McKinsey & Company: Generative AI and CX – The Big Picture – Boost Camp 2023

And I think there's going to be a lot of good chatter
for this first speaker who comes from
there.
That comes from an organization that I really don't think needs any introduction.
But I'm American and I like to hear myself talk, so I'm going to do it anyway.
McKinsey and Company is a global management
consulting company that's leading the way in digital transformation.
John has participated in numerous webinars and research initiatives on how
AI is going to be impacting business and moving forward.
So I'm really looking forward to this talk.
So please welcome John from McKinsey to the stage.
Thanks, Clay.
What a number two or three on this page, which I'll flip to.
But before I kick off a little bit about me,
Clay already did a great introduction to McKinsey and a little bit about myself.
And then thank you also Jerry and the rest of the Booth team for inviting me.
My training and background is actually electrical engineering
and machine learning applications there, so you might be wondering
why did he go into consulting to just give advice to people?
It turns out actually McKinsey, half of our more than half of our hours
today is spent on machine learning, data, science and engineering.
So while we are traditionally a management consulting firm,
we're actually more participating in digital these days.
Now, myself,
I spend my time in customer operations software
and excuse me,
I've been in it for about five years, specifically in conversational
and AI and customer operations more broadly
for about a decade.
So as much as I love engineering and technology,
what I also like about the work that I do at McKinsey is it forces me to
take a step back from
what's the technology itself
to what's the market and the demand and the challenge that it's solving.
And that's why I wanted to start here.
The numbers you see come from a survey that we run
every two years and from last year's survey,
despite the decades that we've heard of digital self-serve, is going to go up.
There's going to be the data voice, what you see there on the left hand side,
contact rates keep going up year over year and demand keeps going up.
It's.
Meanwhile, more recently, we've seen increased employee attrition
and lack of integration across the technology stack.
So while your demand is going up, the ability to solve the that demand
is actually going down and becoming harder than it was previously.
So what are we doing about it?
In that same survey, over half of the people that we polled
and these are executives at some of our clients, over 200 around the world
said that their budget is going to go towards
some form of AI or digital application in the next 2 to 3 years,
and you see it across the entire customer funnel.
How are we going to enhance self-serve?
How are we going to improve access to different channels for our customers?
How are we going to enable our agents to solve
the problems of our customers better?
It's not going to be a single place in this stack.
It's going to be across the stack.
And lastly on the bottom there, how are we actually going
to make our management more effective at solving those problems?
I think one of the biggest problems I see with our clients today
is that there's this operational gridlock for so long.
The way that we've been improving customer operations has been incredibly manual.
How are we randomly sampling calls to then
go into a manual review of the process in the way that we're reviewing it
to the manually redesign and implemented in the knowledge base RPA Business Process
Management, Next Best action, all these different technologies out there.
So actually one of the things that we're most excited about
are some of the tools that are coming to market in the applications
within generative AI that are helping reduce some of that gridlock.
And we'll talk a little bit more about that in a moment.
But before I get there, I think one of the most important things
in these conversations that we have about generative
AI isn't to educate, but it's to level sense because there's
a lot happening in the last six months to what Jerry was saying.
Chat CBT is the lowest common denominator
that the masses around the world understand.
But in fact this technology is well-proven and is going to exist for much longer.
The birth of it is pointed to widely across
the industry with the start of a Google pay for called attention is all you Need.
That came out in 2017.
Google then led the way actually
with the first large language model called Bert,
then open.
I came along now to achieve it.
If we fast forward a little bit, it was not an AI breakthrough.
GPT three was the underlying AI algorithm for chat. CBT
It's then what was it?
It was a user interface breakthrough.
It was the ability to engage with a large language model in the way
that the general public could understand and experience and experiment with.
So it's really important as we look backwards to understand
this is market proven technology.
This isn't something new.
What's new is the experience
that the mass public is starting to understand and engage with.
But then as we look forward,
what you see happening right now with Jeopardy
for Llama Alpaca
and all these other names Palm that I really don't know how they
come up with, but they do.
Performance is increasing at an exponential rate right now
and these large language models that will not go on forever.
So what's happening next?
What are some of these highlighted for
you see geographical expansion with buy these are anybody
you see open sourcing with Stanford's alpaca and met Islamists
and then you see more consolidation around
some of the major players like Amazon and Google.
What does that point to in a market
that's changing rapidly and it's tough to keep track with
looking ahead when performance improvements slow down,
we're going to be looking at cost efficiencies, open sourcing,
how do we make this more available and accessible to everyone?
And we're also going to be looking at distribution
to the point on accessibility.
How are some of the major players creating platforms around these tools
and distributed out to the masses and make it available to enterprises?
Because right now it's mostly consumer products in the way that we see chat, CBT
So then as we look at
what is all the excitement coming from
on the right hand side, I think what you have here is
the way that a lot of people understand generative AI, It's in the title,
it's the ability to create across a number of different modalities
large language models, text
images with Dolly too, and stability coding with GitHub, copilot
and others you don't see here audio, video and the like.
Okay.
But natural language, understanding
and processing existed before generative AI.
So why is generative AI interesting?
And I think it's this here on the left hand side, the ability to understand
and right at the center of it
is that transformer model which came out of the Google paper.
If you think about the first time, you might have interactive the chatbot.
You probably put a question in that you could answer with a Google keyword
search just to kind of test out, see what happens.
Okay. Came back with a good answer
that I get to use Google to do.
That would have taken me 30 seconds.
What if I give you a question that I would ask Wikipedia and spend
maybe 5 to 10 minutes reading Wikipedia, researching to understand?
Okay, well, actually
came back with a pretty decent answer.
What if I ask you a question that would take me 2 to 3 hours
of internet research to understand I have these ingredients in my refrigerator.
Help me plan a healthy meal.
I'm looking to get in a 30 minute workout because I'm jetlagged.
I'm staying in a hotel with, you know, basic dumbbells and some treadmills.
What's good workout? I could do?
Oh, wow, That's an impressive answer.
Came back with.
So it's that ability to understand that's really capturing the minds
of the public right now.
Because while the natural language technology existed,
the level at which it understands the questions we're asking
and how to form a response to it is where the real value
is coming from.
But we're here to talk about customer experience, not technology.
I told you, do you like my technology?
So as McKinsey,
how do we see this coming together in customer experience?
Number one, it's going to touch customers, agents and management.
Management is very important here to the point
I made earlier around that gridlock.
Some of the technologies that we've been deploying more recently
in the last 5 to 10 years, let's say
most of them are technologies for management.
So what does that lead to?
We used to talk about this in waves,
but how quickly we're seeing the space evolve.
And then as we think about what it takes to deploy a use case for test
for generative A.I.
customer self-service, I'll use the dirty word, but
virtual agents, a copilot for your agents,
or even the insights hub for your management.
We think it's actually going across all of these.
First, how do you remove that gridlock
of operational debt and start to work through the backlog?
You're looking at your productivity boosts last agent in management time
spent on non value add technologies and manual tasks.
Let's get rid of those first.
Now that we relieve some of that pressure, where can we go from there?
Service innovation is maybe one of the most important aspects that we see here.
Air itself is different than traditional software.
Put the same thing in twice.
You can get two different things out.
It's very hard to train and train and test that.
So what if we took our workforce and we made them part of that,
made them the contributors today?
Are I training and testing?
Give them that little thumbs up, thumbs down.
Now they're innovating the service, they're providing that feedback
and as
this becomes more standard, what we're seeing is behavioral shifts.
You'll see a chart later.
But at the end of the day,
as people really start to realize that they have these tools available to them,
and their nature of the nature of their role is changing.
We're seeing that they are more engaged.
They want to drive customers and participate in their own work
towards something that is more engaging and they become more affluent
and they become promoters of the product themselves.
So then what's the roadmap
in the sequence for how we do this?
We're seeing some early traction in the use cases
you see up here, but it's a little bit more than that.
While each one of these has value by themselves independently,
the capabilities of generative, I actually actually run across these.
So now we're
starting to think a little bit more about, well,
PTA is the lowest common denominator.
Like I said earlier, people will jump to
I want to do a self-serve, but kind of like churchy PTA.
But then they start to think about it.
I've heard about these hallucination things.
Sometimes it just makes up data.
Maybe I don't want to start with customers.
Maybe I'll give it to my agents first and let them try it out.
But then they start to think like, gosh, all the data
we have that we use to train it, it's pretty bad.
We don't want to replicate a wasteful process.
We want to empower our agents with a new innovative process.
So how is our actual processes designed?
And this is where we get into the sequencing standpoint, starting with
continuously improving your knowledge base, using both the knowledge base itself
as well as other sources like your contact recordings, your chats, your calls,
and tying that to agent performance.
You can figure out
where is this working?
Well, how are my best agents doing this?
I'm going to extract that trap knowledge and I'm going to use content generation
to actually start authoring some of these knowledge base articles.
Great.
Now that I have that real time data, especially on voice,
is incredibly difficult to integrate with and respond very rapidly.
So I'm going to start
to build out use cases they're taking the life data
and then from my agents with that optimized knowledge I designed for them.
Now that I've overcome the real time data hurdle,
I'm going to go back to that customer self-serve case.
But I'm still worried about some of this.
How do I know that I'm ready to go
build the integrations, invest the capital that's required
because a lot of my systems are actually quite disconnected
towards building that self-serve use case.
Well, I mean, ask my experts, agents themselves,
they're going to give that thumbs up, thumbs down.
And by providing that feedback, once I cross that confidence threshold,
that's when I know I'm going to be ready to deploy to customers
and really deliver that self-serve experience.
So as we're looking at the roadmap and how you sequence
your technology deployment, we see it starting with removing that operating debt
and innovating on your processing, empowering your people to be humans
in the loop and provide the feedback and then moving on
to actually deploying it to customers.
So what's the end result?
We looked at this about two months ago to figure out
how the demands would shift in customer contact rates.
June 14th, you're going to see a larger article
come out from us on the overall macroeconomic impacts of Generative IV.
It's part of this research that was going to have a lot more.
So where we see it today,
about 25% of contacts are on made
in the future. We see 50%.
Some of the numbers you might have heard, it's actually a lot
less than that, surprisingly. Why is that?
First of all, the contacts, you don't want to automate
fraud and safety, fraud, health and safety very high touch, sensitive things
where you want that human involvement and that empathy of a bot.
Just the simple knowledge of knowing you're working with a bot
could have a very adverse effect.
Next limitations.
I made the point open I in Chad CBT as a consumer product in its current form.
I'm not sure about sending my PI,
my high off to the cloud.
We're in Europe.
Different countries have different regulations around data
and how it's stored.
That's very hard to control if you're just sending it off to an API.
These are known issues, but they're still working on it for the enterprise.
How are you going to deploy on prem?
We see companies emerging around that, what will be the data security.
But then lastly, as I mentioned, demand shifts earlier.
You see it's relatively flat.
The first three years
and then the manual contact rate goes down, automated goes up.
So one interesting thing, automation can continue to absorb the growth of contact.
So we expect to continue.
That's actually really good news.
A lot of the clients we talk to you, they're at 95% utilization rates.
It's very hard to keep people.
We can absorb that.
Then it goes down.
What's actually happening there though,
that we're not seeing here?
The length of time people are spending on the phone is going up.
We're in conversation.
And that's actually good news because it means with your workforce
today, you can still continue to maintain them.
You can shift them towards these more complex conversations
that actually create better engagement with your customers
and use technology to remove all of the waste that's happening today.
And we're pretty excited about that.
So this is my last page
as you think about how to get started and the different ways
you will deploy this technology, we are at a technology conference.
We see four ways emerging.
On the left hand side, you have your software as a service application.
It is an AI application for a product much like this.
Then you see AI's sorry APIs like Eyes and others that are coming out
and then you see up going to the far right, getting into managed services,
then actually standing up your own private cloud, depending on where
people are in terms of the technology availability, how you're going to spend
your budget, the amount of data scientists you have on your staff.
It helps guide
what you might be selecting.
But one of the things we fundamentally believe is
this will not be a monolithic technology stack.
Every company will have a blend of all of these approaches.
It'll depend on the use case.
And then as you think about the use case,
you're going to be thinking about what's the value I get from the left hand
side versus the right hand side versus the effort it's going to take to put in.
As I think about that roadmap that I showed earlier,
can I be stacking capabilities on themselves
As I stack those capabilities,
can I get increasingly better value out of using those capabilities
and redeploying them in different use cases and contact reasons?
So what are we doing and what are we encouraging our clients to go to?
Number one, get started. Now,
if you wait a year saying, well,
the technology is changing fast, I'm a little nervous,
you will 100% be behind your competitors a year from now.
Okay.
But John,
you just said
the technology is changing fast
and we don't know where it might all land in a year.
And he just told me to get started.
Despite all of that. How do I do that?
The second thing we're doing is we're saying
try a number of use cases and pilots to get going.
It's not usually the McKinsey way.
For those of you
who might have interacted with us, we like a good strategy diagnostic.
We make this nice little chart.
It's two axes, impact and feasibility.
We go through this whole four week ish, 4 to 6 weeks process,
look at all the different use cases.
Say these are the first two you should go after.
Now let's talk about planning and launching
because this
technology is evolving so rapidly and it's so easy
and accessible to get to a pilot within about that same time period,
4 to 6 weeks, we're saying, you know what, Take a different approach.
Think about this like an investment portfolio.
Pick five use cases.
Most the conversations I'm in, if I say pick five, use cases to every executive
in that conversation, 80% of them agree on at least three of those.
It's not that hard.
You don't need a big McKinsey diagnostic to tell you that
you know your business, you know where the challenges have been.
You see the potential of this technology.
Most people agree
from there,
how do we de-risk this in a way, knowing the technology is changing
while still placing strategic bets on building out our technology stack?
Okay.
We're going to launch three of those pilots.
Some of them are going to be strategic bets.
We're on the right hand side.
Some of them are going to be quick wins, obvious value.
We're on the left hand side.
In two months time, we're going to check in.
We're going to say, how is our investment going?
We've set aside more budget than we need for two months,
and we're not going to pick every single pilot that we've run to continue on.
And that's okay.
This is in a learning phase.
It's a volatile growth market.
You can't get that 100% sorry
American analogy.
I'm trying to think of a soccer team even calling it soccer.
You can't get a hat trick every time.
Not every shot on goal is going in the net.
There we go. That's a good analogy.
Not every shot on goal is going in the net.
And that's okay because you're learning about behaviors
and the way that your organization is going to respond.
So in that two month check ins, take the capital you've set aside,
put it towards 1 to 2 of the use cases that have that strategic importance
and are going well.
And think about how you scale,
because I think the thing that I want to leave you with is that scaling
the piloting is incredibly easy with this technology.
Scaling is incredibly hard
because of some of those limitations I mentioned.
This technology is not a silver bullet.
In fact, it's only about 20% of the solution.
There's going to be things of
how do we integrate with other technologies in our stack.
Some of them are in a platform that's cloud deployed.
Some of them might be in mainframes.
Those all exist today.
This isn't greenfield generative.
A.I. does not solve all of it.
But more
importantly, there's going to be customer experience design,
employee experience, design, change management, adoption,
all these other things that sit around the technology
and are required to make it successful.
One of my clients, where we did this at scale in North America,
they ran departments,
one for people answering questions, What's my least price?
Do you allow grills?
They're spending at least 2 hours per person a day
responding these sorts of things.
Okay, well, we automated that great
promise. One for people.
How do you get 20% of value out of that?
So rather than looking at it as a local staffing or sorry,
an independent site staffing model, one of four people for site, we said
what if we did it locally?
What if we made it 5 to 10 apartment buildings?
These people are working across and got value out that way.
Okay,
that'll work.
But what do we need to do as a result of it?
Well,
the way people are working right now,
they've been having this conversation back and forth with this individual
and building a relationship but
they're going to need to shift to a way where they can just log in, understand
where they are in the day, pick up the first task on the list,
pick up the conversations if they know the person.
That wasn't just the guy that did that for them,
that was redesigning their salesforce interface,
that was redesigning the work that they were doing
and training them on how to do it and still feel like they were
providing the same value because people want to provide value in their jobs.
So as you think about how do I get started?
How do I start to pilot this knowing it's moving quickly,
but then where do I continue to go deep and really do what it takes to scale?
Think about it in steps,
not in the near long term, but how you just put your bets,
how you then
test it later, how you measure the performance,
and then how you continue to go deep on those certain use cases.
So with that, I'm at the end of my time
the so much everyone thanks again to the Boost team
will be around more during the day if you'd love to talk and than that
I look forward to having the next I think Clay I'm going to hand off to you.
Thank you.