From: redpointai
Salesforce AI, under the Einstein GPT brand, has developed generative AI applications tailored for enterprise use, particularly in customer service and sales. Salesforce is seen as an incumbent that holds significant data and has rapidly adopted AI [00:00:10].
Key AI Applications in Salesforce Products
Salesforce has incorporated AI across its product suite, including:
- Service GPT for customer service [00:01:59].
- Sales GPT for sales cloud users [00:02:02].
Specific features that have shown significant impact include:
- Service reply recommendations and case summaries [00:02:10]. These features automate time-consuming tasks for customer service representatives and reduce waiting times for customers [00:02:15]. This approach allows humans to focus on tasks they excel at, while AI handles repetitive elements [00:02:40].
- Einstein Co-pilot, a natural language conversational assistant grounded in customer data, metadata, and existing Salesforce flows [00:05:21].
- Co-pilot Studio, which enables customization of co-pilots, creation of “golden prompts” that reference specific data fields, and integration of custom predictive models [00:05:37].
AI is increasingly blurring the traditional lines between job roles, for example, transforming customer service representatives into augmented salespeople [00:02:57].
Case Study: Gucci
Gucci, an early adopter of Salesforce AI, implemented Service GPT for its service team, which operates from Florence, Italy, and New Jersey [00:03:04].
- The implementation led to a reduction in average handle time for customer inquiries [00:03:15].
- AI augmented customer service agents in real-time, enabling them to become salespeople and brand storytellers [00:03:24]. For instance, if a customer called about a belt issue, the AI could recommend other products like handbags based on their browsing history or items left in a cart [00:03:59].
- This augmentation resulted in double-digit increases in conversion rates [00:04:20].
- The customer service team, traditionally a cost center, began to operate as a revenue center [00:04:22].
- Customer service representatives felt empowered and reported doing “the best work of their careers,” with their careers taking on a new life [00:04:43].
The rapid deployment of Service GPT (60 days) was possible due to a year and a half of prior research and development, including early collaboration with Gucci starting in November 2021 [00:14:06].
Impact on Workforce and User Education
The introduction of AI tools like Einstein GPT and co-pilots has addressed fears about job replacement by focusing on empowering users and replacing undesirable job tasks [00:11:02]. Salesforce emphasizes that AI acts as a “co-pilot,” not an “autopilot,” serving as a coworker that users can interact with via natural language to request actions or information [00:11:51].
User education is crucial, involving ethnographic research like sitting with call center users to understand their concerns and addressing them through product design or learning modules [00:11:23]. Initial product messages and the “co-pilot” naming convention are designed to clarify the AI’s role as an empowering tool [00:11:44].
Data and Trust in AI Deployment
Salesforce’s unique position is underpinned by four types of data crucial for AI:
- Structured CRM data: Traditional data records like sales forces’ heritage [00:16:38].
- Unstructured data: Includes knowledge articles, conversation transcripts from Slack channels, contact center voice calls, chats, and emails [00:16:49]. Salesforce’s Data Cloud is expanding with vector search and hybrid reranking to work across both structured and unstructured data [00:17:09].
- Metadata layer: Created 25 years ago for multi-tenancy, this layer provides context to AI, guiding which data objects, tables, or functions to use [00:17:35].
- Feedback data: Salesforce captures customer outcomes (e.g., sales opportunity stages, marketing campaign results) as a data pipeline in Data Cloud. This data serves as a reward function for any AI model, whether predictive or generative [00:18:01].
Salesforce emphasizes that it’s the customer’s data, not Salesforce’s product, and adheres to a strict “zero retention prompts” policy [00:16:19].
Guardrails and Trust Layer
Trust is a primary barrier to enterprise AI adoption [00:24:12]. Salesforce addresses this through a multi-layered approach:
- Technology (Einstein Trust Layer): Features include data masking, data grounding with Data Cloud to reduce hallucinations, citations, audit trails, prompt defense, and zero retention prompts [00:07:56]. Data masking, for example, has been used for years to proactively hide sensitive data fields (like name, gender, zip code) to prevent bias in AI models [00:09:23].
- Acceptable Use Policy: Requires AI bots to self-identify as AI when interacting with consumers [00:08:16].
- Stakeholder Engagement: Developed open-source “trusted AI guiding principles” (accuracy, honesty, empowerment) shared with the industry and government regulators [00:08:36].
Model Selection and Customization
Salesforce adopts an open architecture approach, allowing customers to choose models from their service, bring their own, or integrate third-party models via their Model Builder [00:19:12]. Different models are expected to be used for different tasks and use cases over time [00:19:28].
Salesforce’s own research models are fine-tuned for industry and domain-specific use cases, such as code generation, Salesforce flow generation, or financial services sales [00:19:37]. Customers can select models like Google’s Gemini or OpenAI’s Azure OpenAI [00:19:53]. The goal is to track and benchmark models based on cost, performance, and latency for specific tasks [00:20:01].
Salesforce builds industry-specific benchmarks (e.g., for Pharma or wealth management sales) to evaluate model performance, recognizing that domain-specific evaluation is more relevant than general evaluation [00:20:19].
Barriers to Enterprise Adoption
The biggest barriers to scaling AI adoption in enterprises are:
- Trust: Concerns around data security, data privacy, and ensuring AI respects internal sharing rules and entitlements within a company [00:24:12].
- Business Case: Enterprises need to clearly define the business case, focusing on productivity gains and margin expansion to justify investment [00:24:48]. For instance, the Gucci example highlights reduced average handle time and incremental revenue lift [00:25:07].
Salesforce addresses these barriers by providing turnkey AI use cases (e.g., Service GPT) that are easy to set up and demonstrate immediate business value [00:25:27]. They also allow customers to customize AI behavior, such as brand voice and referencing custom data fields unique to their business [00:25:50]. This customization is facilitated by features like the Prompt Builder and Co-pilot Studio [00:26:29].
Future of AI in the Enterprise
Beyond current “killer use cases” like customer support, chat-your-documents, and code generation, AI is expected to permeate every department [00:31:37]. Every job description will likely need to be rewritten, with functional leads and first-line managers guiding employees on how to leverage AI, similar to how workers learned to use the internet and email in the 1990s [00:32:00].
Salesforce uses generative AI internally for V2MOM (Vision, Values, Methods, Obstacles, Measures) planning in Slack, helping brainstorm initiatives and summarize methods [00:32:36]. This highlights the emerging trend of “team plus AI” workflows, such as group chatting with AI or group coding with AI, where AI augments human conversations and decision-making in real-time [00:33:57].
Slack is seen as a perfect interface for generative AI [00:34:30]. Salesforce has launched Slack AI with conversation and channel summaries [00:34:47]. They are also bringing Einstein Co-pilot into Slack, allowing users to get service reply recommendations or summarize all open support tickets, marketing engagement, and events related to a customer before a meeting, visible to the entire account team [00:35:00].
While the percentage of customer support questions answerable by AI varies by industry, AI’s utility extends beyond mere question answering to workflow orchestration (e.g., initiating returns and sending shipping labels) [00:36:16]. Governance and control layers are crucial, allowing administrators to designate which existing workflows their co-pilot has access to, starting with less harmful actions like data lookups before potentially enabling more impactful actions like issuing refunds [00:36:56].