Generative AI - Driving Resilient Supply Chains

Summary

本次网络研讨会由C3 AI公司主办，介绍了其供应链套件和创新应用——C3生成式AI。C3 AI的供应链套件是一系列针对供应链规划和执行的企业AI应用，涵盖从采购优化到需求预测、生产调度优化、库存优化和供应链风险管理等用例。特别引人注目的是C3供应链数字孪生技术，它整合了来自不同ERP系统和其他企业数据源的供应链数据，提供全面的供应链视图。C3生成式AI作为一个AI驱动的知识助手，允许用户通过自然语言查询访问企业信息，提高了决策效率并减少了数据隐私风险。演示部分展示了C3 AI如何帮助用户通过生成式AI快速获取供应链信息，并采取行动以缓解潜在问题。

Takeaways

• 🚂 数字化转型是企业生存的根本问题，关乎是否能够适应时代发展。
• 🛠️ C3 AI 提供了一整套端到端的企业AI应用，涵盖供应链规划和执行的各个方面。
• 🌐 C3 AI 的供应链数字孪生技术整合了ERP系统、企业数据源和外部数据，为供应链管理提供统一的数字视图。
• 🔮 C3 AI 利用生成式AI技术，帮助企业快速访问和分析大量企业信息，提高决策效率。
• 🛡️ C3 AI 的生成式AI架构通过知识嵌入和检索模型保护原始数据，避免了标准大型语言模型（LLM）的数据安全风险。
• 📊 C3 AI 应用提供可预测的软件套件，帮助企业在全球范围内实现细粒度的可见性，并应用不同类型的AI模型。
• 👥 C3 AI 为供应链中不同角色的用户提供完整的工作流程，包括需求规划师、库存经理、供应商关系经理和采购经理。
• 📈 C3 AI 供应链应用的客户包括航空航天、国防、石油和天然气、零售、消费品、制造业和医疗保健等行业的大型组织。
• 🔄 C3 AI 的生成式AI产品允许用户通过自然语言查询访问跨各种数据库和信息源的数据，提高用户效率。
• 📚 C3 AI 通过案例研究展示了其供应链解决方案如何帮助农业生产商优化供应链，提高准时交货率和服务水平。
• 🔑 C3 AI 通过提供访问控制，确保数据隐私，并防止敏感信息泄露给未经授权的用户或竞争对手。

Q & A

什么是C3 AI供应链套件的核心组成部分？

C3 AI供应链套件的核心组成部分是C3供应链数字孪生(Digital Twin)，它提供了对所有供应链操作的细粒度和时间历史记录，整合了来自不同ERP系统、企业数据源和外部数据源的数据。

C3 AI如何利用人工智能和自动化来改变世界？

C3 AI通过使用包括AI在内的技术，使人们能够创造价值并为客户提供结果，从而利用科学和自动化的潜力来改变整个世界。

C3 AI供应链套件如何帮助企业提高效率？

C3 AI供应链套件通过提供端到端的企业AI应用程序，涵盖从采购优化到需求预测、生产计划优化、库存优化以及供应链风险识别等多个方面，帮助企业提高效率。

C3 AI的供应链数字孪生如何处理不同ERP系统的数据？

C3 AI的供应链数字孪生能够整合不同的ERP系统，无论是Oracle、SAP还是其他系统，将其统一到一个共同的数字视图中，为企业提供完整的供应链数据视图。

C3 AI如何确保数据的安全性和访问控制？

C3 AI通过将大型语言模型(LLM)与数据检索模型分离，确保LLM无法直接访问原始数据，从而保护数据安全。此外，C3 AI还实施了企业级访问控制，确保用户只能访问其被授权的数据。

C3 AI的生成性AI与传统的LLM有何不同？

C3 AI的生成性AI通过知识嵌入和检索模型来提供答案，而不是直接让LLM访问数据。这种方法提供了可追溯性、确定性的答案，并且没有信息泄露的风险。

C3 AI生成性AI如何帮助用户更有效地获取信息？

C3 AI生成性AI作为一个AI驱动的知识助手，允许用户通过聊天或搜索形式提问，系统会将问题转换为查询，并从检索模型中获取数据，以人类可读的格式提供答案。

C3 AI生成性AI如何处理数据缺失的情况？

如果查询的数据缺失，C3 AI的检索模型将返回没有数据的查询结果，并通过LLM告知用户没有数据或无法访问相关信息。

C3 AI生成性AI如何与现有的供应链解决方案集成？

C3 AI生成性AI可以集成到现有的供应链解决方案中，通过查询和分析现有系统中的数据，为用户提供更深入的洞察和决策支持。

C3 AI的实施周期通常需要多长时间？

C3 AI的实施通常从3到6个月的试点项目开始，在这段时间内，将完成数据集成、模型训练和用户工作流程配置，确保在项目结束时拥有一个生产就绪的应用。

Outlines

00:00

🌐 数字化转型与企业生存

本段讨论了数字化转型对于企业生存的重要性。强调了利用新技术，包括人工智能，来创造和提供价值的必要性。提到了数据作为一种核心产品的价值，并强调了理解如何利用数据的重要性。同时指出这是一个大规模且极具挑战性的努力，涉及与世界上一些最大的公司合作，目标是使世界变得更好。

05:00

🤖 C3 AI 供应链套件与创新

介绍了C3 AI供应链套件，这是一系列针对供应链规划和执行的企业AI应用。套件涵盖了从采购优化到需求预测、生产计划优化、库存优化以及供应链风险识别的多种用例。强调了C3供应链数字孪生的重要性，它整合了来自不同ERP系统和其他企业数据源以及外部数据源的数据。此外，还提到了C3 AI的差异化特点，包括其预测性、全球可见性、AI模型的灵活性、用户工作流程的完整性以及可扩展性。

10:01

🛠️ C3 生成式AI与企业风险管理

讨论了C3生成式AI如何为供应链提供创新解决方案。与标准的语言模型不同，C3 AI通过知识嵌入和检索模型提供安全的数据访问，确保了答案的确定性、可追溯性，并符合企业数据隐私控制。解释了C3 AI如何通过整合非结构化和结构化数据，提供比传统语言模型更安全、更可控的AI能力。

15:01

📈 提高供应链效率与案例研究

本段强调了C3生成式AI如何帮助提高供应链效率，通过提供一个统一的真相来源，使供应链团队能够快速访问所有相关信息。讨论了C3 AI如何集成到现有的ERP和规划软件中，并强调了访问控制的重要性。通过一个农业生产商的案例研究，展示了C3 AI如何在六个月的试点项目中帮助统一数据源，使分析师能够提出基于文本的问题，并快速获得测试结果和产品数据。

20:02

🔍 C3 AI 供应链网络风险的实时演示

Justin Kendig进行了C3 AI供应链网络风险的实时演示，展示了如何通过提问和应用程序获取详细信息，预测即将出现的问题，并提供解决方案以避免这些问题成为供应链中的实际问题。演示包括了全球供应商性能的查询、特定供应商的详细信息、预测延迟的采购订单行、机器学习模型的解释性分析以及采取行动以减轻问题的建议。

25:04

📊 C3 AI与PowerBI的区别及ROI问题

讨论了C3 AI与PowerBI的不同之处，强调C3 AI是为特定业务角色量身定制的全工作流程应用程序，而不仅仅是增强版的PowerBI。还讨论了C3 AI生成式AI如何通过提供自然语言接口，使用户更容易访问所需信息。此外，还回答了关于C3 AI投资回报率（ROI）的问题，指出C3 AI可以显著提高预测准确性、降低库存成本、减少采购成本，并提高订单履行率。

30:06

🗓️ C3 AI的实施时间和数据外泄防护

解释了C3 AI的实施通常从3到6个月的试点项目开始，在这期间将整合所需的数据，并配置搜索界面。还讨论了C3 AI如何防止数据外泄，即通过防止大型语言模型（LLM）访问企业数据来实现。最后，讨论了如何处理供应链数据中的空白，包括处理不确定性和使用相似性技术来填补数据缺口。

Keywords

💡数字转型

数字转型是指企业或组织通过采用数字技术和解决方案来改变其运营方式的过程。在视频中，数字转型与企业生存的基本问题相关联，强调了利用人工智能等技术手段来创造和提供价值的重要性。

💡人工智能（AI）

人工智能是指由人造系统所表现出来的智能，它能够执行通常需要人类智能才能完成的任务。视频中提到AI作为推动企业转型和创新的关键技术，展示了其在供应链管理中的应用潜力。

💡供应链优化

供应链优化是指通过改进供应链流程和决策来提高效率和降低成本的过程。视频讨论了C3 AI供应链套件如何涵盖从采购决策到库存优化等多个用例，以实现供应链的端到端优化。

💡数字孪生

数字孪生是物理对象、系统或过程的虚拟表示，用于模拟、分析和控制。在视频中，C3供应链数字孪生提供了对所有供应链操作的粒度和时间历史视图，整合了来自不同ERP系统和外部数据源的数据。

💡生成式AI

生成式AI是一种人工智能技术，它能够创建新的、以前未见过的数据实例，如文本、图像或音乐。视频特别介绍了C3生成式AI如何为供应链提供创新解决方案，通过AI驱动的知识助手提高企业效率。

💡预测分析

预测分析是一种使用统计算法分析数据并预测未来结果的技术。视频中提到C3 AI应用程序提供预测和粒度洞察，帮助企业提前识别风险并做出最优决策。

💡数据集成

数据集成是将来自不同源的数据合并到一个统一的环境中的过程。视频强调了C3 AI如何整合ERP系统和其他企业数据源，以及如何通过外部数据源提供完整的供应链可见性。

💡访问控制

访问控制是限制对资源或数据的访问的机制，以确保只有授权用户才能访问。视频讨论了C3 AI如何应用企业级访问控制，以保护数据隐私并防止信息泄露。

💡机器学习模型

机器学习模型是使用机器学习算法从数据中学习并做出预测或决策的模型。视频演示了C3 AI如何使用机器学习模型来预测供应链中的延迟和风险。

💡自然语言处理

自然语言处理是AI的一个分支，它关注于使计算机能够理解、解释和生成人类语言。视频中C3生成式AI使用自然语言处理技术，允许用户通过聊天界面或搜索形式提出问题，并得到答案。

Highlights

数字化转型是企业生存的基本问题，关乎是否能够跟上时代的步伐。

C3 AI提供端到端的企业AI应用家族，涵盖供应链规划和执行的多种用例。

C3 AI供应链数字孪生提供了所有供应链操作的细粒度和时间历史记录。

C3 AI供应链数字孪生整合了不同的ERP系统，提供了统一的数字视图。

C3 AI供应链应用的关键是其预测性，能够预测风险和提供最优决策参数。

C3 AI的供应链应用提供全球可见性，支持多种AI模型，具有高度灵活性。

C3 AI的供应链应用通过工作流程为不同用户提供定制化的体验。

C3 AI的供应链应用支持大规模AI，与全球大型组织合作，每日改善业务成果。

C3 AI的生成性AI产品允许用户通过聊天界面快速访问企业信息。

C3 AI的生成性AI与标准语言模型不同，提供了数据保护和访问控制。

C3 AI的生成性AI架构通过知识嵌入和检索模型提供安全性。

C3 AI的生成性AI确保了回答的确定性、可追溯性，并防止了数据泄露。

C3 AI的生成性AI通过自然语言界面使用户能够更有效地提出问题并获取信息。

C3 AI的生成性AI可以集成到现有的供应链解决方案中，提供额外的价值。

C3 AI提供3到6个月的试点项目，快速启动并集成数据。

C3 AI的生成性AI能够处理数据缺失问题，提供不确定性处理和相似性方法。

C3 AI的供应链应用通过机器学习模型提供预测和解释能力，帮助用户理解并采取行动。

C3 AI的供应链应用提供双向集成和APIs，实现应用程序之间的实时数据同步。

Transcripts

[Music]

what this digital transformation mandate

is about is a fundamental issue of

corporate survival whether you want to

be on the train or whether you want to

be on the

tracks we're trying to think about New

Visions new models new methods and how

we use technology including AI to enable

our people to to create value and to

deliver results for our customers and

for the company the potential for the

ways that that sort of science and

automation can change the entire world

is really mind-bending the options are

Limitless you need to understand that

the data you have available is a core

product that you have it is value and in

order to realize that value you need to

understand how to leverage that this is

a very large scale effort it's um

extraordinarily challenging we're

working with some of the largest

companies in the world and we're making

the world a better

[Music]

place hello everyone welcome to the C3

AI webinar series my name is Kon I'm a

senior associate outbound product

manager at C3 aai and I'll be your

moderator today I'm excited to be joined

by Layla Fridley and Justin kendick lla

Fridley will introduce the C3 AI supply

chain suite and C3 generative AI for

supply chain and Justin Kendig will help

demonstrate C3 generative AI as part of

C3 AI supply network risk before we get

started I do want to point out that we

have a Q&A window if you have any

questions during the presentation please

feel free to drop those in our team will

either help answer them through chat or

at the live Q&A at the end thank you

again for joining us with that I will

hand it over to

lla thank you Kon I'm thrilled today to

get to discuss with you um our C3 AI

supply chain Suite as well as some

exciting new innovation of applying

generative AI for supply chain

software so I'll start by introducing

our supply chain Suite which is an

endtoend family of Enterprise AI

applications for supply chain planning

and execution our applications cover use

cases from sourcing optimization to

improve procurement decisions demand

forecasting with granular and accurate

demand forecasts production schedule

optimization which drives uh efficiency

in manufacturing production C3 AI

inventory optimization to improve

inventory service levels and reduce

inventory holding costs and C3 AI supply

network risk which identifies um risks

in advance of ordered delays from

suppliers and going out to customers and

I'm really excited that we have Justin

joining us later today to show a demo of

c3i supply chain supply network risk

with C3 generative

AI one thing that is really important

and found foundational to all of our

supply chain applications is the C3

supply chain digital twin the supply

chain digital twin provides a granular

and time history of all of the supply

chain operations um and all of the data

flowing in from various erps from other

Enterprise data sources and from

external data feeds the C3 supply chain

digital twin integrates disparate Erp

systems whether that's Oracle sap or

other systems and unifies them into into

a common digital view of all of your

data across your supply chain this means

that if you're using different erps

across the world across different

business units or if you've invested in

different systems over time our supply

chain digital twin is unifying all of

that information all of that

transactional data and making it a Time

series time history of all of the

different movements across your supply

chain that includes the history of

demand forecasts of order movements

sales histories inventory

um supplier relationships all of the

traditional transactional data is

represented in the supply chain digital

twin on top of that we unify other

Enterprise data sources so that could be

coming from things like your CRM from

your Marketing Systems so you're

building out the context of what's going

on in your business not just looking at

the transactional records and finally as

any supply chain professional knows you

don't operate your supply chain in a

vacuum the world outside is really

affecting the business decisions that

you're making and so our supply chain

digital twin comes pre-piped with

external data feeds that track things

like weather um as well as macroeconomic

data like market indices supplier

Financial Risk ratings vessel movements

and Port traffic so all of this

contextual information is providing the

complete visibility you need to know in

order to manage your supply chain and

make predictive uh predictive decisions

not just reactive

decisions some of the key

differentiators of our supply chain

digital twin and our supply chain suite

are again as I mentioned it is innately

a predictive software suite we no longer

have to manage Supply chains based on

what happened last week or what are

emerging issues today but instead our

applications are providing predictive

and granular insights about what risks

are coming and what are like what is

likely to change what are the optimal

decision parameters in order to account

for that

uncertainty we provide Global visibility

at a very granular granular level with

part level tracking across the entire

supply chain it's very flexible to be

able to apply different types of AI

models I'll talk in a second about how

we apply generative Ai and the

generative models but we also apply

various forms of forecasting models

unsupervised learning supervised

learning very flexible to any kind of

AI every single application provides a

complete workflow for that for the

different types of users across the

supply chain that could be uh demand

planners inventory managers in uh

supplier relationship managers

procurement managers all of the

different people that are involved in

supply chain decisions have uh very

specialized and tailored workflows

within our supply chain applications and

finally we have scalable AI we work with

some of the largest organizations across

the globe whether that's in Aerospace

and defense oil and gas retail and cpg

Manufacturing and Healthcare um our

customers are deploying the C3 AI supply

chain Suite in order to improve their

business outcomes on a daily

basis now I mentioned I'm really excited

to share some exciting new innovation

around how we apply generative AI for

the supply chain our generative AI

product was launched last year um based

on customer feedback and the request to

say how do I have Google for my

Enterprise how can I quickly access all

of that Enterprise information that you

are unifying for my applications that in

that supply chain digital twin across

all of the various data bases and

information sources that we integrate

into C3 make the users more efficient

and give them access to the information

they need in order to make Better

Business decisions so this is the AI

powered knowledge assistant that we are

applying across all of our different

customer deployments and we are

embedding into our uh various

applications

so I'm going to talk a little bit about

what makes the C3 generative AI

capability different than standard

llms as everyone on the planet is aware

of today llms and generative AI are

taking Enterprises by storm and like

many of you I too use generative models

uh to make myself more efficient but

these generative models have some

downsides and some risks the standard

architecture is that a user either um a

a business user or someone who is using

one of these models for personal use

interacts with the llm in a chat-like

interface they could posst a question um

they could ask for you know a travel

itinerary when that llm is coming up

with the response it's actually

searching and querying and is trained on

all of the available data to that llm in

most cases the llm actually has access

to text documents uh to HTML which is

websites and in some cases to code we've

seen examples of that where um employees

are starting to use these llms to help

them write boiler plate code but when

they do that it exposes these companies

to a fair amount of risk first is that

the responses that the llm is generating

are random um the if I ask the same

question twice I'm going to get two

different answers and many of you have

probably experienced this in your

personal use second is there's no

traceability the llm is trained on the

data but it doesn't have a way to

identify what data is actually being

used to generate its response and so

that means that as a user I have no idea

where this answer is coming from do I

trust the data source do I know what the

ground truth is and in the standard

architecture it does

not third is there's no Enterprise

access controls I may be limited in the

type of data I have access to as a

business uner business user I may sit

right now I'm sitting in California I

really shouldn't have access to data

based out of um Europe for example but

the standard llm architecture has no way

to protect the data privacy to make sure

that the right users have access to the

right information and no information

leakage is happening speaking of

information leakage this is a major

concern for our Enterprise customers

where uh they do not want to have their

private data exposed to competitors to

the broader internet again this is

something we're seeing clinging out in

the news right now now where some users

have already um copied and pasted

private IP into these llms in order to

make themselves more efficient but now

it's exposing their organizations to

massive IP risk because their uh I you

know their Source IP their code um their

privacy is being exposed to the entire

internet and finally these models are

prone to hallucination if I ask a

question that the llm doesn't know the

answer to the standard AR architecture

does not provide any controls on that

model to prevent it from making up an

answer what it you know it thinks is the

the next best word in a sentence is

going to predict that is the right

answer to my question there's no way for

it to know whether or not it actually

has the right answer because it doesn't

have the ability to trace down to ground

truth the C3 generative AI architecture

is really different what we do is we

protect the llm from the raw data so I'm

going to walk through this architecture

from the top we incorporate the

unstructured data files like you would

with a standard llm so text HTML and

code but we also incorporate structured

data now this is unique because standard

llms are only looking at text files

we're incorporating the structured data

that we integrate into for example the

supply chain digital twin into the

application object model like tables

application data sensor data log files

transaction data all of that is

represented in the supply chain digital

twin but the llm does not have access to

that data itself instead we apply um

knowledge embedding and the llm only has

access to the retrieval model so when a

user asks a question either in chat form

or in search form say how much inventory

do I have at my Distribution Center the

llm is interpreting that question and

converting it into a query that it

passes through to the retrieval model

and the retrieval model is what has

access to the data itself in that way we

are providing a layer of security to

protect against the risks of the

standard

architecture so what that means is our

answers are deterministic because they

are based on the data the llm is only

interpreting the answer and returning

the result in a human readable format in

a chat-like interface U but the actual

response is coming from that Vector

store coming from the retrieval model

and so it is determinist if you asked

how much inventory do I have at my

Distribution Center if you phrase that

question differently um or if two

different users ask it on two different

days as long as they have access to that

information through the access controls

and the data protection then they'll get

the same

response we have full traceability

because we know exactly where the

response is coming from through the

retrieval model so that means that you

know exactly what the ground truth is

and as a user I can go follow up and

make sure that I agree that this is the

right answer answer um in as you'll see

later in the demo from Justin we have a

Google like interface so the user can

interface with the chat model but then

we surface the ground truth results and

they can click into the underlying

either uh Enterprise data source the

planning software or the C3 AI

application where that information is

coming

from we apply the full Enterprise access

controls I'll walk through this at a

high level in a second but basically

this means that given our 10 years plus

in Enterprise AI applications and

handling unified aggregated data at

scale we have a lot of experience with

Enterprises that need to protect their

data even within their organization so

if one user sitting in California has

different data access than a different

user sitting in London we respect those

data privacy

controls um we do not suffer from llm

caused leakage of proprietary

information because again the llm does

not have access to the data itself

furthermore we Host this either in a C3

hosted Cloud environment or a customer

hosted cloud cloud environment so it's

not exposed to the rest of the internet

but even still the llm itself is not

accessing the data and therefore it's

not learning it's not um being able to

take those insights and then surface

them to a different

user and finally there's no

hallucination again because the llm is

simply playing the role of understanding

the user's question and providing the

response in a human readable format it's

not actually coming up with the answer

itself the retrieval model is finding

the answer and in that way if there's no

response if a user asks a question you

know who's the uh King of England well

your Enterprise data doesn't have the

answer to that question and so in that

way there's there's no response and so

the uh retrieval model will not be able

to surface any data to that question and

the llm will simply respond there's no

response to this

question

so how is this making people more

efficient you know that's really the

name of the game with generative Ai and

the use cases that people are applying

it to in the current supply chain

process a supply chain analyst is faced

a question um you know why is my

supplier contract is the supplier

contract different why why is this um

order coming from a different port than

I expected it to you know all all kinds

of questions they might be asking

they're going to have to go and look

across various disperate data sources

they might look up that transaction in

Erp they're going to go find the

supplier contract they might compare to

other supplier contracts they might have

to look up the bill of materials to

understand what this impacts Downstream

and then ultimately they also are going

to consult with people who are

experienced seasoned professionals that

they work with if they don't have as

much experience with these types of

issues as a result it takes them hours

days weeks to come up with the answer to

various types of questions and to to

keep things on track with C3 generative

AI for supply chain however supply chain

teams can interact with the C3

generative AI application which is

providing them access to all of their

supply chain data all of their

information in one place and they no

longer have to pull together that

information for themselves but rather

have a single source of truth that they

can then go and start their

workflows Justin will show a demo of

this in action in a second uh but again

this helps supply chain teams ask the

most relevant questions that allows them

to provide very seamless access across

all of their Enterprise applications

that are integrated this does not just

apply to the C3 supply chain Suite

applications like C3 AI inventory

optimization which it absolutely is

integrated with but it also pulls that

information from again that whole supply

chain digital twin which is integrating

data from existing erps from existing

planning software regardless of who the

provider is so if you're using sap and

Oracle and blue yonder and canais and

all of the different planning tools this

application the C3 generative AI

application can pull and query data from

across those Source

systems I mentioned that access controls

are really important for our users so

I'm just going to walk through this

quick example of how this works in

action

as a user so let's say I'm a supply

chain analyst I submit a query again how

much inventory do I have at my

distribution center now the way I phrase

that maybe I you know I'm using

different language I'm not going to ask

the same query every time people just

don't right humans interact in in very

colloquial text so the large language

model converts that into a query that

gets pass to the retrieval model but if

I ask that question and I only have

access to some of the data sources the

response I get will be different so in

this example as a supply chain analyst I

might only have access to document one

the application data two application

data three uh but I don't have access to

document two maybe this is a private

legal contract that or a commercial

contract that as a supply chain analyst

I'm not on the commercial team I don't

have access to that and I don't have

application data Maybe I don't have

access to CRM data and so I actually

don't know what were the different

transactions that we had with Downstream

customers so as a result the retrieval

model will return an answer but it's

only based on the data that I have

access to and furthermore the evidence

it provides the ground truth it provides

will tell me exactly where that data is

coming

from so in that way we are making sure

that data privacy and access control

limits are respected across not only C3

generative AI but also all C3 AI

applications where we apply Access

Control

limits so finally I'll end on a case

study where we have a C3 generative AI

for supply chain deployment with one of

our supply chain customers so this is a

large agricultural producer who's using

the C3 supply chain site and our

applications to make their supply chain

more efficient in order to optimize and

predict what's coming and reduce the

variability in order to improve their

ontime info rates and service levels to

their customers

so the problem they faced is that they

had a variety of data sources that again

supply chain analysts were having a hard

time um cating and understanding

synthesizing to make good decisions so

over the course of a six-month pilot

which is how we typically get started

with a new customer in a new use case we

unified data from what what's called

here unclassified data sources so this

is data that the entire Enterprise had

access to um any one could access

publicly like news articles public

statements from the company fact sheets

about their products and their

services and then um we also unified the

their quotequote classified data sources

so this is internal privately protected

IP and data where um the access controls

are really important so that's things

like test result documents test

descriptions and the Ron process test

data so as this agricultural producer is

testing their final product making sure

that it is meaning the quality um and

you know the yield expectations pulling

in that unified data into the C3

generative AI for supply chain

application and as a result these users

are able to ask text based questions ask

for references query things like you

know how many of my batches met my

expected Target performance um am I able

to ship on time uh for this order

ID um and the application returns charts

and graphs to summarize the test results

and provides ranked list of additional

search results so they can again go into

for example the lab information system

Limbs and kind of investigate in more

detail or go straight to the fact sheets

and product data sheets to understand

what are the requirements that matter to

this uh particular product and

test so with that I'm really excited to

turn over to Justin Kendig my colleague

to provide a demo of C3 generative Ai

and C3 supply chain products in

action awesome thank you lla all right

let's switch over to the live demo and

see some software based on what lla just

described all right so as you heard lla

describe this there's a lot of questions

and we as humans naturally just ask a

lot of questions and what I have found

in the supply chain profession is you

fall into two camps you're either the

leader the manager the VP who is asking

questions about your supply chain why

did this happen what should we do next

you know what's going on over here

what's our best practice where am I

failing the most or you're in the other

Camp of you're the one that's being

asked the question and you have to

figure out how am I going to answer this

and there's frustration on both sides

the leaders are saying how come I can't

get this question answered faster and on

the other side it's why are you asking

so many questions these are very

detailed so the ability to be able to

have that Google interface as I'm about

to show you and be able to say hey I

don't know what my questions are going

to be I just want to be able to start my

day figure out here's a situation I get

an email whatever it is I see something

I want to be able to ask a question and

then it's not just the first question

that you want to be able to ask it's

then those follow-up questions and the

drill downs and the next question and so

that's what I'm about to show you so

when we look at our generative Ai and

our Enterprise search applied to our

applications you start to see so let me

navigate you you've got a Google like

interface and so start my day and I want

to say you know what is my current

Global supplier

performance and so very quickly it comes

back and it says hey your Global on time

INF full performance over the last two

months is 82% now let's also see what it

is I can see there's documents there's

news articles there's images videos I

can put context into this I can also see

other search results along with the

initial one about you know supplier

details performance evaluations again

this is tapping into your Enterprise

data to be able to pull back what's the

relevant information for that along with

them what people also ask so what are

other their variations as lla was

describing of that same question well so

great now I know what my Global

performance is let's say I'm the North

American manager I want to know what my

performance is of my suppliers coming in

to North America so I can say which

suppliers are worst performing in North

America so I type that up answer it and

then I quickly get a list of here here

are the top 10 worst performing

suppliers so what's their name what's

their on time and full of the suppliers

delivering into my network where are

they located what's the supplier City

but then I also in the search results

see all right the top supplier is

Bearing Distributors this one has the

lowest on time in full I can also see

that there was a supplier performance

evaluation and so that evaluation may be

in PDF I can then click on it the search

results came back they found and I can

go through and see okay here was this

company it was EV when it was evaluated

what their quality was who the

evaluators were I can get that

information very quickly to now see all

right what do I know about this

particular supplier or I can drill in

and go to information Pages or other

information that's available but now I

really want to know okay what do I need

to worry about today with Bearing

Distributors so which purchase orders

are at risk for Bearing Distributors so

I open this up and it tells me there are

three active purchase order lines where

is their destination that requested

delivery date the predicted delivery

date again this is just going in to the

data that's pulled into our supply chain

digital twin and saying all right these

three are now predicted to be delayed so

we have a model a machine learning model

within our supply network RIS that is

predicting the day that a purchase order

line is placed until when it's actually

going to arrive and every single day

it's updating and so two of these are

predicted to be delayed and one of

them's already delayed but now I can

also go see all right what are those

purchase orders so I can open this up

maybe this is opening up your Oracle

instance your sap maybe you have these

in PDF forms you can now start to see

okay this is a YW motor here's you know

here's who I am that's ordering it where

I'm ordering this from so again just

pulling all that information along with

you know other item details that you

want to order or you can jump into that

particular order line and get more

information so now I'm navigating into

our supply network risk application

where it starts to show this particular

order line for an item a yaw motor it's

coming from this supplier going to my

facility in Joliet okay but before I

jump into here that was me asking

questions about my performance and

ultimately getting to an individual

perch sorder line now I could also jump

back and go in the more traditional

route from our C3 aai applications we

start off with overview Pages where you

can start to see configurable key

performance indicators such as how many

open purchase order lines do I have what

percentage of my order lines are delayed

predicted delayed um predicted to be on

time what's my supplier My overall

supplier performance then you can start

to see what's my open purchase order

lines have I been improving am I getting

worse um but ultimately you can then

start to prioritize you know if I want

to move the needle on my supplier

performance which supplier should I work

on or maybe there are certain items that

I should be focusing atten on that are

causing my performance issues or maybe

there's a specific facility again

prioritize your work to figure out how

am I going to improve my supplier on

time and full performance the other

available option is that say you're a

planner a a material planner and inbound

planner maybe you're in charge of a

specific facility or a group of products

maybe it's a set of suppliers you can

filter down and then see an alert driven

workflow allows you to see what are the

purchases store lines that are predicted

to be delayed so now you can come in

each day and start to see these things

um pop up on your list start to go work

on them prioritize by value how many

days they're predicted to be late along

with an other contextual information and

then you can drill into those purchase

order lines so now I've taken it from

asking questions and getting into a

specific purchase order line detail or I

can go in from a prioritize standpoint

and those two different workflows merge

though into the same location of what do

I need to do about this particular

purchase order line so now that I'm

drilled into this order line I can see

it's predicted to be delayed by 10 days

I can see general information um I can

also see activity feed so this is all

the information that was fed into the

machine learning model from various

Milestones so the day it was order was

created when it was accepted when the

carrier picked it up from the supplier

when it arrived at the Port of origin

when and then now it's on on on a vessel

and so it's been departed but then

sometime after the vessel Departed the

AI model predicted that it was going to

start to be delayed so now I want to

dive in and understand what is the

machine learning model telling me so I

go into my AI explainability to go look

at the evidence package I'm going to

turn off some of this to start to see

and explain what's going on so I'm in

this purchase order line and over time

because every single day our machine

learning models are running for all of

your open purchase order lines I can see

the expected lead time was going to be

45 days and when originally it was

created the machine learning models were

predicting that this particular order

was going to arrive early and then

eventually it got a little bit more

delayed a little bit more delayed and

then about a week ago it started to

predict oh you're going to start being

late and so now we're at the point where

we're 10 days late what I want to

understand though is in the hundreds of

features that are going into this model

how do I categorize them and that's what

these color components start to bring in

is we have grouped features together

into these different areas called

quantity Network time allocation demand

and those help you to start to

understand where are the problems within

your network at least for this

particular order that are causing the

delay another feature I can look at as I

scroll down is I can go pull up any past

model so I said the machine learning

model is running every single day well I

can go back in time to understand what

were the features saying what were the

contributions and so as I look at the

feature contribution I can see that this

network category is causing the biggest

amount of contribution to my delayed

what does Network mean so as I read

through this through our Insight cards

you can start to see that the lead times

between milestones for similar orders

has increased by 22% over the last six

months so over this time maybe you

didn't see it in like a single event but

it started to happen over time you

started to see these increase well what

are all the features that make that up

so now these are the individual features

that make up the machine learning model

port-to-port lead time terminal delivery

supplier promise delivery changes I can

see what a description is of all those

features what the value is and then what

the contribution and so the whole goal

of this explainability or our what we

call our evidence package is to open up

that black box that is a machine

learning model or an AI model and start

to explain this in layman's term

so that everyone can start to understand

and you can start to learn all right

Port toport lead times something's

happening with my vessels that's

probably going to impact more of your

network than just this one purchase

order line so now I understand a lot

more about what's going on with this

model why it's predicting what it is so

now let's go dive in and see what can I

do about it are there actions so I've

predicted that it's going to be delayed

ahead of time now I want to start to

take action and see um what some of my

options are so I'm going to switch over

to a more global view to understand you

know physical premise where is this um

where is this container where is this

particular order line um at today so I

can quickly see here was the supplier

over in China the port of origin the

container because we're tracking as lla

said all of the AIS um data so we can

track where vessels are around the world

so I can see where that container is at

I can see the port of destination and

then I can see the Final Destination um

over here in Joliet if I look at this

particular container I can start to get

information about here's the port here's

the destination but I see that this

container so my Transportation team has

this container as a low priority what I

can do though is we can work with all of

our customers and we can configure

various recommendations one of them

might be you may have a different option

for Port processing priority so maybe I

want to increase the priority of this

container and then I can see because the

predictive model has learned that if you

change this from low to high I can shave

two days off of my processing time so

I'm going to go ahead and do that so I'm

going to change this up to high I'm

going to accept that and what that did

is not only did it change it in our

application because of the bidirectional

integration and the ability for all of

our apis to write we now push this out

to your transportation team they now

have awareness and so you're making this

more efficien for your supply chain

teams to work together I want to look at

some other options for what can I do

about this so now I'm going to zoom in

on my um actual facility where it's

going I can see you know where's the

location it's over in Juliet I can see

there's a stockout risk on this yaw

motor that's part of this purchase order

line and I can see that there are a

couple options there are containers that

have been sitting around for a while um

in one in Houston and then one in um

LaGuardia and so I can start to look at

those I can see one would reduce it by

six days my stock out time one by 10

days so what I'm going to do is I'm

going to pull this one from LaGuardia

I'm going to basically accept it here in

the application that will write out to

your system and what it'll do is it'll

start to reroute that container and say

all right send this one to Juliet

because I need to avoid a stock out risk

and so whether it's containers with

inventory whether it's other facilities

with inventory we're pulling in those

Network views and we can configure these

recommendations so you can start to take

action so let's quick go back and

summarize what we saw was we started

asking questions and then through the

applications I started to get more and

more detailed information about what are

my problems and I started to see

problems that were about to happen and

then the C3 AI Suite of applications

then provided you with recommendations

for how to mitigate those action

mitigate those problems before they

become an actual problem within your

supply chain so I will turn this now

over to Kon and we will start our

Q&A great thank thank you Justin and lla

for the great presentation and demo for

the audience please continue to use the

Q&A window for questions that you want

to ask we'll start off with a few

questions U the first one is for Layla

what's the difference between C3 Ai and

powerbi is C3 aai just an enhanced

version of

powerbi yeah great question um as I

think became evident in Justin's demo

the C3 AI supply chain applications are

not um just an enhanced version of

powerbi these are fully workflow enabled

applications that address primary

business needs um and are specifically

tailored to specific business personas

so the demo that Justin just shared is

really focused on an order manager to

keep orders on track and make sure that

there are not um ins suing delays for

things like manufacturing or for

customer delivery we have similar

application workflows and I urge all

attendees to watch our explainer videos

or follow up with us for additional

demos for our inventory um optimization

application demand forecasting sourcing

optimization and production schedule

optimization what generative AI does is

it provides users with a new interface

to ask questions and get easy access to

the information that they need so if

there is um information that is already

integrated to the C3 supply chain

digital twin or to their um into their

application um instance then they can

start to ask more natural language

questions that traditionally would

require additional reporting or pulling

data in order to to run a query so

instead it's providing a more natural

language interface to allow you to ask

more free form questions but ultimately

neither are in my mind at all similar to

powerbi um or business reporting

dashboards really they are number one

providing endtoend user driven workflows

and number two providing a more natural

language interface to access

information great thank you lla Justin

we have a question for you uh the

question asks I've already made

significant investments in advanced

planning and scheduling Supply Chain

Solutions can C3 aai generate enough Roi

to justify investing in another

tool yeah that's a great question so you

know I actually think the need for

investing into more process driven

Solutions in advanced planning is

absolutely critical but where those

applications start to fall short is the

ability to make decision operational

decisions using Advanced AI ML and

optimization and what we have seen with

many of our customers is when they go in

they do those Investments and then we

come in and sit in parallel and help to

enhance those Solutions is we can see

demand forecast increase of 10 to 20% in

accuracy Improvement we can see

reductions of inventory around 35 to 50%

we can see reductions in sourcing costs

of upwards of 10% and we can ultimately

as you're focused on your customers see

on time and full order fulfillment

increases of around 20% so that's on top

of the Investments that you've already

made those are coming in because the

artific intelligence because the machine

learning models have been specifically

applied along with then some of the um

alert K based um workflows that we sit

on top of your Advanced planning

solutions they can start to enhance and

provide definitely provide that Roi

needed um to make it worth it great

thanks Justin lla we have a question for

you we just saw Justin provide a demo on

J with SNR or Supply C3 AI supply

network risk lla will other C3 AI supply

chain applications have C3 generative AI

capabilities yeah um absolutely we are

incorporating C3 generative AI into all

of our C3 AI applications um we also

sell C3 generative AI as a standalone

product um we will configure the search

interface and ability to chat with the

llm and query the underlying data the C3

generative AI product be deployed on its

own um or it can be um codedeploy and is

part of all of our other C3 AI

applications so if you choose to pursue

C3 AI inventory optimization or C3 AI

supply network risk those applications

come out of the box with C3 generative

AI um the difference is really what are

the data that are Incorporated and are

you applying the predictive models that

are inherent to those other applications

as well as those end user workflows so

it really depends on the business

problem that you need to

solve thanks lla Justin we have a

question for you can C3 generative AI

populate results with different graphs

based on the prompt that you

ask yeah no that's a great question um I

will I go back to a little bit of the

first time I was introduced to

generative AI as we our teams were

starting to show it to us and we learned

that if you ask a question about where

something happens it pulls up a map

because the models are smart enough to

know that if you want to ask where it's

going to bring you a map and so I've

seen in you know we're going to get more

mature in this everyone's getting more

mature I've seen you know different bar

charts line charts I've seen Maps I've

seen sanky diagrams can be generated I

obviously the maps I talked about so

yeah it's it's really cool the power

that can be generated and the way that

information can be shared and what's

really cool is that the application will

start to present you with that data in

the best way that it should be presented

and so and it'll continuously learn over

time and get better at how should it

present the results to specific

questions that are answered and so

really excited to see where this is

going to go um as we continue to enhance

it and make it even better awesome

thanks Justin uh lla question for you

how long does implementation

take yeah we usually get started with

customers with three to sixth month

pilot projects over the course of that 3

to six month Pilot We Will integrate the

data required for the use case now again

if it's just a a generative AI use case

then we are integrating unstructured

data some structured data and

configuring that search interface if we

are deploying a C3 aai supply chain

application that also involves training

the predictive models the optimization

models configuring the end user workflow

such that the end of the 3 to six month

period you have a live production

application with users that have access

to start get getting started you know

immediately um so so it just depends on

which use case is selected to decide if

it's a three or six month

project great Justin a question for you

can C3 generative AI be applied to my

existing Supply Chain

Solutions yeah absolutely and by tapping

it into those databases as L is showing

you can start to pull those results in

and start to query whether it's your

planning systems whether it's your um

powerbi systems your Tableau click you

know whatever your um reporting systems

are and you can take whatever you've

invested in today start to bring that

data in and start to query questions

against it and it'll present it back to

you the way you want to so absolutely

you can get started and then it gets

enhanced even more as you start to bring

in some of the C3 AI supply chain Suite

of applications as

well great thanks Justin lla the next

question is for you uh it relates to

data so how does C3 generative AI

prevent data exfiltration can you speak

a little bit about what happens when

there are gaps in supply chain

data yeah so uh two questions one on

datax filtration which is where I will

start um so I as I presented um in the

slides we actually prevent the llm from

accessing our customer data so the llm

is a trained deep learning model that

knows how to interpret and synthesize

you know human language um but it

doesn't know anything about your data

and so what that llm is doing is it's

simply understanding what is the intent

of the question what information are you

looking for and passes that to the

retrieval model because the llm doesn't

ever have access to the Enterprise data

it never learns from it it doesn't

become um expert about your you know

your operations your supply chain it

doesn't actually know anything about

your suppliers your inventory your

distribution your manufacturing

facilities so there's no information it

has that it could be exposed to

providing out into the world um so that

is a really important part of our our

architecture that prevents all of those

downsides of standard llm architecture

um the second question you asked is how

does it overcome data gaps um I'll

answer in two ways so first is you know

if there's a gap in the data and we're

just talking about the search based

interface you know I ask a question I'm

looking for a response if that data

doesn't exist again the retrieval model

will return no data in the query and

then it presents an basically an empty

array and the llm will say okay my

response is I don't have data I don't

know how to answer that question and

we've you know we've tested this

countless times um if you ask a question

the that C3 generative AI doesn't have

an answer to the L will respond you know

I don't know we don't have access to

that information so that's one part of

the answer the second is well what if I

really need that data to make better

decisions you know I need to know if um

how much inventory I have at my

different different distribution centers

what if that data is missing um and all

of our applications come with pre-built

pipelines that address different types

of problems and data gaps is a really

common problem we run into all the time

um we address that by handling

uncertainties so we know that and if we

see an outlier where maybe there is a

missing data entry point where it looks

like inventory is zero but all of the

previous data um was you know inventory

of 10,000 units we might treat that as

an outlier and then smooth over it um

there's other types of missing data

handling Techni techniques that we use

um we also apply similarity for things

like demand forecasting so you know if

you have no data it's new product

introduction if you're missing data data

from sales we can handle that with

similarity to other products there's a

number of different approaches that we

take depending on the type of missing

data that we're talking about when

applied to specific predictive problem

or specific type of insight that we're

generating um so again it depends on

whether or not you're asking how do you

handle missing data whether it's you

know search-based I'm just trying to

query and get access to that data and

it's missing we'll say we don't have

access to the data but if it's about a

predictive model and a supply chain

workflow then yes we do have ways to um

overcome those data gaps so that you can

still get really good Insight from your

data that you do

have great well thank you Laya and

Justin that's all the time that we have

today as a quick reminder this webinar

will be available on demand and you will

receive an email with the link later

today if you enjoyed today's content

please follow c3ai on LinkedIn Twitter

and YouTube and thank you again for

joining us and we hope you have a great

rest of your

day