# Where Agentic AI Pays Off in the B2B Back Office - with Chris Bradley of Veritiv

- **Show:** The AI in Business Podcast
- **Published:** 1 Sep 2026
- **Source:** https://podcast.emerj.com/where-agentic-ai-pays-off-in-the-b2b-back-office-with-chris-bradley-of-veritiv
- **Transcript:** https://podcasts.opennash.com/shows/ai-in-business/where-agentic-ai-pays-off-in-the-b2b-back-office-with-chris-bradley-of-veritiv-791194/
- **Provider:** OpenNash Podcasts (machine-generated)

## In brief

Agentic AI pays off most in high-volume back-office work, not in tools that help people think. Chris Bradley of Veritiv argues that real agents complete a task from start to finish. He says the best first target is repetitive manual work done by many people, such as the 1.7 million orders Veritiv keys by hand each year, because freeing a large group from drudgery matters more than a small, tidy pilot.

The tension is between trusting an agent and verifying it. Large language models can behave differently from one call to the next, and problems only appeared once volume reached the thousands. Veritiv's answer was a screenshot audit trail for every step and a second agent that checks the first one's work. Bradley also says adoption depends on handing people the grunt work, not the creative work they enjoy.

## Key insights

1. **Agentic AI means finishing the task, not advising** Bradley says many tools labelled agents are still generative AI helpers that draft or analyse for a human to act on. True agentic AI takes a goal and performs every step a human used to do.
2. **Legacy system patchwork blocks agentic AI** Most AI tooling is built for AI-native environments, but large B2B firms run a mix of ERP and homegrown systems, with staff hopping between several monitors. That patchwork keeps agents out of the work.
3. **Screenshots at every step make autonomy trustworthy** Veritiv records a replayable screenshot at each step of every agent mission, so any order can be searched and reviewed. Bradley notes this gives more visibility than the company has for humans doing the same work.
4. **Agent errors appear only at higher volume** The order agent worked well at low volume, but anomalies showed up once it handled thousands of orders. Each agent call is effectively a fresh model, so Veritiv added a second agent to inspect each run.
5. **Pick the process with the most identical jobs** Veritiv used an HR list of jobs across about 8,000 employees to find roles with hundreds of people doing the same work, like customer service. Bradley argues saving a few hours a day for hundreds beats helping two people.

## Questions this episode answers

### What is the difference between agentic AI and the AI chatbots most people use?

Bradley says chatbots like ChatGPT mostly help you draft or think, and a human still does the next step. Agentic AI is given a goal and completes every step of the task itself.

### How can a company trust an AI agent to process orders without anyone watching?

Veritiv records a screenshot at every step of each agent task, so anyone can replay what happened and find out why something went wrong. Bradley says this gave the company the confidence to let the agent run autonomously.

### Why did Veritiv add a second AI agent to check the first one?

The order agent worked well at low volume, but strange errors appeared once it handled thousands of orders, because each AI call starts as a fresh model. So Veritiv built another agent that inspects each run at the end, and Bradley says language models like judging other language models.

### Should a business start with a small AI pilot or something big?

Bradley argues against the smallest pilot. He recommends the process where the most people do the same manual job, because the effort of building an agent only pays off if it saves time for hundreds of people.

### How do you get employees to actually adopt AI tools?

Bradley says to automate the grunt work people feel is beneath them, not the creative work they enjoy. Veritiv also kept the interface simple: staff assign cases to the agent inside Salesforce, the customer management system they already use.

## Facts

- **Host** [0:00:12]: "Today's guest is Chris Bradley, Chief Marketing Officer and Head of AI Transformation at Veritiv."
- **Chris Bradley** [0:06:46]: "order entry within our company, where we manually key 1.7 million orders a year"
- **Chris Bradley** [0:07:17]: "we did kind of our first test of that back in probably late November, early December of 2025"
- **Chris Bradley** [0:08:08]: "every single agentic mission that we deploy, there is a screenshot at every step. And you can play it back to be able to see exactly what it it did."
- **Chris Bradley** [0:08:08]: "Because, you know, L L Ms are non-deterministic, it can do something unexpected sometimes."
- **Chris Bradley** [0:09:08]: "we don't have screen recordings of every time a human does anything on a computer, but we do with agents"
- **Chris Bradley** [0:10:05]: "If it takes twenty different agent call or like large language calls per session, you're actually getting a fresh model every time."
- **Chris Bradley** [0:10:32]: "because it was operating really well at low volume. But once we got to, like, over thousands of orders, then all of a sudden you start to see these anomalies."
- **Chris Bradley** [0:11:02]: "So now we have to actually build another agent that would inspect at the very end of an agent run to make sure that the agent did what we expected it to do."
- **Chris Bradley** [0:11:40]: "they will sign a case and then they will have case comments back and forth with the agent who's actually doing the work on their behalf."
- **Chris Bradley** [0:14:13]: "there's two primary types of leverage that you can get with A I. So you can get decision leverage, where it's really helping you do better analysis."
- **Chris Bradley** [0:14:13]: "We deliberately picked the latter, and it's it's tempting to pick the the other one first."
- **Chris Bradley** [0:14:59]: "I started with our chief H R officer and I got him to give me a list of all of the jobs for like our 8,000 employees."
- **Chris Bradley** [0:14:59]: "we found, like customer service, financial operations, sales, where you literally could have five, 600 people that all have the same job type."
- **Chris Bradley** [0:16:50]: "in all of the big transformations I've done, even prior to A I, the biggest thing that you is always gonna hold you back is gonna be change management."
- **Chris Bradley** [0:03:38]: "how do we use agentic tools to take some of those manual processes and really figure out how to use computer use models"
- **Chris Bradley** [0:02:34]: "most of us live in this world with a patchwork of legacy systems, E R P systems, homegrown systems. And you have a large number of users who have two or three or even I see in our company sometimes four monitors."
- **Chris Bradley** [0:06:15]: "True agentic A I has to be, here's what needs to get done, and the A I goes and do does all the steps that the human was the only one that could do that previously."
- **Chris Bradley** [0:05:05]: "a lot of A I, and even the leadership in A I, usually comes from the A I, M L, analytics, B I functions at most companies."

## Transcript

**[0:00:12] Host:** Welcome everyone to the Emerj AI in Business Podcast. Today's guest is Chris Bradley, Chief Marketing Officer and Head of AI Transformation at Veritiv. Chris walks through why high volume back office processes like order entry still run manually inside large B2B enterprises. And what it took to hand one of those processes, an order entry system processing 1.7 million orders a year, fully over to an AI agent. A quick note for our audience that the views and opinions expressed by Chris Bradley on today's program are his own and do not reflect those of Veritiv or its leadership.

**[0:00:47] Host:** If you offer AI products or services into the enterprise, you need to find enterprise leaders with relevance and readiness. Emerj attracts V P and higher-ending enterprise audiences who are already convinced that they need to move beyond traditional I T. To learn the exact strategies we use to help leading I. A brands and startups connect with their ideal enterprise A. I. buyers, visit Emerj dot com slash a d one. That's e m e r j dot com slash a d one.

**[0:01:14] Host:** Now the conversation with Chris. Chris, welcome to the other side of the mic.

**[0:01:40] Chris Bradley:** Thank you so much. It's good to be here. Appreciate being invited on.

**[0:01:44] Host:** Yeah, so our conversation today is very much about where agentic A I pays off in the B to B back office. And thinking about how much noise there is at the moment about A I transforming every part of the business. But when we look closely at big B to B operations, it's almost as if so much of the day-to-day operations, like capturing an order into a system or closing out the books. It still runs on the same manual processes that we have been following for the last five years, right?

**[0:02:16] Host:** And it's almost as if the transformation conversation and what's actually happening in real life are two very different things. So I'm curious, uh, why does so much of that work still happen manually? Why is it still the old processes, even with all the A I investment out there?

**[0:02:34] Chris Bradley:** Great question. I think one of the challenges that I see, especially with legacy systems, is most of the A I work that in A I transformation software and and tools that people are building are really A I native. But the reality is most of us live in this world with a patchwork of legacy systems, E R P systems, homegrown systems. And you have a large number of users who have two or three or even I see in our company sometimes four monitors.

**[0:03:04] Chris Bradley:** And they're going from system to system to get different pieces of data. And that patchwork effect is really keeping a lot of the potential for agentic A I from some large B to B enterprises.

**[0:03:18] Host:** That's interesting, Anna. You're saying that as I'm sitting here with my three monitors. So what is the real cost of keeping these things manual? I'm I'm thinking about, you know, obviously time, um, human error, all of those things. If you have to think about it, is it more costly to keep it manually than just to bite the bullet and get that A I investment?

**[0:03:38] Chris Bradley:** I think it is, and we're also forcing our employees to really do this manual work that they could be elevated to do such higher level work. So that I am. That's why even I would say, you know, within our own company, we have looked at how do we use agentic tools to take some of those manual processes and really figure out how to use computer use models and some of the new capabilities that exist to identify which wasn't possible just five or six months ago.

**[0:04:11] Host:** Yeah, and you're taking me right to my next question is about the capabilities and knowing how to and when to use it. If if I listen to A I conversations, especially in, you know, when we're having B to B type A I conversations, everything always points to helping people think better, right? Having better research, uh, better first drafts, better analysis. But that's not the same as the A I doing the work itself from start to finish, uh, without shepherding any part of the or any any step of the the process.

**[0:04:44] Host:** So those also feel like two very different categories of tools, having A I just help you out versus having A I do everything. But it's it's labeled the same, both just A I tools. Where do you draw the line? Uh, what changes once A I moves from just helping someone to decide to a point where it just handles the entire task?

**[0:05:05] Chris Bradley:** I think that's really what truly differentiates agentic A I from gen A I. And there's so many tools out there that they all have agent names, which are these fancy titles, but they're actually still gen A I tools. So I think the genesis of, um, A I with chatbots of things like ChatGpt were always helping to draft, helping to be a thought partner, helping to to think. And I would also say that a lot of A I, and even the leadership in A I, usually comes from the A I, M L, analytics, B I functions at most companies.

**[0:05:44] Chris Bradley:** And I think the the promise of actually being able to do real operational work, that's a very different type of A I capability that i is is hard for, I think, a a lot of people to make the jump. And so the way that I like to look at it is when you're actually directing A I with a goal, and it's going all the way through and completing that task end to end. It's not just helping you think through an analysis or doing the analysis for you so that the human can now do the next step.

**[0:06:15] Chris Bradley:** True agentic A I has to be, here's what needs to get done, and the A I goes and do does all the steps that the human was the only one that could do that previously.

**[0:06:26] Host:** I'm gonna get personal uh with you for a moment and now you can share as much as you can without spilling any secrets. But I'm curious to know what was the first task that you handed over completely to an A I? Not just the not just, you know, having it assist you, but say, Okay, A I, you are completely in control of this process.

**[0:06:46] Chris Bradley:** Well, there's a lot of things personally, like inbox triage and responding, doing weekly updates and things like that, that were just more of my personal work. But I would say the the area that I was most interested in was order entry within our company, where we manually key 1.7 million orders a year. So I knew if we could go after something that big, it would have such a lasting and huge impact on the organization.

**[0:07:17] Chris Bradley:** So that was kind of the first real agentic task that I looked at trying to complete. And we did kind of our first test of that back in probably late November, early December of 2025. And to me, that was the the turning point. You know, everybody talks about that, you know, the Christmas holiday that changed everything. And for me, it really was because we were able to kind of get our first order through the system and and I could see the potential of what that would mean.

**[0:07:49] Chris Bradley:** You know, for the future, the types of things that we can even do today.

**[0:07:52] Host:** It's crazy and so intimidating to think, uh, 1.7 million orders all captured manually. So obviously this was a big decision. How did you know that it was ready to run without someone watching it? How did you get to the point? What what was the green flags? I guess of saying, Okay, um, I trust this enough to hand it over.

**[0:08:08] Chris Bradley:** Well, I would say we built in a in an audit trail. So every single agentic mission that we deploy, there is a screenshot at every step. And you can play it back to be able to see exactly what it it did. Because, you know, L L Ms are non-deterministic, it can do something unexpected sometimes. And so we were pretty clear from the beginning that we wanted to have that audit trail. And so for every, uh, again, action that they do, we can go back and see it.

**[0:08:40] Chris Bradley:** And so over time, we've turned, um, a lot of this onto autonomous, you know, where it's actually taking an order and picking it up and processing it fully all the way through the system. But there's always that audit trail, and I think that's what gives us comfort, um, because we can search up any case, any order, and you can go back. And quite honestly, you don't have that. We did. We didn't have that before. You couldn't see what a human did.

**[0:09:08] Chris Bradley:** Um, we don't have screen recordings of every time a human does anything on a computer, but we do with agents. And that gives us the confidence and the ability to go back and see if something did go wrong, why and how do we fix it?

**[0:09:23] Host:** That makes so much sense. So I'm thinking you said that, you know, because L L Ms are non-deterministic and they can sometimes do unexpected things. What about the system? Did it give you anything that you didn't expect, whether it was good or bad?

**[0:09:37] Chris Bradley:** Um, I mean, there were a lot of interesting learnings along the way when you think it's so clear of what it's gonna do, where, I guess, when you train a human how to do something. And I think sometimes in agentic A I, we we build, say, an an agent profile, and that agent is doing something. But every time that agent does something, it's it's in essence an a brand new L L M turn or or or call.

**[0:10:05] Chris Bradley:** And so even though we get the sense over time that we feel like it really is like this agent that we've trained to do it. If it takes twenty different agent call or like large language calls per session, you're actually getting a fresh model every time. And it's having to read the context and read the skills and read all of the guardrails and harness that you've built around it. But it's not the same as as a human.

**[0:10:32] Chris Bradley:** And so I think we can sometimes forget that and think, well, it's done it this the right way of the last 150 times. So the 151, it's gonna do it the same way again, and sometimes it just it just doesn't. So that's why we build guardrails. And in fact, some of the latest features as we've gone on, because it was operating really well at low volume. But once we got to, like, over thousands of orders, then all of a sudden you start to see these anomalies.

**[0:11:02] Chris Bradley:** So now we have to actually build another agent that would inspect at the very end of an agent run to make sure that the agent did what we expected it to do. And we've determined that, you know, large language models, they love judging other large language models. And it's so, uh, that's been an interesting kind of ride in and of itself.

**[0:11:23] Host:** Yeah, having having the one agent supervise the other agent. So I always love when we get to this part of the conversation. Um, get into also how this has changed the human's role. So what changes in role did you see within your team, uh, after this task was automated?

**[0:11:40] Chris Bradley:** You know, we were very deliberate and prescriptive, I would say, when we when we developed this program, that we did not want to say, create a new like agent tool that our customer service professionals would have to interact with. So we actually all of the interaction that our users do is within Salesforce. And so within the year, within the um, the C R M, they will sign a case and then they will have case comments back and forth with the agent who's actually doing the work on their behalf.

**[0:12:14] Chris Bradley:** And so that's really made it, I think, pretty simple and easy for us to get users to to adopt it. Um, and now it's just giving them time back to be able to do better, more interesting things. The biggest problem I would say that I've seen is a lot of the use of A I that people deploy is the fun human work. Don't have an agent go do the incredible creative work for your creative team. They're not gonna like that.

**[0:12:44] Chris Bradley:** But if you get A I to go do the grunt work, the manual labor that they always felt was beneath them, they will love it. They will adopt it. And and that's the danger sometimes. A I is like, look at what A I can write this analysis. Okay, great, but what's the analyst gonna do, you know?

**[0:13:04] Host:** Exactly, yes, I take myself as an example, and um, I think my entire team would agree with this. I absolutely despise anything administrative. So using A I to do my admin is amazing. And all of the sudden, you know, you love your job ten times more because you get to do the things that you're actually interested in. And there should be space for that. So we're heading into the part of this conversation where you get to give our audience members, the leaders, your peers, some some great advice.

**[0:13:31] Host:** And I think, or I imagine that almost every listener has already been pitched about a dozen times about a new A I starting point. And most of these vendor pitches sound exactly the same. Um, they just have a a new fancy name or a different person presenting them. And I'm thinking that the starting point is probably not even the technology question at all. And that the question should be which piece of work is worth the trouble of automating it.

**[0:14:00] Host:** And it fits into what we just said about our humans as well. So for someone in your seat at a company like yours, if they ask you where to look first, what to identify as a starting point, what would you tell them? What advice could you give them?

**[0:14:13] Chris Bradley:** What I would say is, you know, there's two primary types of leverage that you can get with A I. So you can get decision leverage, where it's really helping you do better analysis. It's and then there's also labor leverage, and that labor leverage is where you're taking repetitive manual work and you are automating that. We deliberately picked the latter, and it's it's tempting to pick the the other one first. But we were really looking for efficiency and being able to take our people and, you know, give them a tool that they would honestly like, welcome and and want to use.

**[0:14:59] Chris Bradley:** Uh, so when we started, we actually I started with our chief H R officer and I got him to give me a list of all of the jobs for like our 8,000 employees. And we looked at where do you have the highest concentrations of job types? And so we found, like customer service, financial operations, sales, where you literally could have five, 600 people that all have the same job type. And then how do we automate and find leverage opportunities for that entire population?

**[0:15:31] Chris Bradley:** Cause oftentimes you'll come up with a great use case and it's like, well, there's two people in the company that actually do that. You just created, uh, leverage for those two people. But if you could find an opportunity for a population of four or 500 people and you can take out three hours a day out of their kind of routine manual behavior. Now you have something that's really valuable, cause you can redeploy that expertise and that energy.

**[0:15:59] Chris Bradley:** So in in our case, we look at customer service, we can use that to better serve our customers. They're actually talking to customers and they're solving big problems instead of just keying orders or dealing with credits or so many things that they do that really aren't that value added from a customer experience perspective.

**[0:16:20] Host:** So that's so interesting because I think most of the conversations that we've had, our guests normally would say something. Oh no, start small, start where, uh, you know, your smallest pick one business problem, but pick your smallest one. So, and and what I'm getting from you, and especially saying that you started with an ordering system where you had 1.7 million orders being captured manually. You're saying that it doesn't really matter about the size of the project.

**[0:16:44] Host:** Uh, what matters is where it would have the biggest influence. Is that am I getting that right?

**[0:16:50] Chris Bradley:** Influence, and I think you also have to think about change management. So in all of the big transformations I've done, even prior to A I, the biggest thing that you is always gonna hold you back is gonna be change management. Do users adopt it? Are they resistant? So I think you have to find those places where where naturally people are gonna welcome it. And they will also, they won't be looking for every opportunity to point where it did something wrong.

**[0:17:19] Chris Bradley:** They're gonna be looking for how do I get this? And how does this kind of help me do more interesting work? So that's what really I think drove us. Uh, we were looking for big leverage points because it's gonna take a lot of effort to do to develop the agentic tool, whatever it is. So to do it and solve one person's problem doesn't really give you a lot of of benefit. But again, we were looking for those highest leverage, highest manual labor points where we could get incredible leverage.

**[0:17:55] Chris Bradley:** Um, you still then have to scale it, so it's not like you did it all overnight, like you would with a single person's tool. But I think it it kind of holds the keys to where you can really get value out of agentic.

**[0:18:08] Host:** That debate within your own brain makes so much sense if you think about it that way. Chris, uh, we're running out of time, but I think my my biggest takeaways from our conversation today is first that, you know, we still have so many process processes that are conducted manually. And there is more than just the cost factor of human labor. It's time, it's error, it's everything there. And if you can automate processes where you would empower your team to do more of what they like to do and less of the the silly little little things, like you said, that they feel is completely beneath them.

**[0:18:41] Host:** Um, that is a great starting point, and there is so much more in terms of leveraging than just looking at, like you mentioned, just the scaling or changing one process. There's a lot to think about and a lot of new insights that I think we haven't had on the show before. So it's been a great conversation. Thank you for sharing so much with us, and I hope to have you back on the show soon.

**[0:19:02] Chris Bradley:** Thank you. It was great to be here and, uh, excited always to talk about agentic A I.

**[0:19:22] Host:** Wrapping up today's episode, let's look at the three key takeaways from our conversation with Chris. First, a lot of the value from agentic A I shows up not in new analysis, but in fully removing people from repetitive, high volume tasks they were never adding strategic value to in the first place. Second, trusting an agent with an entire task end to end depends on building in an audit trail, and even a second agent to check the first agent's work, especially once volume scales past a few thousand cases.

**[0:19:49] Host:** Finally, the biggest leverage doesn't come from picking the smallest possible pilot. It comes from finding the process with the highest concentration of people doing the same manual job and automating that first. If you offer A I products or services into the enterprise, you need to find enterprise leaders with relevance and readiness. Emerj attracts V P and higher-ranking enterprise audiences who are already convinced they need to move beyond traditional I T. To learn the exact strategies we use to help leading A I brands and startups connect with their ideal enterprise A I buyers, visit Emerj dot com slash a d one.

**[0:20:24] Host:** That's e m e r j dot com slash a d one. On behalf of the team at Emerj, we'll see you on the next episode.