Yes.

Wei Chen (05:35.874) I think Anthropic might be one of the organizations that's experiencing this and like leading this change. But I think that perspective is still on the task level. This follows from like the Science paper on GPTs and GPTs, right? GPTs are general purpose technologies. That's a very long academic tradition of decomposing work, jobs into tasks.

So like and then like the papers look at how AI is going to change all the tasks. So for like that Anthropic one, I think it's still on the task level.

Paul Tyler (06:17.378) So is it possible to underestimate—I mean, it projects an enormous impact on the workforce. Could even be bigger. Yeah.

Wei Chen (06:26.209) Yes. So my idea is all knowledge work that we are doing today is going to be changed by generative AI. And you can see that from people, right? Like with the twenty billion dollar subscription money that OpenAI gets every year, like seventy percent of that is from individuals who pay twenty bucks a month for that. So that's

Paul Tyler (06:51.744) It's incredible. It's incredible.

Wei Chen (06:53.455) Right, so the adoption is there. So it's just a matter of time of how organizations are going to change.

Paul Tyler (06:59.062) Yeah. Well let's talk about your TWO framework. Do you wanna kinda walk through that? T is for task, W is for workflow, O is for owner, correct?

Wei Chen (07:07.939) Yes.

Wei Chen (07:11.351) Yeah, yeah. So that's what I've been trying to kind of

promote with the business novel. That novel is kind of a body to contain the idea behind this. So like as I said, I think the workflow, so we have the task level that's the lowest tier, we have the workflow that's the middle layer, and the organization which is the top tier. So for tasks we have done quite a lot of research. So we know that. We know the limitation of the current AI, and it's changing very quickly. But

The more AI can do, the longer workflow we're going to give them. So there is no end, right? We're going to give them more and more complicated jobs. So it's still going to be our workflow and they are not going to be perfect. And

Now there are some other problems that's created by these long workflows. When you have a very long workflow, like fixing a bug or implementing a feature in software, and when the outcome gets to the human, it's going to be very hard to review what's generated. Right? So that we call it the validation tax in the book. And

Like what what I realized in real life when I was working with co-authors, working with my students, I see more and more

Wei Chen (08:34.965) of those presentations generated by Claude Code, big fonts, small fonts, very tiny lines of fonts at the bottom. And I also fell into that trap. I found myself in those meetings start to read off the slides. Right? Because I don't have enough time to really evaluate everything and that becomes reading off the slides. And sometimes when my co-author sends me those summaries of papers

that's formatted very well in themes. And like I was like, when I ask them what's the main point of this? They have no idea what's in there. So it looks good, but now right? Leave the job to us to do the review. So like I think it's very common—I experienced this. Those are not bad people. They are very smart people.

Paul Tyler (09:12.302) I have lots of stories that would get me in trouble if I shared them.

Wei Chen (09:32.365) doing very good job, but all of us are falling into this trap of like let the AI generate, but we don't have enough capacity to check it.

Paul Tyler (09:44.142) No, it's insane. But we'll come back to some of those problems that you raise here. I guess maybe we could drill just a little bit into ownership because this actually ties back to literally a discussion I was having with somebody about two hours before now. And this is kind of a question—it's a deeper question I think than it might sound—but like if I'm you know, the vice president of claims or vice president of underwriting or

Wei Chen (09:54.691) Yep.

Paul Tyler (10:13.206) the customer service organization, what does it really mean to walk into a room and say, well, who owns this workflow? 'Cause there's a lot of subtext there. A lot. Like you know, am I asking—is this who owns the workflow, who owns the output? Is it who actually, as you say, owns reviewing the steps along the way? How do you think about owning a workflow?

Wei Chen (10:23.737) Mm-hmm.

Paul Tyler (10:43.084) Is it the beginning and the end? Is it all the pieces in between? How do you do this in a matrix organization?

Wei Chen (10:51.577) So we did this in the book in different steps. At first

the people in the story thought it was a problem of drafting, drafting the denial letter, right? For claims processing. That's drafting. But it's not. It's actually a workflow. So they quickly realized that there are knowledge retrieval steps, the claims classification, what is this claim about? And then make decisions, pull like different information, like historical information about the client.

And then decide how much we're going to allocate for that claim. And then if it's a severe one, maybe book the hotel for the clients to stay out of their home for an extended period of time. So it becomes a workflow. And now if you put people at the end, it's going to like they say they are going to own it.

And if something goes wrong, they are responsible for that. But it's very hard for them to check this non-process like every step there. Because that if if something goes wrong in the middle and they check at the end, it's very hard for them to check. So we started to think about like what what are we

What do we have to do to make this really work? The conclusion is we have to make the intermediate steps checkable by human. So the so-called human in the loop has to have a concrete form. So the information has to be concise enough or structured enough to do this. So then this owner—at first they were just ensuring the outcome is correct, but now it's not just the outcome. They have to be responsible for

Wei Chen (12:45.115) for this AI-human system. So divided into four steps. When a junior comes in, they are responsible for the first one. It's teachable because the output is structured. And it also becomes a job description for the junior. And when they are more experienced, then they move to more steps or own the outcome of the whole workflow.

make sure the outcome is correct and improve the AI system. And the owner of the whole workflow or the architect of the workflow has to change and design this. And then later we discuss like now what it means in a matrix kind of organization. They have to be the glue of that workflow to other parts of the organization.

Paul Tyler (13:32.342) Yeah. Well, you know, to just make it real for me, you know, as I say just a couple of hours ago I was having—and it was a nice friendly debate, but it was, you know, a debate. Who owns the website? I can't believe it's twenty twenty six and we're still having an argument. And you know, I think they think I'm being a little difficult, but ownership's interesting, right? Is it authority? Hey, listen, I have the rights to push this thing out or I have the rights to change this. Is it accountability, you know, hey, if the—

Wei Chen (13:40.644) Yeah.

Wei Chen (13:55.673) Yeah, yeah.

Paul Tyler (14:01.644) You know what, we violated Canadian law, you know, we have a Canadian operation. Guess what? I have to have everything translated to French because we've got a presence there. So there's some rules there. So like if it doesn't work, I'm accountable, okay. Domain expertise, you know, is it who's got the experience or is it all three or am I missing parts?

Wei Chen (14:04.718) Yep.

Wei Chen (14:28.429) Yeah, so this is a great question. In our book we have one scenario. So when they figure out this workflow model, they started to pitch to their clients. Because the main organization is the IT consulting company, right? So they start to pitch to people. And they go to their clients saying, We need—because this owner has to be accountable, so they have to have that authority.

Now you need to give this owner authority across three departments on things they are managing. And the CFOs will be saying, You are crazy. I'm installing a tool and you are asking me to give one person the authority across three departments, changing the organization before I do anything. Right? So then this is crazy. So then they realize the model is okay, but the way we pitch it is wrong.

You have to start with the entry workflow where one person has the authority to do the whole thing. So and they can just sponsor the whole change. Once that workflow is working, you can add another one. People will look at that and say, that's working. Well, I want that as well. So you then you add another one, you add another one that gets into the organization.

Paul Tyler (15:47.011) Yeah. Well, so you know, we do a lot of—you know, Zinnia does a lot of very complex, deep inside a company's software like policy admin systems. You know, we have a platform for policy admin. Martin Cole, you know, I think was looking for a—you know, everything is a platform to him. Now

Wei Chen (16:02.316) Yeah, yeah, yeah.

Paul Tyler (16:11.318) And I would say probably that feels like eighty percent of the customers we see. Now, I did see a very, very large—you know, everybody would know this company's name—they sell life and annuities, a couple years ago where they actually put RFPs out. This is really interesting—for value stream improvements. This really was a pre-AI way. So they said, okay, you own

agent productivity. You own, you know, customer experience. It cut all the way across the organization. And I will say it was disruptive. What I saw, just a lot of executive change. I even saw change in terms of the companies, the consulting companies driving this thing. Do you think that's more the future direction? Are we gonna have to live through that kind of discomfort to successfully deploy AI in a lot of companies?

Wei Chen (17:06.424) So I think we will go through profound organization changes, but probably we can do it less disruptive.

So we have an analogy of the old digital transformation kind of method. It's like it's coming from the old ERP systems, right? So when when you want to implement the ERP within a company, it's like replacing the full jet engines over a flying plane, right? Because it's cutting through every department and it's going to be big change, very disruptive.

That's because the old software is quite rigid. Software is written in rules, logic rules, right? So that's what what's happening there. But for

AI, maybe we can do one jet engine replacement at a time so we don't affect the other three engines. We finish one workflow here, replace one engine, and they say that this engine is better. Let's replace our second one or third one. And one important job of the owner is becoming the person who

glue the replacing jet engine with the other ones. So if executives want to go around them, the owner or the sponsor, the executive sponsor of the owner has to fund them off. Because the as soon as someone can go around it, everybody will go around that.

Paul Tyler (18:41.774) Well let's talk about—sometimes that resistance is like deep in the organization. So I think there's one scene, I think it was chapter six. Okay, you have a character named Angela. I think she's a claims adjuster, right? And she has this—Yeah. Go ahead, yeah.

Wei Chen (18:48.034) Yep.

Wei Chen (18:57.302) Yeah, yeah, yeah. She's the operations manager, claims operations manager, like two ranks down below the chief claims officer.

Paul Tyler (19:08.63) Right, yeah. So it says, okay, let me see my notes if I got this right. So she has this correction. If it ever mentions section four B, delete immediately, do not let it use the word regrettably. You know, I guess because that's gonna set them up for a litigation suit. It makes people sue us. So I guess why does that exist? And what does that tell you?

Like from an organization if people've got this stuff off to the side.

Wei Chen (19:40.025) So for that, I think there are two things. One is about the nature of generative AI. So if we don't break it down into steps, if we only look at the final outcome.

We are inevitably having something that's hallucinated. So that's made up. Either it's the words that's made up or the reference that's made up. Or it just makes mistakes. It's not perfect. It's a probabilistic machine, right? So for that one, that's why they had these binders to check it, to make sure humans are the last gate to ensure the quality of things.

But our second point we want to make there is—this is also happening in programming. People start to use AI to code, at first they check every line of code.

Then they check every three times. Then they check every five times. Then they were like, okay, the AI said the test has passed. Let's push. Right? So it's happening at all the big tech and that's what happened to Amazon, right? They had a very public email about a pretty big blast radius of their code base because of AI. Now because people get used to this and then one incident happens, people turn back their trust.

Paul Tyler (20:51.278) Yeah.

Wei Chen (21:09.54) They don't trust the system anymore. And if they have to check like that, the time it takes to check the results might be even longer than just writing it from scratch. So then adoption becomes a theater and not really something that's useful.

Paul Tyler (21:29.07) Yeah. Listen, I fall into that trap. It's like yeah, yeah, and Claude Code. Would you like to review an edit? Yes. Do I have permission to this? Yes, yes, yes, yes, yes, yes. Run. It's easy. There's something in our head. I guess we get comfortable. You get comfortable driving a car. You start to trust the machine. It's gotta be something psychological here, right, that we—

Wei Chen (21:41.836) Yes.

Yes.

Wei Chen (21:49.101) Yes.

Wei Chen (21:52.462) Mm-hmm.

Wei Chen (21:56.204) Yes.

Paul Tyler (21:56.484) Or some sort of bias, to accept something after a couple of things have happened without a problem.

Wei Chen (22:03.254) Yes, yes. That's that's a nature. That's the nature of human beings, right? So that's how our brain functions.

Paul Tyler (22:09.804) Yeah. Well talk to me a little bit about this—I'll tell people the story about—we'll go to the other one, the regulation later, but the Illinois regulation feature. Maybe set that up to paint the scene and we can dive into that one a little bit.

Wei Chen (22:29.484) Sure, so that's when the company has implemented RAG.

And they were retrieving documents to evaluate claims, to decide which policy applies and then decide the policy in there. And what happens in that scenario is the system retrieved a document that's wrong. Because in Illinois some lawsuits have changed the policy so there is a code memo that's in the system.

But named differently than the usual file names. So the system didn't retrieve it and gave the client a wrong deadline to claim. So it becomes a non-compliance issue for the company and may affect the license of the company. So what we are trying to show there is the system can function perfectly. But when the underlying knowledge

that the system relies on has no owner. No one ensures the latest information is there, the system is always functioning correctly—it will break down. You can have an AI system that's working for three months, but after that, after the implementer leaves, the performance of the system starts to drift, drift, drift, drift. So you need people to manage the system.

and the outcome at the same time and they have to be internal people who are who are managing this.

Paul Tyler (24:07.778) I think in that scene you have something like twenty, I don't know, twenty five or twenty six people and the boss says, Who's accountable, who's responsible? And nobody raises their hand. I guess, you know, is that the point you're making? Let me back this up. What would have had to happen to get one person to raise their hand?

Wei Chen (24:18.816) Yes.

Wei Chen (24:27.913) Mm-hmm.

Wei Chen (24:33.261) So for that I think the person has to have the accountability. So if anything happens with this, they are accountable and they have to have the authority to change things. Because if they are accountable, they have to be able to stop the whole process when

something that's not right happens. So even if it's blocking the whole workflow of the company, they have to have the authority to stop that and say, okay, this is not correct. Let's stop for half a day, let's fix that and then move on. Right. So that's something that usually we don't have in modern organizations. Because why do we have those organizations in the first place? That's what we have been thinking about in our research. That is—

We try to create an organization to reduce the frictions we have, which we cannot do in market. So we use authorities, the hierarchies to deal with that. But a lot of that is using humans as a glue to deal with those frictions, right? And now with AI,

Like the mistakes in the old days with humans, our mistakes are easy to correct. We can correct them because it's slow. But with AI, in one afternoon you can file 400 claims or deal with like 500 claims, right? Then the mistakes are going to be much bigger. And basically it made those implicit organizational frictions explicit. It exposed all the frictions we had before.

And now we have to find a way to deal with that.

Paul Tyler (26:19.318) Yeah. Well let's stay on the regulation topic. I think there was another scene where the regulation changed. And I'm telling you, insurance—I was talking to somebody about a contract, you know, and a friend of mine who's an administrative law judge, he's using some of the AI and talking to me about the problems of summarizing some of these rules. I was telling him about insurance contracts because you know, you ask a question about like a twenty page contract—

You know, the answer really is dependent on this sentence with this comma on the first page. You have to know this clause with a colon in the middle, and then you've got to know something else at the bottom. And probabilistic work or probabilistic math doesn't always separate the comma from the period. So I guess—

How do you deal with this in the insurance arena?

Wei Chen (27:21.601) So what do we imagine? So I have to admit this is not a widespread industry trend yet because we were not satisfied with what's like in reality. Because for now, everything for those large language models, everything stays within the context window. Right? The agent carries the context window forward and everything is implicit. You stuff everything in there.

But to check specific things like whether that clause in a contract is correct, you need structured output on the specific location of that clause. So if that clause is very important that determines the good or bad of our contract, that means the AI has to pick that one out.

And write it down into a simple—we call it the context packet. So maybe it's just five items, and one of them is that clause, where on that page, on what page that one is, and then our evaluator can easily check where that one is and check that original text. It's similar with financial statements, right? With the financial statement, you have a lot of numbers. A lot of times we have to calculate ratios. But if it didn't get the right number,

from the right table in a like a 10-K, everything downstream will be calculated wrong. So that step to output—okay, these two numbers are from page 8 and page 13, and we use these two numbers to calculate the following ones—has to be checked by a human and make sure that's correct. And

make sure the whole process works. So that's what we imagined about what would happen in the future. In our research we call this the intermediate states of the workflow. So that might be a key to make it really reliable.

Paul Tyler (29:30.444) Yeah, you know, it's interesting. It's almost that in some ways it's a knowledge management problem. And knowledge management problems in this industry predated AI. And I think you mentioned—you used the word human glue. I saw a lot of instances in carriers where I've worked where, you know, the person who'd worked in the company for twenty years was that glue. They just knew those regs or they knew those approaches. Underwriting's a great area where you probably have a

guide—underwriting guide you give your agents and it's very thin. You don't really tell what you're doing, you're making them guess. And you've got like an underwriting manual that you've given to the reinsurers, but you know, there's a lot of play in the middle and you've got these unstated rules of a company that kind of live in people's heads. How should that change? And how should companies rethink

those rules or knowledge that are sitting in heads but not really written down someplace.

Wei Chen (30:31.767) I would say there is no easy path. So we just have to work through these workflows like what's happening in the book. Because all these things—there cannot be like one size fits all.

solutions because every company does this differently. The person who is responsible for this is different. So for example in one thing in the book, our experienced adjuster rejected our claim.

And according to the AI and all the standards, it seems to be a perfectly fine one. And like they asked that adjuster what happened here, and he said, this person had a claim like last year, and I was there. I was on his driveway. I told him that yes, we are going to pay for the roof, and you have to find a quality contractor, not your brother.

in order to fix that roof. He didn't. Now like eight months later, he's making a claim again. So that's the personal experience that people—this is never going to be written down.

in our knowledge base or somewhere, right? So it's that work experience. So that's why we still need people there. And we still need people to grow from a junior adjuster to a senior adjuster to people who manage the whole workflow, who architect the workflow. So for that I think it's still human work that we have to go through.

Paul Tyler (32:17.186) Yeah, and I guess when the humans are there and the humans are involved, how do the reactions to risk change? Like I think in that book, the compliance team when they find that wrong case, okay, the compliance person wants to just shut everything down, turn the switches off. The operations team says, No, let's keep going. I guess five years before, when this is a

programmatic thing versus this which is AI, who wins? Who should win? Today?

Wei Chen (32:58.838) That's a great question. I don't think I have a great answer for that one, but

I think the answer will be the same. It still depends on the on the organization where the AI is working in. So the the organization has to be ready to make that changes. So in in that thing that you described, the compliance wanted to stop it and review every claim, but they only have four people. They cannot deal with like one hundred twenty claims every day. So

And the operations want to process it because if they are processing too slowly they get a warning from the insurance department. So there is always these different goals from different parts of the organization that have different incentives. When they have different incentives, they have different objectives about the AI system. And if there is no single owner of this workflow,

This becomes quite fragmented and no one is going to be responsible for the outcome.

Paul Tyler (34:13.88) Yeah, you know, I could make the case and say, you know, it's probabilistic. What are the odds? Let's keep moving. On the other hand, you could say, Wow, if it's making this mistake, it's gonna just multiply and we can't stop it because it's creating so much. So I don't know. Work is changing a lot. Well—

Wei Chen (34:29.834) Yes, yes, right.

Wei Chen (34:36.033) Yeah.

Wei Chen (34:39.454) Yes, yes. Yeah. In one of our papers what we were trying to show is when the AI system mainly does automation, it will actually drive the company to hire more capable people, like so-called upskilling. Because now AI is doing five times the job of people. So it makes five times the mistakes.

And you need capable people to manage that. So that's why you see companies move to hire more experienced workers from the market.

Paul Tyler (35:01.89) Ha ha ha

I love it.

Paul Tyler (35:13.836) That's great. Well, maybe we could kind of flip the script a little bit and you know, you talk about what's broken. What does an AI-native organization look like? Like, you know, what is the operating model? What does it look like or feel like?

Wei Chen (35:31.127) So I think this is a topic that many people are interested in and I've listened to some of the like Y Combinator talks. They are promoting this AI-native startup idea but it's still a system-centric thinking.

If you like—that's Silicon Valley, right? So Silicon Valley is always going to focus on the IT part of things. And they focus on that. But what I think from our book is—

It cannot be just the system. There has to be the people, the organization, the institution that's around the AI system, the AI workflow. So it's a combination of AI and human system. Why? Because if you look at two things, if you look at human history, intelligence is never a constraint for human society.

We had very smart people, very high IQ. But usually they hit a wall in society when they work with normal people because this is a society of normal people. And like AI can do a lot of things, but I don't think that's changing the institutional side of our society. The second one is—for

AI—for IT in the last 20 years, we have platform companies. They have the data, they manage the system. We can do nothing but work with their platform. Like as researchers, we look from outside and evaluate things. But AI is a tool, right? AI is a tool that Anthropic employees have, the same as us. And how

Wei Chen (37:32.918) How do we expect that they are going to do our job better with the same tool? It's like an electrical drill. Yes, they are the manufacturer. But I can be very creative in how I use the electrical drill.

Paul Tyler (37:47.247) I love this. I love it. Well, I guess if you had to predict one industry that gets it right in the next three years, you know, given what you said, which one is it? Which one will like change, morph, and really make the most out of the new tech?

Wei Chen (38:06.304) Well that's a tough question. It's always dangerous to predict the future. But I think software is definitely one of the industries that's moving fast, that's breaking things. So that's from what we see on the junior developer market—things change very quickly. I think that's one industry that's already undergoing significant changes and they are trying to figure out how the incentives are going to work.

But for other industries, I think

Like it's just going to have to come slow because there will be pilots and people will realize it's not that easy when you want to scale. So like that's why in our book we have this line, like one workflow at a time, repeatedly, because we thought back and forth and felt that might be the only way going forward for a lot of industries that have

like high stakes, right? If it's a mistake and you get your license cancelled in one state, that's a no-go.

Paul Tyler (39:08.424) Right.

Paul Tyler (39:16.684) Yeah. Well, you know, actually listen, I'm testing that with a design partner right now. You know, I think agentic software will really change how we support insurance agents, period. But do you do it all at once or just pick one spot? We're picking one spot, one workflow. Now I think when we look at this it will be a complex workflow. You know, the whole case design, case illustration system because

Wei Chen (39:30.569) Mm-hmm. Mm-hmm.

Wei Chen (39:41.366) Yes.

Paul Tyler (39:45.72) That kind of is the core of life insurance. So we'll see.

Wei Chen (39:49.984) Yeah. So that

Exactly, exactly. So I think that's the right approach. That's why in our teaching we're changing the approach. In the past two years, I've been having those lectures and hands-on activities for people so that they get an idea of what's prompt engineering, what's RAG, what's agentic systems. But now I'm changing to a very hands-on process. So we pick one workflow. For example, in a workshop that I—

We did this Excel to PowerPoint pitch, like kind of a workflow. So they get some data, they generate figures, they have one recommendation, make that into a PowerPoint. Right? You can see this is a very natural workflow. At first they do it with Copilot, just

throw in the Excel, generate the PowerPoint. Of course it's crap. And then they break it down and then do it step by step and even have team members doing different tasks. And then they run again. Now the results are much better. They use that to pitch to the executives to see the result. And for the workshops that we are designing for the School of Business, I'm teaching a Gen AI workflow for business class in the spring. We want them to do another workflow from their work. So pick one.

Paul Tyler (40:45.549) Right.

Wei Chen (41:09.471) For example, I was working with a company, they are checking whether their design blueprints are conforming to the building code, which is 800 pages.

Paul Tyler (41:18.74) Interesting. Yeah. Okay. That's a project.

Wei Chen (41:26.525) Yes, yes. So I'm designing a hands-on learning experience with them so that they can just do this themselves and see the difference. So I think for every type of work we have to go through this and people who are doing the work have to own their work. It's not like an IT company comes in and solves their problem for them. I don't think that's going to happen.

Paul Tyler (41:49.261) Yeah. I don't think so either. But hey, listen, it was great to have you on here. I guess I'll put a link to the book. People can purchase this. What else would you like them to know if they want to reach out and either, you know, talk to you, use you for consulting, take one of your classes? How should they reach you?

Wei Chen (42:01.291) Thank you.

Wei Chen (42:14.549) So they can just reach me from our website, the TWO Project.com, or just add me on LinkedIn, Wei Chen from UConn. So just reach out to me and chat with me. I'm very happy to talk. And I really appreciate the real feedback that people have from their real work.

Paul Tyler (42:39.918) That's great. Listen, it was—like I say, a reading that was a little uncomfortable because I've lived some of these scenes myself more than I'd care to admit. But anyway, well listen, thanks so much. Thanks to our listeners. Please leave us a rating or review. Really important. Helps us share the content with more people. And most importantly, be sure to tune in next week for another great episode of the L&A Hub. Thanks. Thanks, Wei.

Wei Chen (43:08.533) Thank you, Paul.