← The Times Tech Podcast
The Times Tech Podcast
Bonus episode: Agentic AI explained – The next phase of artificial intelligence
00:26:03
What Is Agentic AI and is business ready for it?
Loading summary
Transcript66 lines
- [00:00]
A
Foreign.
- [00:06]
B
I'm Katie Prescott and this bonus episode of the Times Tech Podcast is brought to you in partnership with PwC. On this podcast, we talk a lot about artificial intelligence, but increasingly the conversation is shifting. It's no longer just about AI that analyzes information or generates content, but AI systems that can actually act and action things on your behalf. This next phase is often called agentic AI. That's technology essentially, that can make decisions, coordinate tasks, and operate with a degree of autonomy. And while that opens up massive possibilities, it obviously raises some very real questions for business leaders about trust and safety. So today we're stepping back from the hype to ask what agentic AI really means in practice and what needs to be in place for organizations to use it responsibly and effectively. So I'm joined by Lilia Christoffi, who's a partner at PwC specializing in AI and data with over 20 years experience in the financial sector. We're going to talk about how agentic AI is already being applied, why trust is key to its adoption, and how leaders should be thinking about this technology as part of their long term growth strategy. Lilia, welcome to the podcast. Thanks so much for coming in today.
- [01:31]
A
Thank you for having me.
- [01:32]
B
Should we start with the basics then, for people who don't know? How do you describe agentic AI? What is it?
- [01:39]
A
There's a lot of definitions out there about agentic AI, and typically the easiest one is the wrong one, which is you have an agent who, who is invoking a generative AI engagement through a large language model. So a human asks a question, it goes to the language model and comes back with an answer. That is not agentic AI. You need multiple agents and an orchestration engine to be able to have agentic AI. That is similar in terms of analogy, when you have a manager and you have multiple team members, all specializing in different things to create an output.
- [02:12]
B
So I think of it as like a bot inside your computer that can do tasks. Are you saying that that's wrong? Is that the sort of thing we're talking about?
- [02:21]
A
So the difference here is that you have different versions of a bot, if that's what you want to call it, and they represent a different mindset or they have different specialist skills and they have to communicate with one another. That is true agentic because you're orchestrating all these little individuals or individual bots in doing their specific specialist roles within the context of an intent and an outcome.
- [02:49]
B
So what you're talking about here essentially are bots, sorry to use that word, which can act autonomously and also take the initiative to do tasks.
- [03:00]
A
So they're programmed to do specific tasks, and they have been programmed with intent and within a context, and they have to be managed and controlled. That's part of the responsible AI framework. However, of course, you can bypass that and allow them to do just about anything, and that would be unethical. They talk to one another when they have to. You have to define their interactions and how they can use each other as tools, basically. So one agent can use another agent as a tool to be able to come up with an outcome. So think of it like this. You will have a conversational bot that you are having a nice intricate conversation about. Maybe in banking, you want to buy a product and you want to know more information about that product, the eligibility of that product that you might want to know whether you have or you have the levels of creditworthiness required. Those will be activities that will be initiated through that conversation as you start going down the pipeline of that sales process.
- [03:59]
B
I see. And can you give us any other examples of how they might work in the real world?
- [04:04]
A
Yeah. So if you think about research and analysis, agentic systems are really good at that. You have different sources of data that you would consume as a human to be able to come to a deduced outcome of what you're investigating. Correct. So you will have structured data, unstructured data. Structured data may be formats of data that you get in a very formatted way versus unstructured data, which could be any articles and things like that that don't conform to a specific principle and they might have bias in them, where you have to deduce from a human point of view certain aspects of whether it's correct or it's not correct to use and process it. So as you're traversing this world and navigating it of data, you will actually have agents that you can create that will interact with these different types of data sources and then come to a conclusion. And then collectively, the manager agent will deduce what's the real truth behind what has been collected and then present that to a human.
- [05:07]
B
Wow. I mean, this sort of power raises so many questions, doesn't it? But it's an enormous leap forward as well from the current AI tools that organizations are using.
- [05:15]
A
Absolutely. And this is where I think roles are really going to start changing and shaping in terms of specialist knowledge, because humans can consume lots of information, but to really understand the information, they require certain specialist skills and they. For example, if I were to read a medical book, perhaps with some medical background I could understand it. But if a layman were to do that and has no association to the medical world at all and wants to self heal on the basis of information he's read, he may come to the wrong conclusion and deduction. So it's really important that when these things are being designed and when you're powering lots of organizations, whether it's the NHS or the government or financial services, that you're putting the right levels of business insight in collaboration with the technologists. Because a technologist will always assume that the technology is providing an accurate response and deduction out of that content. But it brings into question the human manager's role that's going to oversee the outcome of all these specialist agents. Because before we were segregated very much in our specialist functional areas. So you've got a financial crime expert and they know everything about financial crime. Then you have a financial expert who really looks at accounting and the performance from a financial perspective. Then you look at the product guy who's really good at looking at how his products are being marketed, sold, performing. So you've got all these different perspectives and all of a sudden they're coming into a single outcome. So how can that human that didn't have those variety of different skills be able to deduce if the answer was correct?
- [06:55]
B
And what's the answer to that?
- [06:57]
A
Well, the answer is that there needs to be a level of upskilling and we need to understand with the zero process based redesign, which means that you may come to an outcome or conclusion without having a workflow or a process just through the interaction with the technology. Right. You need to think about how you're going to upskill and based on proximity, these different skills within the human. PWC did a report with the City UK and what we've seen is that 75% of financial services firms and 82% of lawyers reported using AI to create efficiency and process automation rather than just value creation. So we are needing to shift our whole economy and industry in order to be able to create the value, the foresight and the governance that's required in order to have a GDP impact in our country. And if we do that, we are talking about billions worth of substantial GDP increase for our country.
- [07:58]
B
When do you see this being rolled out? Have you got a sense of a sort of time when we might all be experiencing this in the workplace and
- [08:07]
A
in our jobs, but we're rolling it out already.
- [08:11]
B
Right, who are the biggest users? Where are you seeing it being applied?
- [08:16]
A
So the private Market is going to, as we can see from private companies coming up with wild ideas and agents becoming socially aware that obviously they're moving much faster, but in a much more unregulated way. Whereas I think it will take another one to two years before you'll see a lot of this becoming generally available across pretty much every institution through maybe
- [08:39]
B
big technology companies, rather than people downloading something like OpenClaw.
- [08:43]
A
Exactly. And that's because we need to put the right guardrails and we need to put the right controls around that and we need to build in an ethical way. But have you received a call recently from a bot?
- [08:55]
B
Yes, sadly.
- [08:57]
A
Right.
- [08:57]
B
And it's. But it's not very good.
- [08:58]
A
It's not. Not good because technology is not available.
- [09:01]
B
Right.
- [09:02]
A
It's not good because there's a lot of cowboys who think they can build.
- [09:05]
B
Sure, yeah.
- [09:06]
A
This tech.
- [09:07]
B
Yeah, yeah. Well, should we talk about the trust point? Because that's clearly so important when you're painting this vision of bots running the world or certainly being integrated into businesses. How can companies control them and think about safety when they're rolling them out?
- [09:25]
A
You know, my experience over the last year when I've seen sort of agentic come to life and the experimentation take hold, I've seen that a lot of the basic principles of architecture have been forgotten.
- [09:41]
B
And you're talking about digital architecture here.
- [09:44]
A
Well, you need agentic architecture.
- [09:46]
B
Okay.
- [09:47]
A
So digital architecture should really inform agentic architecture principles, but it's too traditional because when you're going from a static environment that is very controlled. So I have a piece of code in a software and that code goes through staged gates and testing, rigorous testing, although testing principles need to evolve even in the digital world. And then that gets released and then we think that code base is locked and we know that that code base works.
- [10:12]
B
Right.
- [10:13]
A
In AI, it's not that static. Right. So your whole testing infrastructure has to change. So what you're doing is you have the code you're building in testing and self healing. That happens real time, all the time. Why? Because there are parameters that are outside of the developer control. What are those parameters? Model fluctuations and versioning infrastructure, which is done by big tech, for example, and they make changes to their code and accessibility of GPUs, for example, to run and process these large language models and all these things. So these are variations that are outside of the developer control, which means that you have to build testing that is real time. So if I'm doing a customer facing bot, if every so often I need to Be running a test on production to see how well the model is actually performing and if it's starting to introduce hallucination. And then I need to have business continuity principles in place to be able to take over if necessary, from a human perspective, switch my model or do something different because I can't allow from a certain threshold point of view for that hallucination to propagate because it will affect my customer.
- [11:22]
B
Companies have to completely rethink how they are testing, putting safeguards around their tech correct.
- [11:29]
A
And so risk departments to date review risk and controls. But tomorrow you're talking about having thousands of agents and you don't have thousands of workers. So you are multiplying the effect of management and span of control. And if a manager is, you know, very good at managing seven people, can you imagine if a manager has to manage a thousand non humans?
- [11:53]
B
Well, I wanted to ask you about that. How should a manager, a human manager, think about that?
- [11:59]
A
So a human manager has to be supported by the underlying design principles of how we should manage and control the technology. And so what the human manager should have is the execution of the features by one set of agents, but then the controls and the real time guardrails that have been standardized and built into a control tower that is supported by its own agents. So you're multiplying it's functions, you're multiplying risk functions, you're multiplying security functions, you're multiplying all these different areas that may have an impact that constitute what your control framework should look like. And you're deploying those alongside all these beautiful agents that want to do all the lovely featured kind of things that you've designed them to do. But you cannot expect that you will just have the same person being able to just take care of the one aspect. And so just one last point on this, and I think it's really important, is that we can get inundated by lots of information that we will not even be able to kind of traverse. So it's important that when you design this, you are managing by escalation. And you need to define those escalation principles for intervention.
- [13:14]
B
It's interesting because I've heard a lot of people say that the rise of agentic AI will democratise business in many ways because it allows people to have access to lots and lots of workers at a cheap cost. But actually from what you're saying, it sounds like it's going to be quite inaccessible to small businesses for quite a long time because of all of this architecture. They need to put in place.
- [13:36]
A
I have to wholeheartedly agree with that. The build cost for the current technology is significant. And until the large tech platform providers actually not just modularize, but introduce agentic services within their architectures, a lot of it is not a buy, it's a build decision by the organizations. And just like when we had a lot of digital platforms coming out, a lot of organizations took the route of not just configure, but build. We are going to be faced with the same challenge that we have with those times we are just repeating.
- [14:11]
B
On that subject. You are really on the front line of this at PwC. It'd be fascinating to hear how your clients are approaching the rise of agentic AI in the real world.
- [14:22]
A
I think there is a lot of love for changing, you know, the nature of work and there's a big openness around learning. There are organizations that I thought would be very traditional and conservative that I'm seeing now changing, you know, their risk factors around accepting or not this, this new tech. And I think the ambitions sometimes are driven in a very positive light in terms of, you know, I have one client who's really keen on making sure that, you know, we increase the value of the pensions and we have to do this change because if we don't and others do, your assets actually and your investment decisions will not be on par with other organizations. And there is a fear of losing out and not making the right decisions and therefore impacting thousands of people's lives. Whereas there are other organizations who very much come at this from a cost and a productivity and efficiency point of view, which is truly worrying because the intent is around experimentation, learning and reinvention. It's a redefining moment for all of us. For example, the banks haven't changed since the 1970s, right? What is the bank of the Future? The bank of the Future needs to grow my wealth. That is really what the bank of the Future should aim towards. It's not about providing me just money at the time I need it, it's understanding my circumstances. And that cannot happen in the context of the construct that we have today and the limitations that we have with the human workforce to be able to do that without agentic services.
- [15:57]
B
Well, on that responsible point, could you tell me about a concrete example perhaps of how AgentIC AI is being implemented responsibly?
- [16:06]
A
One of the programs of work that I'm involved in is a large pensions provider in the uk, which is a rather small entity, but servicing hundreds of thousands of pensioners. And so we need to make sure that we are putting in tech that helps grow their pension schemes and that is servicing the customer service division as well as their assets servicing and investment divisions of that organization. Now the interesting part here is that we would never get into an AI transformation with that organization unless we had the trustee board feel that there is certain safety measures in place, both for the employees, for the impacts to the market, but also for the members who are obviously having their pensions. Because any investment around the technology is actually taking out of the members kitty.
- [16:59]
B
Right. Okay. So you need to get the board to back the plan.
- [17:03]
A
Every AI transformation really needs to have buy in from the bottom. So which we call citizen led approaches to stimulate innovation. But it needs to also be run from the top with very bold ambitions and very concrete ways of working. One of the things that we agreed is to have a responsible AI framework. That framework works in a number of ways. Number one, when we come up with a use case design or capability design for AI, we will take it through a technology that does evaluation on the types of risks and controls that might be required for that AI use case. We can then take that to have a discussion at the AI Council, whether from a governance and steering committee point of view, whether that makes sense.
- [17:44]
B
And the AI Council, sorry, is something that that business creates separately.
- [17:48]
A
Correct.
- [17:48]
B
Like you would have the remuneration committee or something like that.
- [17:51]
A
Exactly. So we are making conscious decisions about the investment that's going in, the cost, the risks associated, the controls we need to put in place, just like you would once you've agreed that that becomes a blueprint of implementation and any other use case that follows that same principle and that same blueprint will be evaluated in a much faster way going forward. So you need to create lean governance and design in there. But then the second part of it is how are we actually monitoring that what we built is really conforming to those controls? And then coming back to what we talked about earlier is creating those control agents to run live to check whether the technology that we've implemented actually is doing the right things or not. And have dashboards that we can make more real time decisions on a day to day operational basis around that tech.
- [18:38]
B
Let's talk about the elephant in the room with all of this, which I think is the impact on jobs. And we've heard the likes of Marc Benioff, the boss of Salesforce, say CEOs are no longer going to lead all human workforces. Which sounds like is exactly what you're saying. What do you think the future of the workforce and organizations looks like. What does it mean for humans?
- [19:02]
A
It's an interesting one and I think it's a concept that will evolve over the next year because as I'm going through these transformations, I'm seeing patterns in relationships I didn't understand necessarily between functional areas. But one thing is very, very clear is that the risk function cannot exist in the way that they're currently executing. So they need a genetic augmentation and they need technology enhancement and modernization. That's one area. Second area is human resources. So if we're going to have non humans, how are you going to performance test those non humans against humans? How do you know what cost those non humans are running against if you know versus offshoring for example, what the case might be around that Unless you
- [19:46]
B
don't need to give them holiday.
- [19:48]
A
Well, that's true. But if you don't build the right measures in place for scale, you may end up paying more in tech than you would in a human. And people tend to build everything agentic because it's fun now and you know, engineers are happy go lucky building stuff, which is great, but it's not necessarily right for the organization. We talked about the legal teams obviously using a lot of AI and the change in face of of legal, tax and financial functional units are going to change significantly in the way that they look at data, analyze data, relationship managers, the way they do research and prepare for client meetings, client and consumer expectations and how we introduce non product based evaluators of performance within an organization. But actually looking at the client more holistically from an experience and a stickiness and a performance against the client point of view, which is completely different thought process to what we have today, especially in financial services. And so I think the shape and face of all these roles is changing. It is quite difficult to train humans to change set behaviors, set ways of thinking, even in it itself. When you're looking at data engineers who've been doing data work for ages, you are asking them to stop building domain models and you're asking them to think about dynamic builds of ontologies and the way we interact with with data. They're not used to that. Or data analysts following lineage when they no longer have to do that. Because agents are going to follow the lineage of the and track the data, where it's coming from, which system, what it's doing, is there a certain level of quality thresholds being met and all this kind of stuff. We're not going to do that work tomorrow. So what are those people going to do? They're going to have to change into those manager roles. They're going to have to change in understanding different facets that they never did before. And that takes time.
- [21:40]
B
It sounds like that adoption, if that's the right word, will take a lot of time. But as you said right at the start, this is coming down the track imminently. It's already being used. It might be being rolled out more widely in one to two years. What mindset shift do you think is needed in organizations in order to make agentic AI work as swiftly as you're predicting that it will be?
- [22:07]
A
So as part of these AI transformation programs, we actually run culture and inspiration series in our experience center. And I think this is really effective. You know, gone are the days where we used to write strategy and then it would take a number of weeks and then that's the strategy and you go on that journey for the next three years, sometimes five years. I mean, we can't do that anymore, right? Even an agentic strategy is 12 months at best. And it's, it's being reshaped literally every quarter based on value and outcomes and behavioral psychology and things like that of the workforce that is actually engaging with the tech because adoption plays a big role. So I guess people are starting to drink from the fire hose because change to your point is at a high velocity. But they're excited by it. I mean, if you look at the statistics, the new entrance into the workforce are actually really optimistic around what this tech can deliver and they're quite open to having changes in their job roles. And when things become a lot more value add kind of jobs for a lot of workers, they're actually really encouraged and enthusiastic. I mean, I have a lot of clients coming into our experience centers and really kind of innovating and thinking about what the future of this technology can do for them. And they always leave extremely happy and excited and they can't wait for it to come out. So I think with that level of enthusiasm, it's not a grudge, like a really slow process. I think you can accelerate that. The one thing that I thought was really endearing is one of my clients turned around in one of the EXCO sessions and said, it's incredible that we're running this AI transformation and we realize how fast we're running and doing things and the amount of change that's happening in a 14 week period. But actually it's having a massive impact on BAU business as usual. So they're saying, so if we can go through a transformation program at such pace. Why can't we execute our daily duties at a faster pace? So they're starting to question that and then they're starting to make the same accelerated form of decision making in the business as usual activities that they didn't feel comfortable to do before.
- [24:22]
B
Does it scare you, the rise of agentic AI?
- [24:25]
A
No.
- [24:26]
B
Are you excited by it?
- [24:27]
A
I am very excited by it because I think if we build correctly, it can create a lot of grounding in terms of fairness and transparency in society, which I think is much, much needed. And there will be a consistency in the service and a standardization, which I think is important. Unfortunately, as humans we all have our own different experiences, our own biases and things like that. And of course you can build bias into systems, but there are countermeasures to that and there's thought being put behind to make sure that that isn't the case. You can't control the human in the same way that you can control tech if you were to invest in it properly.
- [25:05]
B
Lilia, thank you so much for joining us on the Times Tech Podcast. That was absolutely fascinating. I can't believe personally, how quickly this is all happening. It feels like just yesterday that ChatGPT was made public and here we are talking about agentic AI.
- [25:20]
A
Yeah, indeed. Very exciting times for us all.
- [25:26]
B
That was Lilia Christoffi, partner for AI and data at PwC. And this episode was a sponsored bonus edition of the Times Tech Podcast, brought to you by PwC. Accelerating what's possible so you can turn vision into value, turn promise into performance, and put AI to work for your business to create real value and unlock growth. Discover more@pwc.co.uk.