AWS re:Invent 2023 - Transforming the consumer packaged goods industry with generative AI (CPG203)
Summary
TLDRMichael Connor, leading the generative AI practice for a global CPG retail team, shares insights from engaging with 1,000 C-suite and board members on the transformative impact of generative AI on business. He discusses the technology's potential $7 trillion GDP impact, use cases in marketing, sales, and product R&D, and the importance of a robust data strategy. Connor also highlights innovative applications like advertising personalization and product concept generation, emphasizing the need for change management and responsible use frameworks to leverage generative AI effectively.
Takeaways
- π Michael Connor leads the generative AI practice for the global CPG retail team and has observed significant interest and adoption of generative AI in the last six months.
- π‘ The technology is predicted to have a $7 trillion impact on GDP, indicating a massive potential for business transformation.
- π Connor has engaged with over 1,000 C-suite and board members, noting a significant shift in business strategies to incorporate generative AI.
- π Investments in generative AI are prominent, with examples such as the $4 billion investment in Anthropic, highlighting the technology's potential.
- π’ Advertising is being revolutionized by generative AI, with the ability to create personalized content and engage consumers in new ways.
- π οΈ Use cases for generative AI span across various departments including marketing, sales, customer operations, product R&D, and software engineering.
- π There's a clear progression in customer engagement with generative AI, moving from understanding the technology to executing and integrating it into business operations.
- π§βπ» Generative AI differs from traditional machine learning by utilizing a vast array of internet data to create models capable of multiple tasks, such as code generation and image creation.
- πΌ A robust data strategy is essential for leveraging generative AI, emphasizing the need for data hygiene, governance, and a chief data officer role.
- π§ Tools like ComfyUI and LangChain enable the creation of AI-driven solutions without coding, facilitating rapid development and deployment of generative AI applications.
- π± The future of retail may involve conversational interfaces with product catalogs, allowing for more personalized and dynamic shopping experiences.
Q & A
What is Michael Connor's role at the company?
-Michael Connor leads the generative AI practice for the CPG retail team globally.
What is the estimated impact of generative AI on GDP according to the script?
-The estimated impact of generative AI on GDP is $7 trillion.
What is the significance of the investment in Anthropic mentioned in the script?
-The company invested $4 billion in Anthropic, which makes Claude, Michael Connor's favorite model.
How is generative AI changing advertising according to Michael Connor?
-Generative AI is changing advertising by allowing for the creation of personalized ads and product concepts, potentially leading to a permanent shift in the industry.
What is the role of generative AI in product development according to the transcript?
-Generative AI is used in product development for creating product concepts, generating 2D images, 3D CAD models, and even drafting press releases and product descriptions.
What are the three phases customers go through when engaging with generative AI as described in the script?
-The three phases are: understanding what the technology is, discovering use cases, and executing on the technology by organizing as a business and considering ethics and bias.
What is the 'zero-ETL era' mentioned by Michael Connor?
-The 'zero-ETL era' refers to a future state where the need for ETL jobs, which involve transferring data from one system to another, will be eliminated due to advancements in technology.
How does Michael Connor describe the difference between machine learning and generative AI?
-Machine learning involves training a model on one task with a dataset, whereas generative AI uses data from the internet to train a model capable of doing many different tasks.
What is the importance of a good data strategy for leveraging generative AI, as emphasized in the script?
-A good data strategy is crucial for taking advantage of generative AI because it ensures that the data is well-organized, clean, and accessible for the AI to generate accurate and relevant outputs.
What is LangChain and how does it relate to generative AI?
-LangChain is a framework that helps glue various systems together, allowing generative AI models to interact with databases, knowledge bases, product catalogs, and other resources to provide more informed and context-aware responses.
What is the recommended approach for companies to start utilizing generative AI according to the script?
-The recommended approach is to start with high-impact, low-effort use cases, such as call center applications with a human-in-the-loop, and to develop a responsible use framework that considers ethics, bias, and legal implications.
Outlines
π Generative AI's Impact on Business
Michael Connor, leader of the generative AI practice for a global CPG retail team, shares his insights from the past six months. He discusses the significant interest in generative AI among business leaders and the technology's projected $7 trillion impact on GDP. Connor highlights the engagement of board members and executives who are keen on integrating AI into their business strategies, emphasizing the importance of aligning with business goals, considering privacy, and ethics. He also mentions investments in generative AI models like Anthropic's Claude and Meta's celebrity integrations, suggesting a transformative shift in advertising and retail environments.
π Accelerating Business with Generative AI
Connor discusses the evolution of customer engagement with generative AI, moving from understanding the technology to identifying use cases and finally executing on these cases. He outlines the transformative applications of AI in areas like product development, advertising, and merchandising. Connor also shares his observations on the three phases of customer interaction with AI: initial curiosity, exploration of use cases, and the current phase of production implementation. He emphasizes the importance of a robust data strategy to leverage the full potential of generative AI and shares examples of how companies are using AI to enhance marketing, sales, and product R&D.
π οΈ Practical Applications of Generative AI
In this section, Connor provides practical examples of how generative AI can be applied across various business functions without the need for coding. He demonstrates how AI can be used to generate product descriptions tailored to specific demographics, clean up product catalogs, and even assist in pricing analysis. Connor also showcases how AI can speed up the product development process by generating numerous product ideas and CAD models quickly. Additionally, he discusses the use of AI in manufacturing to help troubleshoot equipment issues by understanding complex manuals, highlighting the versatility of generative AI in enhancing business operations.
π The Future of Retail and Customer Engagement
Connor envisions a future where product catalogs are interactive, allowing customers to have conversations with AI to find products based on their needs and preferences. He discusses how AI can revolutionize the shopping experience by making personalized recommendations. He also touches on the use of AI in sales enablement, helping retail staff recommend products effectively. Connor further explores the potential of AI in call centers, automating quality assurance and summarization of calls, and even translating these summaries into different languages. He also demonstrates how AI can be used to understand and convert legacy code, making modernization of old systems more accessible.
π§ Implementing AI Responsibly and Efficiently
In the final paragraph, Connor emphasizes the importance of a human-in-the-loop approach to ensure the quality and appropriateness of AI-generated content. He advises against over-reliance on training models and suggests starting with out-of-the-box solutions for high impact and low effort. Connor also stresses the need for change management aligned with business outcomes and recommends starting small and scaling up responsibly. He shares a responsible use framework that companies can adopt and customize for their AI initiatives, covering aspects like team diversity, risk management, and legal implications. Connor concludes by inviting the audience to connect with him for further discussions on AI's role in business.
Mindmap
Keywords
π‘Generative AI
π‘Chief Architect
π‘CPG Retail
π‘Data Science
π‘ETL Jobs
π‘Amazon Connect
π‘LangChain
π‘Responsible Use Framework
π‘Human-in-the-loop
π‘Product Descriptions
π‘Change Management
Highlights
Introduction to Michael Connor, leader of generative AI practice for CPG retail globally.
Engagement with board members and senior executives on generative AI's impact on business.
Investment of $4 billion in Anthropic and the development of the Claude model.
Meta's announcement of celebrity interaction through generative AI.
Transformation of advertising through generative AI with the example of Kate's Island lager.
Generative AI's application in retail for reviews and product descriptions.
Customers' shift from understanding technology to executing and organizing for production.
Top use cases for CPG identified as marketing, sales, customer operations, product R&D, and software engineering.
The zero-ETL era and the potential disappearance of ETL jobs.
Generative AI's impact on revenue growth and cost reduction.
Use cases generated for a retailer in a single day showcasing the technology's versatility.
The necessity for a scalable pipeline to roll out numerous use cases across an organization.
The importance of a good data strategy for the success of generative AI.
Examples of generative AI in advertising using Stable Diffusion and ComfyUI.
Concerns about generating human faces and the legal implications.
Demonstration of generating product descriptions tailored to specific demographics.
The role of generative AI in cleaning up product catalogs and generating SEO keywords.
Generative AI's potential to speed up the product development process from concept to market.
Use of generative AI for troubleshooting and repairing manufacturing equipment.
LangChain as a framework for integrating generative AI with business databases and knowledge bases.
Tools like Flowise and Langflow enabling rapid development of AI workflows without coding.
The future of shopping involving conversational interactions with product catalogs.
Application of generative AI in call centers for quality assurance and summarization of calls.
Generative AI's ability to understand and convert code, aiding in migration from legacy systems.
The use of generative AI to create cloud formation templates for rapid infrastructure deployment.
Emphasis on human-in-the-loop use cases for quality assurance.
Advice on starting with generative AI: focus on business outcomes, use out-of-the-box models, and avoid getting stuck in proof-of-concept land.
Responsible use framework for generative AI, including considerations for team diversity, risk, bias, harm, and legal implications.
Transcripts
- It's good to see everyone.
Quick introduction, I'm Michael Connor,
and I lead the generative AI practice
for our CPG retail team globally,
so, last six months have been really fun.
This is all anyone wants to talk about,
so that's what we're gonna cover today.
I have a CPG background, I came from Coca-Cola.
I was chief architect,
and I led a data science innovation team,
so I work with a lot of CPGs,
and I want to talk through, basically,
what I'm seeing with customers.
I've talked to 1,000 C-suite and board members
over the last six months,
and so, I've seen a lot in terms of what people are doing,
use cases, how they're approaching it,
how they're aligning with their business,
how they're thinking through, like, privacy,
and ethics, and things like that,
so, I wanna share a lot about what I've learned, basically,
in the last six months.
So, the biggest things I've learned
is that there's like, huge predictions
in terms of what this technology's gonna do for business.
There's a $7 trillion estimated impact to GDP.
This is like, the first time that I've seen board members
really getting engaged, and senior executives,
where they're seeing a huge shift in their business coming,
and they wanna make sure that everyone
in the organization has a plan,
that they have funding for it,
and they really know what they're doing.
And so, engaging with board members
has been something that's a bit new.
So, you've seen the announcements,
we invested 4 billion in Anthropic.
They make my favorite model, Claude.
Meta announced that they're doing celebrities.
You can now talk to Kendall Jenner 24 hours a day
if that's your thing, (chuckles)
and so, a lot of changes to advertising,
that image on the top-right.
Kate's Island lager, affectionately named
after (chuckles) Kate Wiley, our marketing lead.
It's really changing advertising,
and I think it will permanently,
so, I want to go through some of that stuff,
share use cases, do some demos,
and kind of just share what we're seeing
across the industry, so,
we really think that this is a fundamental change
in technology. I've had a lot of friends that say, hey,
this is NFTs, or it's Bitcoins, or it's Metaverse,
and so, we already see customers going into production.
This is gonna be a transformational shift.
You've seen the announcements, we're all-in on Amazon.
Every group within Amazon is thinking
about how to apply generative AI,
so, we're using this in our retail environment for reviews,
for product descriptions, all across our business,
so, we really think it'll be transformative.
This is just a simple demo that we put together
using a couple things.
One, it's kind of an R&D use case
where we're creating a product concept,
but it's bringing together two things, like imagery,
generating a product concept to 2D images.
Also, we're generating 3D CAD models from text to CAD,
and then also building the press release,
the product descriptions, internal communications,
the social media piece,
and so, across like, products,
we're seeing a lot of generative AI use cases,
especially around the merchandising, advertising,
photo studios, and things like that.
So, I've seen three shifts with customers.
The first two months, they wanted to know
what the technology was.
The second phase was we wanna know what the use cases are,
and this is kind of the use case roll up,
and I'll talk to this in a second.
The third phase is how do we actually execute on this,
organize as a business, get things into production,
think about ethics, and bias, and stuff like that?
So, I'm really seeing a lot of people go into production
in the last couple months.
We did this with McKinsey about five months ago,
but it really holds up,
and so for CPG, we're seeing like, marketing, sales,
customer operations, product R&D,
and software engineering as being the top use cases,
and I thought that this would change,
but it's really stood up.
Retail is kind of similar.
Software engineering is an obvious one, ETL jobs.
When I was at Coca-Cola, we paid hundreds of millions
to companies to do ETL jobs,
transferring data from one system to the next.
I think that that's gonna go away, the zero-ETL era,
and so, marketing and sales is the big one,
so we'll show a couple use cases around there.
Customers always want to grow revenue if they can.
There are some times where there's opportunities,
like in call center, to cut costs,
but I think revenue growth is always where people
wanna focus.
This is a customer I worked with, a COO of a retailer.
We worked with their team, and in one day,
these are all the use cases that we came up with.
And so, a lot of people are thinking, okay,
there's ChatGPT, we can have chat bots,
and they're having a hard time making the leap
from what they see with the tech
to actually how it can help their business,
and so, what we're seeing is every single aspect
of a CPG retail business,
and actually, I think all businesses, is gonna be affected,
and this is just one day of work
coming up with use cases for this business.
So, I think what happens is companies go from thinking,
okay, how do I do a use case,
and how do I think about generative AI
for a specific use case or two,
to all of a sudden you're thinking
about how do I create a pipeline internally,
a scalable pipeline to roll out 500, 1,000,
3,000 use cases across the organization?
So, I'll show you a little bit about drag and drop tools,
how to move fast, trying to get out of code,
but I see a lot of things slowing people down,
and we'll talk about that.
So, I think there's a bit of a trough
of disillusionment with some customers,
where they're thinking GenAI,
they're gonna feed all their SAP data,
and magic is gonna happen,
and so, it's really important to really share with people
the difference between generative AI and AI,
and why they're both important, right?
And so, this is the way I describe it to people,
and I found it helpful.
Machine learning is where we took one data set,
and we trained a model on one task.
The post office did this, of course,
with handwriting recognition,
and of course, deep learning is where we took
a lot of different data,
and we used that to train a specific model,
but we use a lot of different data sources.
So, demand forecast is a good example of that,
so you're looking at historical sales, adjacent sales,
competitor activity, your planogram, weather,
the holiday calendar, all those types of things,
but you still have one model.
So at Amazon, we have a lot of these,
thousands of models around the organization,
our demand forecast, product recommendations,
we've got a lot of these.
They're very good at math, they handle complex tasks.
Generative AI is very different.
So, with generative AI, it's basically taking
all the data from the internet and using it
to train a model that can do a lot
of different stuff, right?
And so, code generation,
these are some of the examples we show,
but code generation, image generation, text, summarization,
and so, it's amazing to me that we have a single model
that can do thousands of things, right?
But we tell people, you have to have the generative AI,
and you have to have the deep learning as well.
So, the punchline to all of this
is that none of it works without a good data strategy.
It'd be good stewards of your data.
So, I talked to companies and executives that said,
"Look, we wanna be data driven,
"We have all this data, but no insights,"
and I talked and preached about the data lake,
and data mesh, and a data catalog,
and data hygiene, and a chief data officer,
and some people took it to heart, some people didn't.
Now, they're taking it more seriously,
'cause we're telling them, look,
to take advantage of GenAI,
you really have to get your house in order
from a data perspective.
So, I think that that gives weight to teams
that wanna do this to go back to the board and say,
all right, time to take this really seriously.
So, I'll show you a couple fun examples.
Advertising is a game of of decimal points,
and if you can build an ad that catches people's eye,
it can really make a big difference,
and return on an ad spend,
and so, this is using Stable Diffusion
on Amazon G4dn instance,
and it's different from something like DALL-E 3
or a Midjourney,
and the reason is because those products,
you give it text, and it spits something out,
it's either good or it's not.
With this, we actually have a lot of control.
This is a drag and drop tool called ComfyUI,
really brilliant stuff,
and it gives us the ability to pull together
a lot of different AI to produce a result.
So in this case, what we're doing is generating an ad,
and what I asked it to do is, you can see this,
I'm naming a specific park, Piedmont Park
in Atlanta, Georgia.
I wanna show the skyline in the background, and I wanna,
I'm actually telling it what the dog name is, too.
So, imagine generating an ad for every single park
in the world, with every single breed,
but especially the breeds that are popular
in that neighborhood, so,
I went ahead and changed it over to a Goldendoodle,
I have a lot of Goldendoodles in my neighborhood,
but essentially, what this does is SDXL
is a model with Stable Diffusion.
It generates the new image, it does a base pass,
it has a refiner pass,
so, it really adds, like, you'll see skin pores,
and nuanced things, like every little detail
in the eye is perfect,
and then it generates the Goldendoodle, right?
So, the AI is really bad at generating product images,
and so, what we've done is we created an overlay,
a PNG here with a product image, a QR code, the messaging,
and then it's overlaying,
and this is what I like about these tools here,
it's overlaying this product.
So, it's kind of fun to go over to Bedrock, and say,
look, gimme a list of all the parks,
gimme a list of all the dog breeds,
and so, we're using the text generation
to then generate all these.
So what we can do now is generate thousands,
thousands of these images.
This is actually the way that the Amazon ads platform
works now.
It generates the background images,
it also generates messaging for each different product
in each different audience segment,
so, I love this stuff.
So, go to the next one.
Yeah, so this is an example of all the different dogs
that we created, right?
I've noticed that a lot of brands are really squeamish
about generating human faces.
The concern is that if you generate a human face
that looks like someone,
and you don't have rights to that person's image,
is there a legal liability?
So, you heard Adam Selipsky talking
about the indemnification?
That's something that we're kind of talking through,
but I think that that's the way that things are gonna go.
Dogs are a little bit (chuckles) less
of a third rail, right?
A little bit less sensitive.
So, I wanna blow through a bunch of examples
on how you can use this stuff, like right now.
All of these were done without any coding whatsoever.
Most of these demos were done within 20 minutes.
I'll show you how to do that with the drag and drop tools,
but I'll give you some examples.
In this example, we've got a product,
so we take the image, and the details from the image.
In this case, we don't have a good product description,
we just have these kind of bullet points.
Instead of generating a product description for everyone,
we generate it based on a Nielsen segment, a PRIZM segment.
So, we know where these rising singles,
we know where they live, we know where they shop,
what they drink,
and so, we're actually asking the model
to create a product description given that product
for that target demographic.
We can also do it in French, 10 other languages,
so, I had a customer with 300,000 items
in their product catalog.
It was a mess, there were special characters, HTML codes,
and imagine being able to clean that up
really quickly with this model.
Same thing with SEO keywords.
One of my customers last week said that all of their people
that were copywriters are now copy editors.
The models are generating all of the copy for products.
So, we're seeing some role changes,
and some change management that needs to happen,
but there's a lot you can do around the products.
Pricing analysis.
This is kind of meant to be generic,
but a lot of times, you have a data scientist
in your organization that's doing something
like pricing analysis,
but really getting the business to understand it
is difficult.
So, in this case, what we're doing is we have
a Python notebook,
and we're looking at what are the most important things
about our product that affect the price and sell through.
It could be reviews, the quality of the product,
the total price, competitor pricing,
and we've got all that data,
and so here, what I've done is I've said, look,
explain this analysis to me,
and I did another one where I said,
explain it to me like I'm a 10-year-old,
but so, what I see is planograms, pricing,
assortment, a lot of these complex things
that were really hard for people to understand
at the business level.
You can now have the data in your data lake,
you can have this really sophisticated analysis,
but a business user can talk to that analysis
in really simple plain English,
or 10 other languages, right?
Here's another fun one.
I had a customer that said that it takes them six weeks,
they make athletic wear, yoga pants and things,
they said it takes them six week, or six quarters
to get a product to market, right?
So, it's the concepts, the naming,
the colors, all the CAD models,
and so, this is an example where, basically,
we could give it a product description,
and we could generate in seconds,
we can generate hundreds of different products,
and the idea is that we would probably change this
before we go to market, right?
But the ability to generate hundreds of product ideas,
generate the CAD models, the descriptions,
is really super powerful, right?
I love that one.
So, this is another good one.
In a manufacturing environment, a lot of my customers
tell me, when a machine goes down, it costs them money,
but the machines are really complex,
and they don't have people on staff that know
how to operate 'em and fix them.
So what they end up doing is they end up going to a room
where they pull the manual, they flip through it,
and they're trying to figure that out,
so, a really complex set of information
about how to repair equipment,
and so, what I did was I gave the model the PDF
of the manuals of the equipment,
you can see this is really complex stuff,
and then we can start to ask questions of the equipment,
so, if the line goes down,
and the bottle filler's giving us trouble,
we can go in, and someone on the shop floor
can ask questions like, how do I bypass the bottle detector?
That comes back, and what the model's doing
is it's going into the manual, reading through things,
and then it's gonna answer.
You can ask something like, help me troubleshoot this.
It can actually look across multiple documents
and database to formulate that answer.
And of course, you can do it in 10 languages,
so if you have people that are Spanish-speaking,
French-speaking, it's really powerful
for them to be able to get access.
So, I'm gonna show you how to build this.
This was built in about 15 to 20 minutes.
I'll show you exactly how we do it.
There's something called, well, let me back up.
There's something called LangChain,
which is the thing, the framework that helps you
glue all these systems together.
So, the model doesn't know anything about your business,
it doesn't know anything about the equipment,
so what LangChain does, it helps you,
it helps the model talk to your database,
to your knowledge base, to the reviews,
and your product catalog,
and so, LangChain stitches it all together,
but you have to know Python, it's complex.
So, these tools, this one's called Flowise,
allow you to drag and drop components,
and to build that demo that I just showed you,
this was all built in about 15 minutes
using no coding tools.
There's another one called Langflow, if you like Python,
I'm a Node.js developer.
I love Node.js, right?
So here, we gave it the document, the PDF file,
I'm using HuggingFace embeddings, the mini LLM.
I gave it a Bedrock component running in US East.
My favorite models are the Anthropic Claude models.
Sorry, Titan, but I love these models.
They're really fast, very, very, very sophisticated,
and then at the end, we give it a conversational QA chain,
and that's all we need to build that flow,
and you could take that flow,
and you could put it in a mobile device,
you could put it in a webpage,
or on a tablet in the factory floor,
and the idea is, back to that list of use cases,
if you've got thousands of use cases in your business,
you gotta figure out how to deliver these really quickly.
If every one of these use cases is a Python developer
checking into the the CICD platform, running through QA,
it's really gonna be hard to get the value out of it, right?
So, product catalog is super interesting.
Right now, when you go into an e-commerce site,
you're looking for specific items, right?
So, if you're going hiking, you may look for boots,
a jacket, a hat.
The platform doesn't know what your intent is.
I think the future of shopping is to really have
a conversation with a product catalog to say,
hey, I'm going hiking in North Georgia Mountains
this weekend.
The model can know the temperature, it can know the weather,
it can make product recommendations based on that,
and I think that's where things are gonna go,
and Instacart announced something like this, we did as well.
You're gonna tell the system what you're looking for,
and you'll have a conversation with it.
In this use case, basically, I did the same thing
I did with manufacturing.
You give it the product catalog,
and you can ask it questions.
Can you recommend a face wash good for sensitive skin?
All my health and beauty customers,
they're all about the regimen, right?
So, you want a moisturizer, a face wash,
a makeup remover, they should work well together,
and so, you can ask it in multiple languages,
and I think that this is the way that retail is gonna go,
and you can ask follow up questions, it's really easy to do.
One of my customers, they have 100,000 products
in their catalog,
and we're building something with them
around the sales enablement,
so when you go into their retail store,
their seller, they just started,
they don't know 100,000 products in the catalog.
They probably don't know 100,
and so, the ability for the seller
to also use this app to be able to help the customer
be thoughtful about product recommendations
is super powerful.
We were talking about call center earlier,
and every single industry has this,
but basically, when you get a phone call to a call center,
a lot of 'em are QA'd after the fact.
5% of calls typically are QA'd.
It's expensive, it takes a long time.
Agents are are tasked with summarizing the call afterwards,
it takes two to three minutes.
They don't do a good job a lot of times.
So we have Amazon Connect, which is our call center.
There's all kinds of crazy cool use cases
you can light up right away.
The basic architecture is you have a Connect instance,
we transcribe the call to text,
then we start running it through the models, right?
What could the agent have done better?
We did no training on this, by the way,
everyone wants to talk about training models.
None of this was using training.
Create a call summary of this call with 100 words or less.
Do it in French, right?
Super powerful, this is using the Claude Instant model,
which is kind of comparable to ChatGPT 3.5 Turbo.
Super fast, low-cost model.
I use the Claude 2 when I need something more powerful.
I like this, too.
What was the result of the call?
Or how about, did the agent commit to anything?
Because if the agent said, "Look,
"we'll follow up in two days,"
then you wanna maybe open a Salesforce ticket
to make sure that they follow up, right?
So, I love some of the code stuff.
I had a mainframe at Coca-Cola, we migrated,
and we couldn't find the COBOL developers that wrote it,
so, it's really cool to be able to ask the system,
given this COBOL, or any code, tell me what this code does.
All right?
And then once you know what it does,
you could say, hey, how about converting that to Python?
So, all these old legacy systems you have,
your ability to now migrate that to a new platform
is really super powerful.
And I was doing the same thing with cloud formation,
this is the last demo I'll show,
but it used to be it took three months
to get an application up and running in your data center.
Amazon made it faster, we took it down to like, maybe,
you know, a few days to get the architecture.
Now, with cloud formation, you can actually have Claude
or the models generate your cloud formation template.
You could do the same thing with CodeWhisperer,
but here, I say, create an Amazon Connect instance
in the US East, and give me an example flow.
So, it's building the cloud formation template in real time.
We can then run that and get all the infrastructure
stood up within minutes,
so you got a call center in just a few minutes.
Super powerful.
I would QA this stuff, you know,
assume it's like a really brilliant, hardworking intern.
They're really smart, you wanna give 'em direction,
but when the stuff comes back, you wanna double-check it.
You wanna sanity check it, right?
So, never assume the model is perfect.
Most of my customers are using human-in-the-loop use cases,
so, the model will generate something,
and there's a human that's looking at it and validating it
before you put it out into production.
I would still recommend that at this point.
It's not the technology that's gonna hold you back,
it's the change management aligning with the business,
focused on the business outcomes.
A lot of times, I think the best thing you can do
is work with the business,
figure out what their goals are for the year,
and focus on this kinda like management,
you know, the McKinsey, Bain,
what are the things that we could do with GenAI
that are high impact, that are low effort?
Start here, don't train the models.
Use stuff that's out of the box.
Start with call center, human-in-the-loop.
I see a lot of companies that are stuck in POC land,
they're building hundreds of POCs, nothing in production.
Start with a small set.
Get your responsible use framework up and running,
and if you don't have one, you can use ours.
So, this is our responsible use framework.
You can find it online.
So, most of my customers are taking what we have
and they're adding to it, right?
Rather than starting from scratch,
and that talks about how do you get
this thing into production?
How do you monitor it?
How do you think about your team diversity,
risk, bias, harm, and all the legal implications?
So, again, I'm Michael Connor,
and please connect with me in LinkedIn.
I'll be here after too, I'd love to to follow up,
have some deep discussions with you,
but thank you for being here at re:Invent.
It's good to see you all. (audience clapping)
5.0 / 5 (0 votes)