Are AI agents changing the rules of business?

0:01There's like an entire class of problems that we didn't think we'd get to address, perhaps sometimes in our lifetime, that are now possible because we have these teams of agents at our disposal. Making things work in a technology environment today is going to get easier and easier. What's going to be harder is actually rethinking the way that work has been done for so long. An idea that you had on Thursday can be working product and shipped to tens of millions of users on Tuesday. AI agents are here, working alongside people, solving difficult problems, helping teams innovate and revolutionizing old ways of working.

0:42Rob Seaman:But only 20 % of businesses are capturing a majority of AI's economic value, according to PwC Research. So is your business building momentum around this next wave of AI and truly taking advantage? I'm Lizzie O 'Leary, a podcaster and journalist. And I'm Femi Oke, a broadcaster and journalist, and this is Take On Tomorrow, the podcast from PwC. Today, could AI agents transform the way your business works?

1:16Shortly, we'll be hearing why Slack, the work operating system, is embedding AI agents into everything they do.

1:23Rob Seaman:But first, to understand the opportunity of this technology, we're delighted to be joined by Joe Atkinson, PwC's Global Chief AI Officer. Joe, welcome to the show. Thanks, Lizzie. Great to be with you both. Okay, we have to start with the basics. I think we need to understand what makes an AI agent, what is it, and how is it different from the AI tools that many people have already been using. Why does this type of software have so much traction behind it right now? Everybody's talking about agents. Well, I think start with the way we use AI today. Traditional AI tools that most of us are used to include a chat window.

2:00We ask it to do something, it provides something back. The agent is a very, very different construct. Think of the agent in the background just like you would another human worker. It gets an assignment. It gets some rules to follow. It gets some data that's available to it. And it's set off to go find out the answer or complete the transaction. What kind of work can AI agents do, Joe, that would have been completely unrealistic just two or three years ago? Think of an insurance agent that's answering a call from a customer or a client, and the customer's asking them something, and they're tapping on the other side into a chatbot to get information.

2:36It's kind of the way most of us think about it. But now think of the agent being a technology agent and not an insurance agent. And instead, the customer is providing the input. but the agent is making a determination in the background as to what information is necessary to answer that customer's question. They may be referencing policy. It may be referencing data sets around claims adjustment or damage or whatever the information might be necessary to respond to that customer. And then providing the customer not just with information, but with a completed transaction. The way that we experience AI today and the way that AI will continue to function in the future, very, very different.

3:11Thank you, Joe. We'll come back to you shortly. But first, to get a first-hand account from a company using these tools in their everyday work, I spoke to Rob Seaman, the general manager of Slack at Salesforce, who are incorporating AI agents into their platform at scale. I started by asking him about how businesses can get the most out of agentic AI.

3:36One, I think they have to have an experimentation mindset. I mean, one of the things that's most important is you're not going to get it right from day one. And very often, you know, we are wrong the first four, five, six, seven attempts. And it's about figuring out how wrong you are as quick as you can in adjusting. And so in that regard, I would say people should have a very liberal approach to deploying agents in different areas and seeing what works. So like sitting around and waiting for the perfect use case is the wrong approach. Have you been able to use Agentic AI to transform the capacity at Slack?

4:14When we went into the year, we had a goal and a set of things we wanted to achieve in the year. We delivered our entire roadmap in the first quarter for the year. And so it was like, okay, we've got to kind of rethink how we do things now. So the interesting thing is code is no longer the bottleneck. Ideas and craft and judgment are the bottleneck. And we've seen that play out throughout all of Slack. And so we've done a number of things. We wanted to increase the overall AI proficiency of the team at Slack. And so we took our entire team, thousands of people, offline for three days to do what we call builder days.

4:49All jamming and vibing in Slack and AI to figure out how they can actually improve how they do their work. I mean, what's interesting is we actually have 50 % more users of the agendic coding tools within Slack than we have engineers. And that shows you like our product managers, our designers, our EAs, our HR team, everybody is contributing to the success of our customers and our product itself. And so we've seen just like a tremendous productivity improvement. So we've been able to deliver more. We've been able to iterate faster. And we've had to kind of restructure how we work as a result of it.

5:23Let me go big picture, which is why agentic AI will change the way we work. What are the opportunities? The opportunity for agentic AI is ultimately to increase human output and allow us to get back to the things that we are uniquely positioned to do, which is exercise our judgment, our experience, our craft, be creative. You know, these are things that the agents are not great at right now, and we don't have a path to them being as good as humans. And so I don't see it being about replacing people at all. Like I see the bull case for this being augmenting people and ultimately increasing the upper bound of what's possible for companies and individuals that work within companies.

6:11And so what's interesting, and we see within some of the most advanced customers of Slack that are also pushing the envelope from an AI adoption perspective, they have as many agents in their Slack as they have humans. They are teammates, like they coexist in your Slack. And as a result of it, every single person in every single company can achieve more than they thought was possible. An idea that you had on Thursday can be working product and shipped to tens of millions of users on Tuesday. And the cool thing about that is we don't know what's next because of it. You know, there's like an entire class of problems that we didn't think we'd get to address certainly this year, maybe not this month, or perhaps sometimes in our lifetime that are now possible because we have these teams of agents that are at our disposal.

6:58So for AI agents to really do their best work, what needs to change in the way that we typically are doing business? What are you noticing for yourself in Slack about how you're maybe having to work differently? One, these agents need context And Slack is a perfect system for context because it's where people have conversations. It's where all your topics, teams, work items, the priorities of your company are reflected and the conversations are happening. So that's valuable context for the agents. So if you look at some of the most proficient customers of Slack, their employees think out loud in Slack.

7:36And all of that context is actually available then to agents. because if you think about it, it immediately disseminates all knowledge within a company to all of your employees and to all of your agents. It reduces the duplication of effort and it increases the probability that ideas are going to compound and build upon each other. Give us an example of an organization, you cannot name your own company, where AI agents are being used really well, where you're just looking at them with admiration for what they're doing. So I'll probably go with Anthropic. I'd say 65-70 % of the code written at Anthropic right now is actually happening through channels in Slack, which I think is very, very different than what you see from the rest of the world.

8:20So as we're going back and forth and having a conversation and being like, okay, we have an idea, or you know what, I saw this thing that's broken over there. They just tag it in Claude and then boom, it writes the code and it's done. And they just keep carrying on the conversation. I think that's so far ahead of where many companies are at this point. And this comes back a little bit to the term vibe. I feel like vibe coding has gotten a bad name, but vibing, you get people together and they just kind of jam and they riff off of each other and the outcome is amazing. And to me, that's what vibe coding is.

8:52It's like human beings working together in public in a multiplayer fashion, bouncing ideas off of each other, playing to each other's strengths and it all happening because AI is right there in the conversation with them. I'm going to throw a piece of research at you from PwC. Research finds that 74 % of the benefits of AI are going to 20 % of companies. Yep. What are the organizations getting the greatest value from AI agents doing? Honestly, this is what we are aiming to help solve. If you look at the 20 % of companies that are getting the value, one, I think they were early adopters and they didn't try to make a choice about which AI was right necessarily or perfect for everybody.

9:40They tried a bunch of different things. Two, they used these things as teams. I think one of the downfalls of AI tools is that many of them are single player, you know, like you leave it to an individual like in a silo to go use the AI and get value out of it. And they can't see what each other are doing and that value doesn't accrue to the broader companies. I feel optimism from you. It's like exuding all the way through our conversation. I'm going to add a but into our conversation because I know you'll be very candid about the but. There are stories, examples of AI agents being destructive.

10:20How do you manage the risk as you're playing, as you're experimenting, as you're being creative? How do you manage the risk? So one, I think building in mechanisms for trust and verification. I think it's very important that the agents operate in systems with the access rights of the users that are instructing them. We don't give agents unilateral or universal access to systems. Like they operate on behalf of the human that's instructing them and are restricted to what that particular human can do. So I think that helps. The human needs to know to trust the system, but also verify it. What do the leaders need to understand?

10:57What do they need to do? They need to use it themselves. And I think that is paramount. Like you have to lead by example. Like you're never going to be the best. But I use it and play with it every single day. And it's almost like a muscle you have to exercise. And I think that would be the biggest thing I'd say to leaders. Don't be afraid of this stuff. Solve a very real problem. You're so optimistic about AI. What is possible in the future beyond business? Well, I mean, this comes back to something I'm not satisfied with, with AI. So if you look at just even the nature of this conversation and the questions we're asking and the answers that we're having, we're so much focused on business, we're focused on revenue, we're focused on expense reduction.

11:37But there's some very real world problems that are out there. We're potentially exacerbating with AI as opposed to solving with AI. And I kind of just as a human being wish we would pivot our focus a little bit more towards those problems. Like you look at the climate change that's happening. And it's like, okay, we can increase revenue and we can reduce expense. But if we took all the world's creativity and power and all this AI and pointed it at the problem of climate change, we could probably fix it, right? AI could fix climate change. AI can help with poverty. AI can help with hunger. But we all got to decide to apply it to those things or it's not going to do it.

12:17Rob Seaman, it has been a pleasure chatting with you. Thanks for joining us on Take On Tomorrow. Thank you very much for having me. It's a pleasure to be here.

12:32Rob Seaman:so joe rob described how ai agents have inspired slack to rethink the way it works i wonder if you are seeing the same thing with clients you work with we absolutely are lizzie and i loved what rob said because he talked about this idea that coding isn't the bottleneck anymore it's ideas and judgment, and I think he's spot on. The reality is that making things work in a technology environment today is going to get easier and easier. What's going to be harder is actually rethinking the way that work has been done for so long. There's that classic innovation story about when we moved from steam to electricity in the factories and everybody put the machinery where the steam pipes were because they hadn't really rethought about the fact that you didn't have to have the factory floor flow around the steam pipes anymore.

13:20We're living that exact same situation right now in the business flow, in the workflows, in organizations. We have to rethink the way work gets done. And that is much more an inputs and outcomes discussion than it is a process and steps discussion. That is easy to say. It's easy to describe. It is actually devilishly difficult to do. So what is the difference between automating something that a business is already doing and then redesigning a business around AI? I love that question because for so long we've all been focused on automation, but think of automation as doing the things we do today faster with fewer keystrokes.

13:55So whatever the processes we have today, we can automate that process and execute steps one through ten in a lot more speed with a lot less keystrokes than we used to do. And that's automation, and that has real value, and companies should continue to drive automation wherever they can. But redesign challenges the ten steps in the first place and says, what is it actually that we're trying to accomplish? And for me, I always come back to the roots of many of us that have been in professional services and consulting for a long time. When we used to do the flow charts and figure out where the data sets were and where the process points were and the decision points, the reality is those steps now are getting compressed by agentic workflow.

14:32Humans will still be there, to be really clear. We're not talking about taking humans out of this. But what we are talking about, letting agents do the work that agents are great at, letting humans do the work that humans are great at, which is judgment and clarity and visioning and strategy. And if you bring those together, it's a very powerful opportunity to go from automation to redesign.

14:49Rob Seaman:I mean, it makes so much sense conceptually, but I wonder if you could give us an example of a workflow, maybe a complex workflow, where you are seeing agents change how the work gets done. Well, let me go. I'm going to go back to insurance because I think insurance, interestingly, insurance and financial services are actually leading in a lot of ways here. So think about a multi-step case handling for a life insurer. A life insurer has got to figure out whether or not they know who you are. Do you have coverage? Is that coverage in place? They've got to then, once they get to the identity validation, they've got to figure out, well, has Lizzie asked us for any help on this particular claim before?

15:24Is there history here on this claim? Have we reviewed this claim request before? That's all things today that in I'll call the legacy automation world, you may have some automation that speeds that up, but generally speaking, that's somebody that's stepping in. Well, now you could have an agent on identity that says, let me go validate and confirm that Lizzie is a covered individual in our organization and that she's eligible for the services that we're talking about. And you could have an agent that's really built on the medical data that's necessary to evaluate the nature of the claim. Is this in our covered services?

15:56Is this something that's experimental? Is this something that is not yet in the book of services that we cover that then therefore needs additional evaluation? And all the while, Lizzie needs communication. Liz needs to know where the process is and whether her claim is being evaluated and what's happening. So start to think about that as multi-agent orchestration as opposed to what most of us are still thinking about, which is an agent using a chatbot to serve Lizzie better. Both things can be done and both things can improve. One is going to give you that outsized performance that we talk about with the AI performance study.

16:27If businesses are using AI agents to make individual employees more productive, how do they then scale that up to their wider business so the business becomes more productive and we don't just end up giving individual employees more work to do? So one of the things I've said many times is that today, with general productivity tools, the chatbots and copilots, et cetera, all important, improving the way people work. But what organizations have experienced to date is that the benefits of those tools are accruing more to the individual than they are to the enterprise. And this goes exactly to your point, right?

17:06Why is it occurring to the individual? It's because most of us, if I could figure out how to make my 10-hour or my 12-hour day eight hours, I'm probably not going to raise my hand and say, in those other four hours, what do you need me to do? I'm going to go have dinner with my wife or spend some time with my kids. That, by the way, is a good thing. I think that means work will get better over time. But the challenge for management is how do we open up the right capacity and redirect it to the things that are actually going to move the proverbial needle from a performance and growth perspective?

17:32That's a management problem. And that management problem really becomes the way organizations make choices. How do I start to frame out what I expect you to be doing with the productivity gains that you have?

17:43Rob Seaman:But then I guess the idea is to tackle bigger problems that couldn't be tackled before. And how do leaders do that? Because that requires a real mental reframe. There are problems on the plate that generally speaking, if you look back over five, 10 years, that most of us would say, that's a really big problem, but I have no idea how to tackle it. So I'm going to work pragmatically on the stuff that's on the plate that I know I can move. Today, this is moving in a very different direction. problems that before would have said to us that we don't have the resources to do it, at some point in the future we can, need to come forward on the agenda.

18:18And they may come forward on the agenda outside of the typical planning cycle. Most organizations plan on one-year budget cycles. Most of us have looked at three-year or five-year strategies and three-year or five-year plans. What this means is that you're much more into milestone planning. You're into multiple looks at the strategic plan over the course of a year versus, let's see what we think about this in October as we plan our next calendar year. I think those days are gone. And again, that's going to create much more pressure on the C-suite to have a lot more agility in the plan. I want to talk about responsibility and risk.

18:52When an AI agent makes a decision or takes action, and it's a bad decision and it's a risky action, who's responsible for that? Yeah, Femi, it's such an important point. Unfortunately, we have public examples already that are well known about agents taking actions that were perhaps, at best, unanticipated, at worst, shouldn't have happened. And unfortunately, I think we're going to see more of that. One of the things I say to my teams all the time is, you were responsible for the output of your work yesterday. Today you may be using a very different set of tools and a set of agents to perform your work.

19:30You are still responsible for the output of your work. Ultimately, humans have to be accountable for the execution of work, no matter what degree of agent orchestration is being used to do it. There's no accountability on the shoulder of an agent. That accountability rests with the organization, the team, the individual that's architected and executing that workflow with that agent. That's a very different management skill set. It's going to require different thinking from all of us. And I think it's one of the most important things in terms of the horizon on responsible use of AI that we all have to get used to pretty quickly, which is what's actually happening in the background of this process and do I understand it.

20:07Rob Seaman:If someone is listening to this right now and they run a business and they're thinking, oh, I don't know, agentic AI is fascinating, but I don't quite know how I should be thinking about this. What are your top two or three takeaways for them? So a couple. One is, and I'm going to echo something that Rob said about leaders and organizations. I think this is true for all of us. and if you're owning a business, I think it's incredibly important, is play with tools. So even before you worry about the advances of agentic AI, the reality is the capability of the tools today can do an awful lot for your business today.

20:38But it's very difficult to see the art of the possible until you start to play and you start to see the capability. Many, many people, for better or for worse, are evaluating the capability of the technology based on what they read in the press and online, etc. And that's fine and it's valuable. but the experience of working with the tools can really help to open your eyes about what the art of the possible is. The second thing that I think is really important is don't wait for perfect. So if you're building a big organization, you've got a lot of work underway, it's easy to look across the horizon and say, okay, well, I'm not quite sure it's ready to do that yet.

21:11If you wait, you're likely going to wake up at the moment that it's too late. And so this idea of iteration and four or five, six, ten failures before you get it perfect, I think also means start right now. So don't wait for the perfect. And then the last two, maybe three points, I'd say the redesign over automate. I've been talking a lot with clients and organizations that actually we've maybe at risk of framing this as a false choice. Don't automate, just redesign. You actually have to do both. There's a lot of efficiency to be gained in the organization to continue advancing the base capabilities that most organizations have.

21:44But at the same time, you have to be looking at those places where redesign could open up not only new efficiency, but new growth. And then the last, accountability and governance. You have to be driving that throughout the organization at every single level. If people don't understand their accountability, then the risk of agents doing things that are unexpected or AI doing things that don't meet the quality or the values goes up. So accountability and governance becomes paramount. Joe Atkinson, thank you so much for joining us. My pleasure. Thanks for having me.

22:19Rob Seaman:Femi, what I loved about these two conversations is that they are both simultaneously about kind of the day-to-day of running a business, but also big, thinky questions about who is responsible for what and how do you have new kinds of thoughts if you've cleared some work off of your plate. And that, I don't know, there's a creativity there that I think is really fascinating. And also the optimism, the possibilities, what this can do for work and for our lives, the idea that an agentic AI tool can half your work. That's extraordinary. But also the be careful, be creative, be open minded, and the possibilities are endless.