- 25 minutes ago
AI agents are rapidly evolving from copilots into autonomous digital teammates, reshaping how we work, collaborate, and communicate. Inside organizations, they support research, drafting, and decision-making. But their most visible impact is at the front lines: customer service, sales, and client engagement.
From banking assistants to AI-powered support agents handling millions of interactions, companies are redefining how they connect with customers. But do users actually want to interact with AI? How much autonomy should these systems have, and how do we maintain trust, empathy, and accountability when machines become the first point of contact? This session explores what it takes to make this new interface work for both businesses and consumers.
From banking assistants to AI-powered support agents handling millions of interactions, companies are redefining how they connect with customers. But do users actually want to interact with AI? How much autonomy should these systems have, and how do we maintain trust, empathy, and accountability when machines become the first point of contact? This session explores what it takes to make this new interface work for both businesses and consumers.
Category
🤖
TechTranscript
00:23Sit over there.
00:30Hello. Good morning, everyone, and welcome to Sounds Like AI, How Smart Agents Are Reshaping the Way We Work and
00:39Communicate.
00:40I'm Kelsey Chang from Caixin Global, and we are a finance and business news organization based in China.
00:47Now, for the past two years, much of the conversation around AI agents has been about co-pilots.
00:53But today, they are moving from being our assistants to something that's increasingly making decisions on our behalf.
01:04So today's discussion is really about three questions.
01:09What makes an AI system a true agent?
01:12Where can these agents create real value?
01:15And how do we build trust and human judgment into these autonomous systems?
01:19Now, to help us unpack this, we have a panel of fantastic speakers.
01:24First, we have Mohamed Ali, Senior VP and Head of IBM Consulting, which helps large enterprises strategize and deploy AI
01:33tech transformation plans.
01:36Then, we have Eleanor Crespo, Co-Founder and CEO of Pigment.
01:41Pigment is an AI-powered business planning platform used by companies to connect teams, data, and decision-making.
01:50And we have Yang Su, Head of AI for Transaction Banking and Head of AI and IT Innovation at PNB
01:57Paribas.
01:58Su brings the view from a highly regulated sector where the potential is huge, but so are the risks.
02:05And last but not least, we have Bar Winkler, CEO of Wonderful AI.
02:10Wonderful focuses on customer-facing AI agents, helping businesses communicate with their customers in more than 30 markets.
02:17Welcome, and thank you so much all for being here with us.
02:20Now, let's begin with the broader question and setting down some definitions.
02:25This question is for everyone.
02:26So, from your perspective, what makes an AI system a true agent?
02:32And are companies actually ready to deploy these agents these days?
02:36Mohamed, could you kick us off, please?
02:39Sure. First of all, thank you for having me here.
02:44So, there are a lot of terms in AI these days.
02:47But when we think of agents, we think of a piece of software that calls an LLM, so it's non
02:55-deterministic.
02:57And it can take actions, right?
03:00And in order for it to take actions, you need to create guardrails for it.
03:03So, the question is, are companies ready to deploy these things?
03:07And the answer is absolutely yes.
03:09And I think my colleagues are all examples of that.
03:12We have over 500 engagements with large corporations doing exactly this, representing about $10 billion of business.
03:21And it's not an experiment.
03:23For example, we have a set of 28 agents that do threat investigations.
03:30And they did 70,000 threat investigations last month for our clients.
03:38And Eleanor?
03:39Yes, I think you gave the definition of agents.
03:42I think probably most of you know by now what an agent is.
03:45An agent is super autonomous.
03:47He can get decisions on your behalf.
03:49And he can really act as a digital teammate for you.
03:52So, what we see at Pigment, to make it concrete, is that you can now give tasks that are, you
03:57know, high-level goals of this is what I want to achieve.
04:00I want to create this business model within the platform.
04:02And you can close your laptop, and you come back tomorrow morning, and the model is built for you end
04:07-to-end.
04:07And the agent has chosen itself the different steps to get there.
04:12So, the capabilities, you know, in the past six months for my world, which is the world of finance, have
04:18been absolutely incredible.
04:19And what we are able to do today is completely unparalleled to what we were able to do six months
04:23ago with agents.
04:24And I think we are just scratching the surface here.
04:28And Sue?
04:29Yeah.
04:29So, we were in the prep room, and there was an advertisement where people just talk with AI all the
04:36time.
04:36So, it's the same with the agents as well.
04:39We're definitely hearing about it all the time.
04:42So, which is why I really wanted to start with maybe something else that we're more familiar with.
04:47I want to talk about self-driving cars.
04:50Back in 2006, the research lab at Ecole des Minions in France already knew how to build and operate the
04:58self-autonomous vehicles.
05:00On the countryside mostly, because the technology at that time was really working on looking at the markings on the
05:07road.
05:07And things get more complicated when we have to go to run in the city because of the pedestrians, because
05:15of the other vehicles,
05:16and because, well, things are just a little bit more messy in the city in general, right?
05:22And that is exactly the same for the agents.
05:26It's very easy to build and deploy agents in a secure and controlled environment.
05:31The challenge arises when you have to meet the messy world, the real world, which is precisely why that despite
05:40having agents in productions with their limited level of decisions,
05:44we are really focusing on working out the condition to scale the agency capabilities in a secure and cost-effective
05:54way.
05:54So, on the one hand, we have those agents in production.
05:56On the other, we have something called My Companion, the AI assistant that delivers the LLM's capabilities to all the
06:03employees that we roll out before the end of the year.
06:06And when you put both together, we know that we have the ability to definitely deliver the agency capabilities into
06:14the hands of our employees.
06:15So, the bottom line is not really deploying autonomous agents into the open market, but is to make sure that
06:24they actually deliver value in real world for our customers, for our employees,
06:30while inside the framework that we can fully control.
06:33Thank you, Sue.
06:34And, Bar, for you, what do you think is a true AI agent?
06:42Thanks for having me.
06:44Yeah, obviously, all the technical explanations are on point.
06:48I think the more puristic way to look at it from my perspective is that an agent is an intelligence
06:54that can conduct a synchronous or an unsynchronous process
06:59using LLMs and other technologies that are made available to it.
07:03I think what we're seeing in terms of the second half of your question, which is, are enterprises ready to
07:09deploy those or not?
07:13It depends on how wide you want to give, like, customer support, I think, is a great example.
07:19So, we started in customer support, and first thing, we build agents that answer simple questions.
07:24They needed to go to a database and answer questions.
07:27Then, we wanted to start answering questions regarding a bill.
07:30So, suddenly, you need to allow the agent to access a bill.
07:32Now, you want to start allowing the agent to restart a router.
07:35So, the agent has just become more capable, similarly to a human that was going through training in a contact
07:42center.
07:43That was last year.
07:45I think what we're seeing now in 2026 is that people understand that the modalities that are now supported by
07:51AI are not limited to voice and text.
07:53They can now operate computers, as an example.
07:56So, why not continue the same agent process and allow it to access back-end systems and conduct actions that
08:02are complex?
08:03Now, you can start solving asynchronous processes, like an insurance claim, right?
08:09Or a mortgage, which can include talking with a customer, talking with an employee, uploading a document, editing something in
08:16a system, et cetera.
08:18So, wonderful is really focused on those, like, complex cross-departmental processes.
08:22So, there is an agent that has the entire process end-to-end, but in reality, you can also call
08:27it many agents that talk to each other and hand over tasks.
08:30We think that's the highest way to unlock value, and it's possible today, technologically.
08:36Enterprises are trying to do it.
08:38It will take a bit of time for some enterprises, but I think the main thing we're seeing is that
08:42when there is bullishness from the management of an enterprise,
08:46and the perceived risk of not moving forward is bigger than the perceived risk of moving forward and making a
08:51mistake here and there,
08:53there's no real limitation to what can be achieved.
08:55And I think in the next few months, we're going to start seeing full end-to-end workflows done with
08:59agents that were completely imaginary last year,
09:04because the technology is there. It's mostly an implementation game.
09:07Thank you, Bar. So, thank you for the first round. That gives us a really good foundation.
09:13Now, let's move to deployment and actually where these agents are actually creating value today.
09:19Wanted to stick with you, Bar.
09:22Because for many people, their first interaction with AI agents might be as a customer, right?
09:28When they call their telecom provider, when they contact the retailer.
09:32Based on your observations across the various markets that you operate in,
09:37do actually consumers want to talk to AI agents, and under what conditions are they actually the preferred choice over
09:44a human?
09:47I think, I mean, no one really wants to talk with customer support, like, regardless if it's a human or
09:53agent.
09:54Like, people in general are not happy to have this conversation, and it's highly transactional.
09:58I think the only reason why a human wouldn't want to talk to an AI entity to solve a problem
10:03is because they lack the belief that the entity will be able to solve the problem,
10:06or the entity doesn't sound like a human, doesn't give you a feel that it can help.
10:12We see that the containment rates across agents that are exactly the same is increasing over time,
10:18which tells you a lot about the human psychology aspect of it.
10:21So in that aspect, I think people in a transactional situation are happy to speak with whoever they think can
10:27solve their problem.
10:28And if they speak with an AI agent, and they don't give it a chance to solve their problem,
10:32and then they have to wait six or nine or 20 minutes for a rep, maybe the next time they'll
10:36try to do it.
10:37And we believe that there's, like, this is a one-way street.
10:39Like, once a human gets instant help from an agent across, like, a voice or chat session, they will not
10:44want to talk to a human.
10:45And I think, again, in 2027, maybe the second half of it, it will start becoming weird to talk to
10:50a human in context-centered environments.
10:53I think the more interesting aspect that we're seeing is with regards to agents that are conducting sales.
11:00So what we learn...
11:01So at first, I was pretty much anti-doing sales with agents because, as a consumer, I didn't think it
11:06would work on me.
11:07So I'm like, why invest our time there? The technology is not there, psychology is not there, etc.
11:11But what we're discovering now is that when there is intent, it works as good or better than a human.
11:17Now, there are some pseudo-sales situations, like debt collection.
11:21So in Latin America, we have offices in five different markets in Latin America.
11:25And collections is a huge use case there. People sometimes don't pay their bill on time.
11:30And collections is very much a sales process. Like, you're negotiating with a person to try and get them to
11:34pay you a bill.
11:35You offer them something in return. Now, we see that agents are better than humans at doing this.
11:41Now, you can tell yourself that's because they call more times, they're more aggressive, etc.
11:45But for some reason, they're better at it, and that's what the data shows.
11:49So I think that's pretty amazing. Also, on pure sales, we now have an agent with one of the bank
11:55customers, actually,
11:56that is identifying when a user is getting close to their limit in their credit card, and they call them
12:01to offer them a loan.
12:02Their conversion rate is higher than a human.
12:04Now, I wouldn't say it's because they're more human than a human or because they have more empathy.
12:09They just have perfect timing. So when there is an intent and you get a call, like, so intent triumphs
12:15everything.
12:15I think this is only going to improve over time. And yeah, I hope this answers your question.
12:24Thank you. So usefulness in the AI agent themselves and tracking the data and actually finding the right point to
12:33intercept.
12:34So Eleanor, thank you for the external side. For the internal side, you have a very interesting product.
12:40So when inside companies, when AI agents enter the business planning process, are they just simply making teams more efficient,
12:48or are they actually making the decision-making as well?
12:51Yeah, so I think I see two steps in the way I think about ROI in our context, in the
12:57context of AI business planning, analytics, etc.
13:00The first step definitely is productivity. So it's really about making all of you more productive, more effective, faster.
13:08So we see it already today. You can model 80% faster than before. The agent can do it for
13:14you. Analytics is the same.
13:16So that's obviously a very, very strong ROI and a very strong value prop.
13:21And we're going to keep seeing more ROI there because now we can create, as Bar mentioned, some end-to
13:26-end workflows
13:26where you can really replace a task or a daily day of the analysis completely end-to-end thanks to
13:33workflows.
13:34So that's great. Now, I think the second step that we are working towards is decision quality.
13:40And it's really to help you way beyond productivity about making any CEO or any CFO smarter about what they
13:48want to do.
13:49So, you know, when it comes to thinking about how to re-forecast your business, thinking about where to add
13:54headcount,
13:54thinking about how to optimize your margin, agents are now going to be able to do that for you.
13:59So they're going to be able to, you know, run for us, for instance, thousands of scenarios in parallel
14:03and help you really optimize where your company should go.
14:06And so this is where I'm extremely bullish about, you know, the next year and where I think every company
14:13is going to try to go.
14:14Thank you, Eleanor. Mohamed, let's bring it to the enterprise level.
14:18So when a large company comes to you and say, we want to deploy AI agents,
14:22what is the framework you go through to make sure you're asking the right questions or you can share some
14:30examples on how you do that?
14:31Yeah. So a lot of companies want to do this.
14:34And a lot of companies are actually not ready to do this.
14:37And so there are actually projects we will walk away from because the company doesn't have the right mentality to
14:46make it successful.
14:47So what do I mean by this?
14:49Even before you start talking about AI and agents, you have to be willing to decompose the processes in your
14:56organization.
14:57Right. So process workflow reengineering is actually super important.
15:03And if all you want to do is buy a bunch of licenses and hand it out to employees and
15:08say, go be productive, that just does not work.
15:10And we don't want to do that project. So we actually did this to ourselves three years ago.
15:15We decomposed the IBM company into 490 workflows.
15:20Then we took 70 that represented a $25 billion spend.
15:24And we reengineered it for what we call human plus digital labor, which is insertion of agents.
15:31And that $25 billion, three years later, is now approximately $20 billion.
15:39So about $4.5 billion of savings.
15:41You could see that in our financials.
15:43Free cash flow went from $9 billion to $13 billion.
15:46But that only happened because we were willing to reengineer it.
15:49The other thing that we did that we focused on heavily are guardrails.
15:54Because in an enterprise, you need to do this in a secure and governed way.
16:00So we actually built an Uber harness for all the AI that was underneath.
16:05So whether you're using, you know, Gemini or Anthropic or OpenAI or IBM's Watson X, there is a harness that
16:14manages all of that.
16:16And that is actually what we've now deployed at scale.
16:20So you take this governed way of approaching this along with the process reengineering.
16:27And then you start getting results.
16:29You know, Ben talked about some of those results.
16:31Like, we are actually seeing that, you know, at enterprise scale.
16:35And like I said, we did it to ourselves.
16:37We're now working with about 500 clients.
16:39L'Oreal is a great client.
16:41We started working with them two years ago.
16:43And the process that they wanted to reengineer was their R&D process.
16:48You know, raise your hand if you have ever bought a L'Oreal product.
16:52Look at that.
16:53Like half the room, right?
16:55There are many products out there.
16:57That R&D process is a sometimes long process.
17:02But unless you are willing to actually decompose the process, you can't effectively insert AI.
17:09Okay.
17:11So guardrails and the shift in mindset for readiness.
17:15So let's bring in the finance perspective here.
17:19One of the sectors where the opportunities are very, very big, but the margin of error is very small.
17:25So in transaction banking, what is the right level of autonomy for AI agents?
17:31And what should still remain as a human judgment?
17:36So let me remind everyone what transaction banking is.
17:40So it's the core plumbing or the bedrock of financial services for a bank.
17:44It's where we help the customers answer the question, how my money flows, how my money is aggregated, and how
17:51we put it to use to work for your business.
17:54So it's e-banking, it's payment, it's trade finance, cash management, a lot of those things where B&B Parabers
18:00is the leader in Europe.
18:03For those who know me well, I do like some quotes.
18:07So I want to make a quote from someone called Seneca.
18:12Erare humanum est, which roughly translates into the arrow is human.
18:19I want to add a personal twist to it.
18:22The arrow is machine too.
18:24So AI makes mistakes.
18:26AI will make mistakes.
18:28And that is very reason we have to put the right governance around the process to make sure that those
18:35mistakes do not translate into catastrophic failures.
18:41In other words, we're managing risks.
18:43And managing risks just happens to be B&B Parabers business over the last 180 years.
18:48We put layers of defenses to make sure we supervise the actions of the people.
18:53And we're going to obviously put defenses of layers, layers of defenses to supervise the actions of the agents who,
19:01if we take the service and we're answering the request of a customer,
19:06or let's cooperate, the most time consuming part is definitely going to gather the information, making the sense out of
19:13it, making where the payment is for the customers, not the very answering part.
19:18So I think in the near future, there will always be going to be someone to take that action to
19:25validate the final answers to the customer.
19:29So going back really directly to the question, the key distinction when we talk about the autonomy, since agent is
19:37part of a system, part of a system where there is a human in the loop,
19:41well, we're actually talking about the autonomy of an agent on the given systems, not in general of a whole
19:48system.
19:49And I believe the right answer to the question isn't a specific level of autonomy, but the right autonomy on
19:56the right process inside a framework that we fully control.
20:01Thank you, Sue.
20:03Now, we've heard very good things about where agents can create value, but the more capable these systems become, the
20:11more important the trust issue is, right?
20:13So, Eleanor, let's start with you for this round. From your experience in enterprise planning, how should companies redesign their
20:20operating models or their workflows so that AI agents become a trusted part of the team?
20:27Yeah, so there is a lot of work to do to get there, a lot of work for every company.
20:32So if I look at all of you, you know you are all on your individual phone right now, and
20:37this is what's happening today with AI.
20:38It's that everybody is experimenting on their side, they are doing their pilots, it's great, CEOs are pushing, so everybody
20:44is trying to do as much AI as possible.
20:46But clearly, what we need from there, especially in the world of finance, supply chain and decision making, is one
20:53single source of truth.
20:55One single infrastructure where you are all going to work together with the same context, with the right governance, the
21:01right access rights, to be able to really talk the same language.
21:05That's absolutely critical in our world. So that's the first thing you need, so proper infrastructure.
21:11The second is a technology that can scale, because today obviously, the goal of agents, given the power of agents,
21:17is to be as powerful as possible within your org,
21:19and being able to handle amounts of data that were almost difficult for a human to really make a sense
21:24of.
21:25So again, you need a technology that can scale.
21:27Third is about transparency and accountability, auditability, because guess what?
21:33If one of you does something, I want to be able to make sure, if an agent starts to do
21:38a task on behalf of one of you,
21:39I want to be able to make sure I understand perfectly what happened here, step by step, and even more
21:45potentially than with a human.
21:46So this is also something very big that we provide with Pigment is a technology to be able to bring
21:51that auditability.
21:53So it's great because now, you know, there are platforms like us that can provide these things.
21:57But the last one that is still an organization is change management, and it's the fact that you completely have
22:03to change the mindset in which you work when you start working with agents.
22:07Because now it's not anymore about, you know, micromanaging task by task, it's really about thinking about the strategic goal
22:13that you want to achieve,
22:15give that to the agent, and make sure you're here as a manager to validate, stay in the loop, but
22:20in the right way to make this possible.
22:22And this requires, obviously, a lot of training and a lot of changes in processes for any company out there.
22:28But look, I am very impressed, you know, with how our customers are starting to adopt it.
22:33We see companies, for instance, like Uber, Palo Alto Networks within Pigment Portfolio that have managed to deploy that at
22:39scale extremely fast.
22:40So I think, you know, if these big American players can do it, I think everybody can do it as
22:44well.
22:44Thank you, Eleanor. That's a great framework.
22:47Bar, turning to you for customer service agents, also accountability issue here.
22:52So when an agent makes a wrong recommendation, deals, mishandles a sensitive customer, who should be accountable and what framework
23:01should there be?
23:06I think there's two ways to, as an enterprise, you can deploy those agents either by buying an agent like
23:14in a black box,
23:15in which case the accountability would obviously be on the vendor.
23:20That's not the way that we usually like to operate in Wonderful.
23:22So we have like a co-build mentality. We believe that's the right path for enterprises as well, to build
23:28their internal AI muscles and not be dependent on any specific vendor for any specific use case.
23:35So to answer your question, obviously, you need to lower the chances of a mistake happening by testing and doing
23:43as much like evaluations, et cetera, to make sure that you lower the chance of a mistake on error happening.
23:48But the real way to be excellent in that aspect is to be on the ground with your customers and
23:55be fully accountable like you are a part of their team.
23:59So the way we operate in Wonderful is we are extremely local. So we have 30 offices around the world.
24:04We don't have the concept of a sales office. So we either have a full service team on the ground
24:09or we don't operate in this market.
24:10And this means that our teams are consisting of a GM who's like a CEO, a CTO, FDs, deployment strategists,
24:20solution architects, even talent acquisition and marketing.
24:22These are like full service teams. And in their world, their customers in their market are the most important thing
24:28in the world.
24:29So our customers don't need to wait for someone in San Francisco to wake up to solve some production challenge.
24:34Like they have people that are fully embedded in their team to solve those problems.
24:38And this creates accountability. I hope this answers your question.
24:43Thank you. So basically the importance of localization will be key in the markets that you operate then.
24:50Yeah. If you want to be able to handle situations quickly as they arise,
24:53and if you want to have a super high level of context with your partners, then these partners, in my
24:59opinion, have to be local,
25:01especially if you're in a critical vertical, like a health care or utility or a telco or a bank or
25:05an insurance company.
25:06You cannot treat agents, like buying agents is not the same as buying SaaS for workflow for an internal employee.
25:15It's usually the weight that sits on the shoulders of those agents is a lot heavier.
25:19Therefore, you need to treat it a lot, a lot more seriously.
25:22And it's very different from buying a traditional piece of software.
25:24So we don't like to be in the vendor box. We call it in wonderful. We try to be partners.
25:29And as partners, we take accountability on the results of our projects together with the customer.
25:34And we don't really get to the finger pointing game. We are too embedded for that to happen.
25:39So just a quick follow up. So how do you find these local partners then?
25:43Like how do I identify like the appropriate partners?
25:48So in partners, you mean customers?
25:51Partners that you work with locally.
25:53Oh, so we don't work with partners.
25:55Okay.
25:55We currently don't work with system integrators or implementation partners.
25:59The only formal partnership we have is with McKinsey because they're just great at strategy and it helps us deploy
26:06better.
26:06Okay.
26:07But we have, so those 30 teams around the world are all wonderful employees.
26:12And I think that's the only way, at least in the next year or two, everything that has to do
26:18with agents and AI is too bespoke and bleeding edge and hard for a company like Wonderful that tries to
26:25be at the forefront and ahead of the curve to trust anyone but like our own people that we hire.
26:31And we make very clear, we make it very clear that in the next two years, they're not going to
26:36have great work life balance and they join regardless.
26:40And then they're, they're super committed to making it work with the customer.
26:43Okay.
26:44So turning to the accountability question and banking, this becomes a more technical question because it's also affects infrastructure.
26:52So I know you have an engineering background.
26:56So far from that perspective, what does it take to move AI agents from pilot to scale?
27:02And what is the major technical challenge there?
27:05Okay.
27:06So that's the exciting question.
27:08First of all, I want to state that the model often is not the problems.
27:13Choosing the model is not the difficulty part.
27:15It's all the harness that you have to put around and make it work.
27:18That's the very important part because, um, I have to say because most people are obsessed, uh, having the latest
27:24model coming up.
27:25Now, I'll quote three, uh, technical challenge.
27:28The first one is identity and permissions.
27:30So agents represent a new actor.
27:33It is not exactly a traditional applications and it's definitely not a human.
27:38So managing its identity is very key to the success of its deployments.
27:42Um, and, and the authorization of what it can do sits really at the intersection of three elements.
27:50The first is what the agent claims it can do.
27:54What our policy allows it to do.
27:58And what is the entity behind the request, either being a human in applications or another agent is authorized to
28:06do.
28:07And being able in real time, uh, on the fly to answer this intersections for every action is the core
28:14engineering problem.
28:16Um, the second thing I would mention is governance versus, um, execution.
28:20So the governance has to be humane, slow and rigorous to make sure that the legal, that the compliance, that
28:26the race departments have approved anything before it goes into production.
28:29But the execution has to be swift.
28:33Uh, we're talking about limited second swift.
28:35It means that the enterprise architectures to make it happen has to separate them at the same time.
28:40Make sure that the, um, some sort of coherence between them, right?
28:44That's why then usually in your architectures, you're going to have a static governance and a more operational one.
28:50The third challenge I would mention, um, is going to be control versus, um, behavioral stability.
28:56So you, you need to have some sort of, uh, brick that, uh, the brick circuit that allows you to,
29:01uh, revoke an agent's right whenever you want.
29:05Uh, even in the middle of a, of something that the, the agent is doing something in the middle of
29:09something.
29:10So, but on the other hand, you want to make sure that you have behavioral stability.
29:14You don't want to anger a session to be, uh, uh, secretly altered without the user's knowledge.
29:19So being able to balance those two is really key to make sure that the, uh, agency platforms, uh, is
29:26production ready for a bank.
29:27Well, I can talk about orchestrations.
29:29I can talk about or guard rails.
29:30I can talk about 70 clears.
29:32There's, there's so many things.
29:33And I believe the gift is really that it's a very exciting, very exciting topic.
29:38And maybe one last thing is that evaluations is really, really key.
29:42Uh, and, uh, no agent should go into production, obviously, before pre-production testing.
29:46But also the, um, deployment must be, uh, the deployment must be gradual.
29:51And that we have the capabilities to really, uh, roll back if the, if the quality really degrades.
29:58Uh, and I hope this make people remember.
30:02Um, I, I believe we can even not say that we're deploying agents, but we're admitting agents.
30:08Admitting.
30:09Yes.
30:10And, uh, just, uh, to address quickly the challenges that you just mentioned, uh, Sue.
30:15So for how at the bank are you actually tackling those challenges that you just mentioned then?
30:22Um, so, uh, from a technical?
30:26Yeah.
30:26Okay.
30:27Uh, well, it's a lot of work.
30:30Um, uh, so we, we really focus on, um, I think for one of the previous questions, I answered that
30:38we really focus on, uh, how we're putting up the conditions.
30:41Which means that there's really a, uh, agentic program at the group levels that works out what is the, um,
30:48enterprise architecture should be.
30:50What are the different blocks?
30:52Um, how, what are the, uh, communication protocols?
30:55What are we going to use for the, uh, agent registries?
30:57How, uh, what cryptographic, um, uh, methods we're going to use to ensure that we, we can trace back the
31:05intent, uh, of the entities
31:07behind the agent, uh, in relation to its actions.
31:10Uh, that we're, well, defining the universe as of applications and databases we can interact with.
31:16What is the appropriate level of observability?
31:18And what is the operating model around it to make sure that we can remediate in, in the world where
31:24agents going to move thousands of times per second.
31:28And we have the appropriate, uh, orchestration layers to make sure that different agents are doing things in relation to
31:34our goal that we set to it.
31:36Um, and in the, in the most efficient way because, uh, cost is definitely a, a very important topic, uh,
31:43uh, for the whole industry, uh, uh, right now.
31:46Thank you, Sue.
31:48And Mohammed, I want to, uh, bring this very important question.
31:51I think in a lot of people's minds, um, connecting this back to the future of labor and work.
31:58IBM has described AI as changing how work gets done across sectors, not just how tasks are automated.
32:05So what does that mean for the future consultant or manager?
32:09Yeah, no, that's a, that's a great question.
32:11And, um, it's sort of at the heart of not just consulting, but every labor-based business.
32:15But I actually want to make a comment on what Sue said.
32:19I mean, you can, just by listening to Sue, you know he has implemented this because he's been through the,
32:25the, the heartaches and pains.
32:27And one of the things that he said is that it's now less about the exact model that you use
32:31and more about that harness to, um, to oversee, to govern, to observe, uh, to roll back.
32:41And, you know, I'd say in our experience with, with clients, many of our large clients are a hundred billion
32:49dollar clients have actually built such a harness internally.
32:53And the most mid-sized companies, you know, even a $10 billion company have, have not been able to build
32:59those kinds of harnesses.
33:00And so before you actually go try to deploy this in a productive way, you should think about that harness
33:08that, uh, that Sue was talking about.
33:09Okay. Now let's go to the question about how all of this AI is going to affect labor.
33:14So, you know, I've spent 30 years of my career in software and I've never been a consultant before.
33:20And about three years ago, the CEO of IBM called me and said, Hey, would you come to IBM and,
33:25and run the consulting business?
33:26And the idea there was that, um, why would you want somebody with a software background running a consulting business?
33:33Because, uh, the world is moving from, from just human labor to a combination of human plus digital labor.
33:40And digital labor is a bunch of bits of software, right? That have to be integrated into a management system.
33:47And so we need to take a software mentality to building this.
33:50This, that's actually why we started three years ago with building this, um, kind of Uber harness,
33:55because in some ways that's your management platform for your digital labor.
34:00And managers are going to have to learn how to manage a team of human laborers and digital laborers.
34:06Now, you know, there's this question about, like, is, is AI going to replace human labor?
34:12And, and to be quite honest, I don't think we know, right?
34:15I mean, I'm actually quite an optimist about this because every time in the last hundred years, thousand years,
34:21there have been any kind of major technological inflection, the GDP has actually gone up.
34:27And what we've seen in the service, in the IT services business is not that long ago, 20 years ago,
34:33we were building hundreds of large applications for clients, like SAP applications and so forth.
34:38Then we started building thousands of small applications called mobile apps,
34:44because the iPhone and so forth came into existence.
34:47And now we're building millions of agents for clients.
34:52And, um, as Ben talked about, a lot of this is co-developed.
34:56Like, it's actually, like, IT work to develop this, to integrate it, to put it into the harnesses,
35:03to make it safe, uh, to make it productive, to get the returns.
35:07And so we're actually seeing, like, an acceleration on that part of our business.
35:11And that part of our business now is about a $10 billion book of business.
35:14In first quarter represented 40% of all of our signings.
35:19It's growing substantially faster than the rest of the business.
35:23And so I think the companies that sort of lean into AI will actually see growth.
35:29And we're actually seeing that, and I'll mention just one client that's doing exactly this.
35:33Has anybody here heard of a company called Pearson Learning?
35:36Raise your hand if you have.
35:38Okay, that's great.
35:39So they're one of the largest textbook publishers, right?
35:41So they started out with hard paper textbooks, and then they went to digital textbooks.
35:47Then they realized that there's this whole certification of skills.
35:52So now they're the largest certification of skills company.
35:55So in your own company, if you have to take, you have to pass your cybersecurity training,
36:00you might get a badge from Pearson Learning in a tool called Credly.
36:04So they've been certifying human beings.
36:06So they came to us and said, hey, there are going to be millions of digital workers out there, agents.
36:11We need to build a whole brand new business to certify digital workers.
36:17And so we have worked with them to help them build a brand new business.
36:20So there are entire new categories of businesses coming into existence that couldn't have existed before.
36:26And those are creating jobs.
36:29That's very optimistic and a great example as well.
36:32Now we have a couple of minutes left.
36:34Let's end with the quick round.
36:37What is, this question is for everybody.
36:40So what is the one task you would happily hand over to an AI agent today?
36:45And what is the one task that you would absolutely not give to an AI agent?
36:50Let's start with you, Bar.
36:54Well, so I think in Wonderful, like we are less than two years old and as AI native as it
37:01gets.
37:01So we do almost everything with AI agents and apps and workflows.
37:06Like we barely have any business software.
37:08We build everything on top of Wonderful.
37:10So a good example of something that I use an agent all the time for is recapping my site visit.
37:17So we have 30 offices around the world.
37:19I came here from South America the week before I was in APAC.
37:23So I visit between three to four different countries a week in 2026.
37:28My wife is very heavy about it.
37:32And I want to write a proper summary of the visit.
37:35And like this, it has depth, right?
37:37Who did you meet?
37:37What did you feel?
37:38Et cetera, et cetera.
37:39So I have an agent that prepares the entire thing and sends it for me.
37:42There is a wonderful app, sends it for me to approval.
37:44I edit it, et cetera.
37:45This is something that to get to a level of consistency when I'm like half dead on a plane would
37:50be impossible to do without an AI.
37:52What would happen is I would have to wait until I'm back and then start to remember what happened two
37:56weeks ago.
37:57So this is something that I'm trying to identify completely.
38:00And it's almost there because I still want to review it to make sure that I'm not saying anything bad
38:05about anyone I'm seeing.
38:06One thing that I'm not looking to identify at all is hiring.
38:10I'm obsessed with hiring hands on.
38:14So I interview, me and Rui, my co-founder, we interview 100% of all employees.
38:18So we have 600 people now and we interview between five and seven interviews a day.
38:24They are short.
38:25And I'm not going to give this up anytime soon.
38:28I guess that's because being the final interviewer for a candidate that you want to bring to a company that
38:35is like your baby, it's like it requires an intelligence level that I have.
38:42I'm not yet seeing from from AI, right, like extremely small nuances like facial expressions delays between like so that's
38:50something that I'm not looking to to move over to.
38:52I wouldn't have an AI agent interview an employee that I want to bring to wonderful yet.
38:59I think that's a reassuring to a lot of people here.
39:02The hiring remains 100% human and the recap agentic I think is actually useful to a lot of us
39:09at Viva Tech here as well.
39:11So thank you.
39:13And Sue, one task that you would absolutely identify and one that you would not.
39:19Sure.
39:20I'm going to be very short.
39:22Minutes for my meetings.
39:24Oh, my God.
39:24AI, please take it.
39:27But the choice of the restaurant, well, when you work that long for the financial sector, you know that some
39:33risks are just too high.
39:34So definitely not.
39:36And Eleanor, for you.
39:37So in my words, I would give to an agent pretty much anything an analyst would do.
39:43From modeling to analysis to planning, I would give everything.
39:45But the one thing I would never give up is a final decision making.
39:49At least for now.
39:50Okay.
39:50Maybe next year I'll tell you I give that too.
39:52But for now, I keep the final decision.
39:53And we'll hear from you next year, Viva Tech.
39:56Yeah.
39:56So I think this is similar, right?
39:57So things that could be like processes that could be automated or well-defined, you know, agents can certainly help
40:03there.
40:04Things that require deep judgment, that's the realm of human beings.
40:08And it will always be.
40:11You know, but even recruiting, I think, falls into the first category for certain types of recruiting.
40:19So there's a hospital in the United States that has 120,000 employees.
40:23And the head of recruiting said, my biggest challenge is recruiting nurses.
40:27Like, we just don't have enough of them.
40:29And so we did a project with them and now she can recruit nurses 12 days faster, right?
40:35And that's because we've inserted AI into certain parts of the process.
40:38There's still human judgment.
40:41I would say that there are other types of scenarios where you have to be very careful.
40:45We have a healthcare insurance company where we're now handling tens of millions of inbound calls.
40:52And, you know, when people call in and say, I have this diagnosis and I need to go see this
40:56doctor.
40:56Well, the agents need to have empathy and they need to go find the doctors that they can provide to
41:03the person.
41:04But they can't go to judgment.
41:07They can't be making, like, medical recommendations and so forth, right?
41:10So we have to be very careful about where that line is, especially since we are actually implementing this for
41:16clients.
41:17Thank you, Mohamed.
41:18What a great panel.
41:19This conversation really showed us four very different but connected sides of the AI agent story.
41:25Mohamed, thank you for reminding us the importance of mindset shift and readiness and the guard wells in place for
41:32AI agents.
41:33Eleanor highlighting working together on one infrastructure is really important.
41:38Technology to scale and transparency and accountability to help companies work better.
41:47Sue, thank you for your banking perspective.
41:50You know, authorization access, governance has to be remain human.
41:55And, but execution has to be swift, right?
41:58And then bar, customer facing agents, really important.
42:01Only when they are useful, they can take over human operators.
42:05So, thank you so much.
42:07Whether you are an enterprise leader hoping to revolutionize your operating workflows,
42:13a founder building the next wave of innovation,
42:15or just a consumer like myself looking forward to have an agent that's actually useful and trustworthy.
42:22I hope this session has been helpful to you.
42:24Thank you so much for joining us and enjoy the rest of VivaTech.
42:27Goodbye.
Comments