Integration of AI in Salesforce products

From: redpointai

Salesforce, an incumbent with extensive data, has rapidly advanced its AI integration within its products, as discussed by Clara Shih, CEO of Salesforce AI on the “Unsupervised Learning” podcast [00:00:06]. The company aims to make AI an integral part of how businesses operate, predicting that in 10 years, current work methods will seem “completely different” due to AI’s influence [00:01:22].

Key AI Products and Features

Salesforce has developed several generative AI applications under the Einstein GPT brand [00:01:54]:

• Service GPT for customer service [00:01:59]
• Sales GPT for Sales Cloud users [00:02:02]
• Commerce GPT for e-commerce [00:25:43]

Recently, Salesforce launched its Einstein Co-pilot and Co-pilot Studio [00:05:19].

• Einstein Co-pilot is a natural language conversational assistant grounded in customer data, metadata, and Salesforce flows [00:05:22].
• Co-pilot Studio allows customers to customize their own co-pilot [00:05:37]. It includes a prompt builder for creating and saving “golden prompts” referencing specific Salesforce data fields [00:05:42]. It also features a model builder, enabling customers to fine-tune or integrate their own predictive models into the Salesforce AI stack [00:05:49].

AI’s Impact on Customer Support

A standout application has been service reply recommendations and case summaries within Service GPT, which significantly reduce time-consuming tasks for customer service personnel [00:02:10]. This system grounds queries in customer data within Salesforce, making it feel “magical” for users [00:02:28].

Gucci Case Study

Gucci, an early adopter of Salesforce AI, deployed Service GPT to its customer service teams in Florence, Italy, and New Jersey [00:03:04]. The AI augmented service representatives in real-time, allowing them to:

• Reduce average handle time for customer issues [00:03:15].
• Uplevel to become salespeople and brand storytellers [00:03:21]. For example, if a customer called about a belt issue, the AI could recommend new products like handbags they had browsed, coaching the representative in real-time [00:03:59].
• Increase conversion rates by double digits [00:04:18].
• Transform the cost center of customer service into a revenue center [00:04:24]. This shift empowers customer service representatives, making them feel they are doing “the best work of their careers” [00:04:43].

Organizational Structure for AI Development

Salesforce’s AI team structure has continuously evolved [00:06:20]. Initially, AI folks were largely decentralized within each application cloud [00:06:22]. However, recognizing the need for shared infrastructure (like the Einstein Trust Layer and model gateway) across all apps, Salesforce created a new shared services AI platform team [00:06:48]. This central team builds foundational AI components (Trust Layer, Gateway, Prompt Builder, Einstein Co-pilot) [00:06:58], while individual product teams (e.g., Sales Cloud, Service Cloud, Slack) focus on predictive AI for their specific use cases and build actions on the central platform [00:07:10].

Trust and Guardrails in AI Deployment

Given its large enterprise client base, Salesforce prioritizes trust, security, and ethical considerations for AI deployment [00:07:28].

1. Technology: The Einstein Trust Layer is engineered into the product, providing features such as:
   • Data masking [00:07:58]
   • Data grounding with Data Cloud to reduce hallucinations [00:08:00]
   • Citations [00:08:03]
   • Audit trails [00:08:04]
   • Prompt defense [00:08:05]
   • Zero retention prompts [00:08:06] These features mitigate data security, privacy, and ethical risks [00:08:09]. Salesforce also proactively surfaces sensitive data fields and recommends masking them to prevent bias in AI models [00:09:42].
2. Acceptable Use Policy: For AI Bots, the AI is required to self-identify as an AI [00:08:23].
3. Stakeholder Engagement: Salesforce has developed open-source trusted AI guiding principles (accuracy, honesty, empowerment), shared with the industry and government regulators [00:08:39].

Data Strategy

Salesforce leverages four unique types of data crucial for AI [00:16:31]:

1. Structured CRM Data: The traditional Salesforce heritage of data records [00:16:39].
2. Unstructured Data: Including knowledge articles, conversation transcripts from Slack, contact center voice calls, and emails [00:16:49]. Salesforce is expanding its Data Cloud for vector search and hybrid reranking across both structured and unstructured data [00:17:09]. They also have zero ETL partnerships with major data lake providers (e.g., BigQuery, Redshift, Snowflake, Databricks) [00:17:18].
3. Metadata Layer: Created 25 years ago for multi-tenancy, this layer is crucial for providing context to AI, helping it identify relevant data objects, tables, and functions [00:17:35].
4. Feedback Data: As the world’s largest database of customer outcomes (e.g., sales opportunity stages, marketing campaign conversions), this data serves as a reward function for any AI model, both predictive and generative [00:18:01].

Model Strategy and Custom Model Integration

Salesforce adopts an open architecture approach to models, allowing customers to choose models on their service, or bring their own first-party or third-party models via the model builder [00:19:12]. This acknowledges that different models will be optimal for different tasks and use cases [00:19:28].

Salesforce’s own research models are fine-tuned for industry-specific and domain-specific use cases, such as code generation, Salesforce flow generation, financial services sales, and high-tech sales [00:19:36]. Customers can select models from providers like Google, OpenAI, or Azure OpenAI [00:19:52]. Salesforce also benchmarks models for cost, performance, and latency per task, creating specific benchmarks for industries like Pharma or Wealth Management [00:20:16].

Barriers to Enterprise AI Adoption

While cost is a consideration, especially for scaling, it is not the primary barrier [00:23:48]. The biggest barriers for enterprises to scale AI adoption are:

1. Trust: Concerns around data security, data privacy risks, and honoring sharing rules and entitlements within the organization [00:24:12].
2. Business Case: Clearly defining how AI will drive productivity, margin expansion, and a positive ROI [00:24:48].

Salesforce addresses these by providing “turn-key AI use cases” with pre-built prompts, allowing customers to quickly see business value [00:25:27]. Customers can then customize these prompts with their brand voice and custom data fields (e.g., Gucci’s “style” field vs. Ford’s “car model” field) [00:25:50].

Future of AI in Business Operations

Salesforce envisions AI transforming every department and requiring everyone in a company to “rewrite their job description” [00:32:01]. They believe the “team + AI” paradigm will be the future, where humans and AI collaborate seamlessly [00:34:26].

Internal AI Use

Salesforce uses generative AI internally in various ways [00:32:54]:

• V2MOM Planning: Brainstorming and summarizing Vision, Values, Methods, Obstacles, and Measures within departments and roles [00:32:32].
• Product Launch Planning: Using AI as a co-pilot in Slack for brainstorming and collaboration [00:33:07].

AI in Slack

Salesforce has launched Slack AI, which includes conversation and channel summaries [00:34:47]. They are working to bring Einstein Co-pilot into Slack, enabling features like service reply recommendations, summarizing customer data, and sales opportunities before meetings [00:34:58]. This allows entire account teams to access comprehensive customer information, augmenting collaborative workflows like group coding or sales account closing with AI insights [00:34:00].

Workflow Orchestration

AI’s utility extends beyond question answering to workflow orchestration [00:36:27]. For example, a co-pilot can be empowered to initiate returns or send shipping labels. However, this is done with a layer of governance: admins must explicitly designate which existing flows the co-pilot has access to, starting with “least harmful” actions [00:36:57]. This permissioning ensures that the AI’s actions align with organizational rules and entitlements [00:37:39].

Incumbents vs. Startups

Salesforce believes there is ample space for innovation and value creation in the AI ecosystem, both for incumbents and startups [00:28:29]. While incumbents like Salesforce have an advantage due to existing data and platform logic, they also invest in startups through Salesforce Ventures [00:28:51]. They encourage startups to build within the Salesforce ecosystem and on its platform, leveraging its data cloud and UI [00:29:08].

For more information, visit salesforce.com/Einstein [00:40:46].