Transforming the consumer packaged goods industry with generative AI

Michael Connor: Good to see everyone. A quick introduction - I'm Michael Connor and I lead the Generative AI practice for our CPG retail team globally. The 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 where I was Chief Architect and 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 1000 C-suite and board members over the last six months. I've seen a lot in terms of what people are doing, use cases, how they're approaching it, aligning with their business, thinking through privacy and ethics.

I want to share what I've learned in the last six months. The big things:

• Huge predictions in terms of what this technology will do for business - $7 trillion estimated impact to GDP.

• First time I've seen board members and executives really engaged - they see a huge shift in their business coming and want to make sure everyone has a plan, funding, and knows what they're doing. Engaging with board members is new.

You've seen the announcements:

• We invested $4 billion in Anthropic - they make my favorite model, Claude.

• Meta announced celebrities - you can talk to Kendall Jenner 24/7.

• Lots of changes to advertising - that Kate's Island Lager image, affectionately named after Kate Wiley. It's changing advertising permanently.

I'll go through some use cases, do demos, and share what we're seeing across the industry. This is a fundamental change in technology. Friends say it's like NFTs, crypto, metaverse. But we already see customers going into production.

Amazon is all in - every group is applying generative AI. We're using it for reviews, descriptions - all across retail. We think it'll be transformative.

This demo uses Stable Diffusion to generate product concepts - imagery, 2D images, 3D CAD models from text, press releases, descriptions, communications, social media. Across products we're seeing use cases in merchandizing, advertising, photo studios.

I've seen three shifts with customers:

1. First 2 months - what is the technology?

2. Second phase - what are the use cases? I'll cover the roll up.

3. Now - how do we execute, organize, get into production, address ethics and bias?

I'm seeing people move into production. For CPG, top use cases are marketing, sales, operations, R&D, software engineering. Retail is similar - software engineering, ETL, marketing, sales. Customers want revenue growth. Sometimes cost savings in call centers. But revenue growth is the focus.

I worked with a retail COO - in one day we came up with all these use cases. People are trying to connect the tech to the business. We're seeing it affect every aspect of CPG and retail. It goes from 1-2 use cases to a scalable pipeline of thousands across the organization.

I'll show drag and drop tools to move fast, get out of code. But I see things slowing people down - I'll cover that.

Some customers think they'll feed SAP data to AI and magic will happen. Important to explain the difference between generative AI and ML/DL, and why both are important.

Machine learning - one data set, one model, one task. Deep learning - more data to train a model. Generative AI takes all data from internet to train a model that can do many things - text, images, code. Amazing we have a single model that can do thousands of things.

But you need both generative AI and deep learning. The punchline - none of it works without good data. Be good stewards of your data. Customers want to be data-driven but have no insights. I've preached about data lakes, data mesh, hygiene, CDOs. Now with generative AI they're taking it seriously. So it's a way for teams to get board support for data.

I'll show some fun examples. Advertising is about decimal points - if you can make an ad that catches eyes, it really helps. This uses Stable Diffusion on AWS to generate ads. It's different than DALL-E or Midjourney - you have more control with drag and drop tools like Anthropic's Claude.

We're generating an ad - I named a specific park, asked for the skyline, dog name. Imagine generating for every park, every breed. I changed it to a golden doodle. SD generates the image - base pass, refiner pass, every detail perfect. It generates the doodle. AI is bad at product images so we overlay PNGs with product, QR code, messaging. Fun to go to Bedrock - give me all parks, dog breeds, generate thousands of localized ads.

This is how Amazon Ads works - generates backgrounds, messaging for each product and audience. I love this stuff.

Here's an example of all the dogs we created. Brands are squeamish about human faces - concerns over rights to someone's image. We're discussing indemnification. Dogs are less sensitive.

Let me run through more examples of how to use this without coding in 20 minutes:

• Product description based on Nielsen/Prism segment - know their demographics and tailor description. Do it in 10 languages.

• Clean up product catalog with 300,000 messy items - special characters, HTML.

• SEO keywords - customer said copywriters are now copy editors, AI generates all copy. Role changes happening.

• Athletic wear maker said it takes 6 weeks to get product to market - concepts, naming, CAD. We can generate hundreds of products in seconds to spark ideas and accelerate process.

• Manufacturing - when a machine goes down it costs money. Complex manuals. We gave the model all the manuals, can now ask it questions in multiple languages to troubleshoot issues.

Let me show how this was built in 15-20 mins with no code:

• Gave it PDF file, used Hugging Face, Bedrock, Anthropic Claude, conversational QA chain

• Langchain helps model talk to your data - doesn't know your business otherwise

• Tools like Flowwise let you drag and drop to build quickly without coding

• With thousands of use cases, you need to deliver quickly - not every one through CICD and QA

Product catalog - today you search for specifics on ecomm sites. Future is a conversation - "I'm going hiking this weekend" and it makes recommendations based on weather, your needs. Instacart and Amazon announced this. Tell the system what you're looking for.

• Give product catalog, ask questions - face wash for sensitive skin? Get recommendations. Follow ups in multiple languages.

• Seller enablement - 100,000 products. Help sellers make thoughtful recommendations.

Call center - 5% of calls are QA after the fact. Expensive, 2-3 mins to summarize. We transcribe with Amazon Connect, run through models with no training:

• What could the agent have done better?
• Summarize the call in 100 words
• Do it in French

Using Claude model - fast, low cost. Also CLAW. Ask questions:

• What was the result of the call?
• Did the agent commit to anything? Open ticket if they promised follow up.

Fun code stuff - had COBOL mainframe at Coke, migrated, couldn't find developers. Ask system to explain code, convert COBOL to Python. Migrate legacy systems.

Last demo - cloud formation. Took months to get apps running in data centers. AWS made faster - days. Now models can generate cloud formation.

Ask for Connect instance in us-east, example flow. Builds template in real time, spin up call center in minutes.

I'd QA the output, assume it's a brilliant intern - give direction, double check. Never assume perfect. Most customers use human in the loop - model generates, human validates before production.

Not the tech holding you back - it's change management, aligning with business, focusing on outcomes. Figure out business goals for the year - what high impact, low effort AI initiatives support that? Start simple - call center, human in the loop. Avoid POC land.

Get responsible use framework - ours is online, many add to it. Covers getting into production, monitoring, team diversity, risk, bias, harm, legal.

Please connect on LinkedIn. Happy to follow up and discuss more. Thank you!