McKinsey’s CFO on the benefits and costs of AI

**Peter Thal Larsen** (0:04)
If you talk to the Chief Financial Officer at any large company these days, the conversation will very quickly turn to artificial intelligence, and specifically, the cost of AI. Corporate executives everywhere are trying to work out how AI could upend their business or give them a competitive advantage. But CFOs have to weigh up the costs as well as the potential benefits. In other words, they must think about the return on investment, the ROI on AI.
So this week on The Big View, we're going to dig into the debits and credits of corporate AI. It's what we do at Reuters Breakingviews. We tap our best sources around the world for fresh insights into the biggest questions in global business, finance and economics. I'm your host, Peter Thal Larsen.
And today, we've got the consumer guest to help us tackle this question. His name is Yuval Atsmon, and he's the Chief Financial Officer of McKinsey & Company, the global consulting firm. In addition to his management role, Yuval is also a senior partner in the firm's technology, media and telecommunications, and consumer packaged goods practices. So he's not only overseeing the implementation of AI at McKinsey, but also advising clients on how to do it at the same time. Based in London, he previously spent six years in Shanghai. He has an MBA from Harvard Business School and a law degree from Tel Aviv University. He joins me in the studio in London. Yuval Atsmon, welcome to The Big View.

**Yuval Atsmon** (1:37)
Thank you. It's great to be here.

**Peter Thal Larsen** (1:40)
Let's start with McKinsey. McKinsey spends a lot of time advising other companies about their strategy and their business, and obviously how to use AI. But I want to turn the lens back on McKinsey itself.
What are some of the most tangible tasks and applications where McKinsey is really using AI at the moment?

**Yuval Atsmon** (2:05)
So we are at a moment, I think, that we don't even fully know what would be the full potential of AI in terms of changing how we work, and one of the mindset that we have taken very early, and if you would like to sense the CHPT moment, is we want to be the fastest learning, fastest adopter organization, knowing to some degree in advance that we are going to try a lot of stuff that is not going to fully work. So very early, we decided that we are going to create our own GenAI platform, still, of course, supported, not our own large language model, but still supported by what is available from others, but with a significant ability to, of course, protect our IP and leverage a lot of our own knowledge base.
We even named it after the first professional woman that we've ever hired in McKinsey. Her name was Lillian. We called it Lillie. We launched it maybe less than a year after, I think, seven, eight months after. It was probably the biggest internal and fastest delivered. It was sort of our own vaccine to COVID in terms of kind of energy and speed.

**Peter Thal Larsen** (3:11)
So seven, eight months after ChatGPT.

**Yuval Atsmon** (3:15)
Yes, we already had that released. It is, we've iterated in a lot since.
So it's very different now than what it was in the beginning. But even from the start, it gave us a real opportunity to test almost anything that we wanted to do all the way from, can we create our own slides quickly from that tool? The answer was no for a while. Can we create a deliverable for a client in terms of, even if not in a PowerPoint, but in other ways?
But also we've learned that there's a lot we can do. Even before some of the models have improved into reasoning and gentrification, we were already able to, for example, create people with their own unique profile of their own mini LLMs to know all the materials that they already had. We were able to drive pretty quickly to close to 95% weekly user adoption of that tool. And I think that has given us already a start of taking that stuff faster to the way we change the way we work with clients. Of course, over the last 12 months, we've seen another big change of the technology. We've seen an acceleration of the models, mostly most notably from Anthropic, OpenAI with things like Cloud Co-work and Cloud Code and Codecs. And generally, they have become a lot more enterprise ready in terms of the way that they support our clients, but also us and tools like Cursor and others that started with coding tools, but they are bringing coding capabilities into also, how do you accelerate the way that you research things? How do you accelerate the way that you prepare a full document on a topic? How do you accelerate the dashboards that we can now bring into clients? And rather than sending them a PowerPoint to look at, we can actually spend time in a dynamic way, because our consultants are basically programmed an interactive tool, which is still at the back as the same insights we had previously put in to other forms of presentation. But now we can work real time, including, we typically have a process that before we get to the senior executives, we would want to make sure that we have validated some of the work, or we have helped their own teams to ensure this is something they can implement, they can apply, or in some cases even helping them implement and apply. So those tools are now enabling us to go faster, if you'd like, into the stage of, in a more granular way, into the stage of how is this going to shift the way you work? I mean, just to give you an example, if you imagine that you're an automotive company that can apply into the planning that goes into the manufacturing lines, that can apply into the target setting and way of managing dealerships. So instead of the classic, maybe for us, multi-step process where we would start by mapping, where could you create more value? You can produce at a lower cost, you can sell better to your customer and everything in between, and then we're going to spend three months prioritizing between those opportunities. We can then go and create tangible specific initiatives against that with the broader organization, and we're going to set a three-year implementation plan. Now, we can use those dashboards that are AI-powered and the analysis that is AI-powered to go faster so we can actually be testing stuff with the manufacturing team or with the sales or with the distribution dealership teams, et cetera, already within a few weeks from the moment we started.