Salesforce AI Associate

[su3inni]

1. AI Fundamentals

17%

Machine Learning

• learns and adapts based on patterns in data to make accurate predictions
  • Key characteristic of machine learning
  • uses algorithm to learn from data
  • and make decisions
  • without explicitly programmed
  • analyze data identify pattern and make predictions/recommendations

    Generative AI

  • Creates new content based on existing data prompted by a human

    Natural Language Processing : NLP

  • understand and generate human languages
  • interaction between computer and human language
  • speech recognition, machine translation, sentiment analysis, text summarize, chatbot

Algorithm

• a simple set of rules for turning an input an output

  Neural Network

• a web of connections, guided by weight and biases

Key benefit of effective Interaction between humans and ai systems benefits decision making

• leads to more informed and balanced decision making
• enhance cognitive abilities and performance

[su3inni]

2. AI Capabilities in CRM

8%

Improve sales processes using AI

• Lead Scoring and opportunity forecasting

  • Part of Einstein
  • Portfolio of AI features
  • Lead Scoring은 과거 데이터 분석 후 고객 전환 가능성의 기회를 점수화하여 예측한다.

• representatives better understand previous customer interactions

  • Using Call summaries
    • brief report , analyzing audio or text
    • sales representative review custormer's satisfaction

• decrease workload by implementing chatbot

  • chatbot using NLP
    • understanding, generating language
    • create, conversational interface
    • interaction with users

• Commonly Used three

  1. Predictive scoring : 우선순위 정하기
     • lead or opportunity based likelihood of conversion or closing
     • help prioritize
  2. Forecasting : 수요 예측
     • mitigate risk
     • forecasting future sales revenue or demand
  3. Recommendations

Salesforce Guideline 종류

• Honesty Guideline
  • ⭐Honesty Guideline⭐when Training AI models
  • ensure appropriate consent , transparency about AI-generated responses

    인공지능이 만든 답변을 사용할 때 적절한 동의와 투명성을 보장해야한다.

  • 데이터 출처가 중요하며, 오픈소스 또는 사용자 제공 데이터 동의가 필요하다.
  • AI가 생성했음을 사용자에게 공지해야한다.
• Sustainability Guideline
  • minimize model's carbon footprint
  • reduce energy consumption , greenhouse gas emission
• Safety Guideline
  • control bias, toxicity, harmful content with embedded guardrails and guidelines
    • create guardrails to mitigate toxicity and protect PII guardrails 는 AI 시스템에 의해 유해하거나 부적절한 컨텐츠 생성을 방지한다. ex. filter, threshold, warnings, human reviews
  • mitigate/prevent potential harms(unfairness, discrimination, manipulation, protect rights)

Einstein = CRM AI Assistant

https://trailhead.salesforce.com/ko/content/learn/modules/get_smart_einstein_feat

• Einstein Bots
• Einstein Predict Builder
• Einstein Next Best Action
• Einstein Discovery
• Einstein GPT
• Einstein Engagement

Einstein Features

• boost sales efficiency and effectiveness
• features
  • Opportunity Scoring
  • Lead Scoring : 데이터를 분석하고 가능성에 영향을 미친 요소를 식별하는 과정
    • 점수와 이유를 부여해서 고객이 될 가능성과 이유를 확인
    • lead에 우선순위를 구하고 유망한 부분에 집중할 수 있다.
  • Account Insights
  • Opportunity forecasting
  • case classification
    • 속성을 기준 사례를 분류하는 과정으로
    • NLP 로 사용자 데이터를 분석해서 유형,이유,우선순위,제품 등 필드에 올바른 값을 자동으로 할당할 수 있다.

Einstein Generate

• generative AI capability for CRM
• NLP for analyze
• composing emails, writing knowledge articles, generating code

Einstein Bots

• AI-powered chatbots
• provided self-serve application
  • customer resolve their issues quicker
  • reduce workload
  • improve satisfaction
• using NLP, NLU
• can leverage other Einstein features ( case classification, recommendations, provide personalized, relevant solution )

Einstein Engagement

마케팅 관리자가 마케팅 캠페인 및 채널을 최적화하는 방법에 대한 통찰력과 권장사항 제공 best solution for better engage

• understanding customer's behavior
• providing insights & recommendations on how to optimize

Einstein Discovery

• Requires CRM Anaytics Plus license or Einstien Predictions license ( 추가 비용 필요 )
• 코드가 필요없는 빠른 반복 환경에서 통계적 모델링과 지도형 머신러닝을 통해 비즈니스 인텔리전스를 강화한다.
  1. 개선할 비즈니스 결과 정하기 ( 회귀분석, 이진분류, 다중클래스 분류 )
  2. 분석을 위한 데이터 랭글링
  3. 데이터를 기반으로 예측모델 생성
  4. 데이터에 대한 인사이트 제공
  5. 예측 및 개선 사항 생성
  6. 데이터 알림으로 통찰력 및 모델 개선
• 사용 방향
  • 편향성을 분석할 민감 변수 선택하기
  • 불평등 효과 해석하기
  • 프록시 변수 살펴보기
  • 모델 카드로 예측 정확도 증대하기

[su3inni]

3. Ethical considerations of AI

39%

• Example of ethical debt : 윤리적 부채
• launching an AI feature after discovering a harmful bias

Ethical Challenges

• potential for human bias in ml-algorithms
  • affect ai development process, data collection, data labeling, model selection
  • train with not presentative diverse customer
• transparency in decision-making process

Bias leads Ethical Challenges

• unfair or discriminatory treatment of customers
• violation of customer privacy, erosion of trust

Disparate Impact

• Discriminatory practices toward a particular demographic

Bias

• an inaccuracy that is systematically incorrect in the same direction and affects the success of the person or algorithm that contains it
  • 편향 제거를 위해 진행하는 사항
  • pre-mortem 진행
    • a pre-project meeting to set measured and realistic expectations and catch the 'what went wrong' before it happens
    • 편향 제거를 위해
  • 데이터 집합에서 배제되거나 실제보다 비율이 너무 높은 요인 식별
  • 주기적으로 교육 데이터 평가
  • Einstein discovery 사용하여 모델의 편향성을 감지하고 방지하기

⭐Bias Type⭐

Bias는 데이터 수집, 분석, 해석에 영향을 미치는 요인 affect quality (reliability of data)

• Societal Bias : 사회적 편견
  • data being labeled according to stereotypes or prejutices reinforce existing inequalities, discrimination
• Confirmation Bias : 확증 편향 기존 관념에 따른 확증 편향
  • from their purchase history
    • remember informations & existing belief/preference
  • limit diversity , prevent discovering, reduce satisfaction
• Gender Bias

How to minimize Bias

• Implement salesforce's Trusted AI Principles
  • set of guidelines that help salesforce and customer develop and AI system responsible, ethical and beneficial for all
• use diverse & representative datasets to train
• test & monitor for accuracy, reliability, fairness, identify
• mitigate errors/anomalies/bias
• provide clear & transparent explanations & sources
• use AI to augment & empower human
• respect and value the diversity and dignity
• do not exclude, discriminate, harm

Salesforce's Ethical AI Process Maturity Model

AI 개발/배포 관련된 윤리적이고 책임있는 관행/프로세스 평가 및 개선

• technical team is assessing the effectiveness of AI development process
  • use guide the development of trusted AI solution
• Trusted AI > Process Maturity Model

⭐Salesforce Trusted AI Principle > Practices⭐

Guiding ethical & responsible use of AI

• Responsible
• Accountable 본인 행동의 영향을 염두하고 피해나 오류를 피하기 위한 정책 seek and leverage feedback for continuous improvement Taking responsibility for one's actions toward customers, partners and society
• Seeking feedback
• mitigate harm
• engaging with 3rd party
• participate industry
• Transparent & Explainability
• explanation of the prediction's rationale ( user's profile, behavior, preference )
• a model card : how the model was created
• model characteristics & performance ( data source, metrics, limitation )
• understand how AI works , what data is use, what outcome produces
• building trust, accountability
• Empowering & Education
• All Skill level to build AI application with clicks, no code
• Free AI education with Trailhead
• Deliver AI research breakthroughs
• ⭐ Inclusive & Diversity 모든 사람의 사회적 가치를 존중하기 ( 다양성 존중 ), 상황에 따라 차별받는 사람이 없도록 보장하는 것이 목표
• support diverse datasets
• consequence scanning Workshops 제품이나 서비스가 사회에 미치는 영향을 조사하고 피해를 예방하는 과정
• build inclusive teams
• Diverse Dataset
• models are fair, accurate, representative of diverse customer
• prevent & mitigate bias , avoid harmful, discriminatory outcomes

[su3inni]

4. Data for AI

36%

Avoid bias from Demographic Data

• Demograhpic data should be omitted to avoid unintended bias to an AI model
  • 인구 통계 데이터는 성별, 연령, 인종, 민족, 소득, 교육, 직업과 관련된 개인의 정체성이나 속성에 따라 차별을 일으킬 수 있다.
• 각 항목에 따라 데이터 균형을 맞추거나, 마스킹 및 정규화를 진행한다.
• 인구 데이터가 아닌 행동 데이터를 사용해야한다.

⭐Data quality⭐

얼마나 현실을 반영하는지, 얼마나 안전하고 일관성있는지, 데이터 오류와 이상값이 얼마나 없는지

• Key components
• Accuracy , completeness, consistency : fundamental
• Validity, Uniqueness, Timeliness, Fitness
• Key factor influences the ethical use of AI
• Performance, Accuracy, Reliability
• Data quality assessment
• provides a benchmark for AI predictions : compare predictions/real
• identify and address data : performance & reliability
• data quality assessment 항목
• Accuracy
• Completeness
• Consistency
• Timeliness

• Essential data quality dimension

  • Reliability : 정확하고 일관된 정도
  • Trustworthy, credible

• How to Ensure Data Quality

  1. By Field type
     • Using Picklist Field Type
       • enforcing consistency, accuracy, completeness
     • by limiting the possible inputs
     • reducing errors
     • standard data
     • Using validation rule
       • allows to define criteria for data entry
       • prevent saving records that do not meet the criteria
  2. Page Layout
     • 일부 레코드에 사용하지 않는 필드 제거
  3. 중복 관리
     • Reduce communication inefficiency with Data Quality
       • Duplication > confusion, errors, waste

High Quality Data

• Accurate, Complete, Valid, Consistent, Unique, Timely, fit for purpose
• essential for ensuring unbiased&fair AI decisions
• preventing discrimination
• trustworthiness of AI-Driven Decision

Benefits of High-Quality data

1. improve the reliability and credibility of AI-driven decisions, fostering trust
2. improve customer trust & satisfaction better leads, better understanding of customers, better customer relationships
3. enhanced customer value & growth provide more value & opportunities / identify and target their profitable customers ( new & expand good )
4. Increase customer engagement & retention communicate timely & personalized manner

Poor Quality Data

• produce flawed, misleading results
• Harm or injustice to someone
• lead to biased : inadvertently learned and amplified
• produce erroneous result
• reduce accuracy
• limiting generalization , robustness
• increasing complexity, opacity
• decrease efficiency

Ensure Data Quality for AI initiatives

• collect and curate high-quality data from reliable sources
  • data cleansing
  • data validation
  • data integration
  • data monitoring
• High-quality data
  • trustworthy, credible, high-quality from reliable resource
  • affect performance and accuracy
  • completeness, correctness, consistency, timeliness, relevance
• Data quality affects ai models
  • Data's role in training and fine-tuning salesforce AI models
  • Einstein Prediction Builder
  • Einstein Discovery
  • Einstein Vision & Language

Incomplete data quality ( missing info )

• the accuracy of recommendation is hindered
• accuracy , performance, reliability 에 영향을 줌

Security Principle

• Right of Least Privilege > reduce risk of handling sensitive data
  • by limiting how many people have access to data
    • simplifying management, auditing of data access
  • protect sensitive personal data from unauthorized access, misuse, leakage
• Data minimization
  • only necessary & relevant data should be stored
  • decreasing the storage, processing costs, complexity of data
  • enhancing privacy, protection data by limiting scope
• Data retention
  • should be kept as long as needed
  • preventing accumulation, duplication of outdated or obsolete data
  • improving accuracy and quality data
  • respecting rights and preferences of data - who request deletion & modification

Data Backup

key component of data management strategy

• 백업 빈도, 범위, 방법, 위치 선정
• 별도의 위치에 데이터 복사본 생성 및 저장으로 가용성 보장
• Using Cloud service, hard drive, tape
• ensure security, availability in case of any loss, corruption, disaster
• help recover

Essential performance/relevance of AI model

• Volume Volume은 AI 품질 차원이 아니라 모델의 성능이나 관련성에 영향을 주는 요인
  • How to select suitable dataset for training an AI model > crucial factor
• Size of the dataset affect feasibility 타당성 , quality of model
• sufficient information

Protect AI Data Privacy

• User Consent 사용자 동의
  • 어떤 데이터를 수집하는지, 왜 수집하는지, 어떻게 처리하는지, 제 3자 공유 여부
    • AI infringes on privacy when user consent is not obtained

Custom Service Analytics

• data should be fresh and up-to-date
• age : recent data is important to reflects current situation

Words

• Data Leakage ( Hindsight bias )

  • results from using variables in your prediction model that are known only after the act of interest

• Data Insights

  • finding in your data based on thorough analysis generated by Einstein Discovery

데이터 관리 방법

• 이름 지정 : 레코드에 대한 명명 규칙 설정
• 형식 지정 : 날짜와 비용 표현방식을 지정
• 워크플로 : 레코드 생성, 검토, 업데이트 및 보관 프로세스를 결정
• 품질 : 레코드를 측정하거나 점수매기는 기능을 포함하여 데이터 품질에 대한 표준을 설정
  • 기간, 완전성, 사용, 정확성, 일관성 및 중복, 비즈니스에 맞는 다른 품질
• 역할 및 소유권 : 레코드를 소유한 사람, 데이터 변경을 책임지는 사람, 데이터가 변경될 때 알림을 수신하는 사람을 결정
• 보안 및 권한 : 데이터에 대한 적절한 개인 정보 보호 수준을 결정
• 모니터링 : 데이터 품질 관리를 보장하기 위한 프로세스

[su3inni]

• Representaives 담당자
• Debt 부채
• Consent 동의
• Maturity 성숙도
• Demographic 인구 통계
• leverage 영향력
• Accessment 평가
• Reliability , Trustworthy, Credible : 신뢰할 수 있는
• inadvertently 무심코
• amplified 증폭
• rationale 이론적 해석