TrailblazerDX: Innovation, Investment and Integrity
Salesforce recently hosted TrailblazerDX, an annual conference designed for developers, architects, partners, engineers, and entrepreneurs looking to build the new class of generative AI applications and leverage Salesforce’s latest platform features. Below, I’ve shared a recap of the keynote, an overview of how new product features can be used in Life Sciences, Anthropic’s approach to Constitutional AI, and how companies can close the trust gap with the Einstein Trust Layer.
Einsten Copilot: Conversational AI for your CRM
The keynote introduced the public beta release of Einstein Copilot, which uses trusted AI grounded with your own company’s data. Einstein Copilot enables Salesforce customers to generate responses using their own private and trusted data, while maintaining strict data governance and without requiring expensive AI model training. As a result, Einstein Copilot can answer questions, summarize content, create new content, interpret complex conversations, and dynamically automate tasks on behalf of a user, all from a single, consistent user experience embedded directly within Salesforce’s CRM applications.
To enable the use of Einstein Copilot, admins and developers can utilize 3 low-code tools that comprise Einstein 1 Studio: Prompt Builder (Generally Available), Copilot Builder (Beta), and Model Builder (Generally Available).
Einstein Copilot is the UI that serves as the Conversational AI for your CRM across every app.
Einstein 1 Studio are the low-code tools purpose-built to rapidly build AI apps grounded in trusted data.
Einstein Trust Layer are features, processes, and guardrails guiding privacy, security, safety, accuracy, and responsible use of AI.
With Prompt Builder, admins can create, customize, and test prompts to deploy in Einstein Copilot or a Lightning Web Component. Users simply describe in plain language the desire content, such as customer success stories, FAQs, or marketing messages, and Einstein will suggest tones, length, and other attributes to craft impactful prompts.
With Copilot Builder, users have access to out-of-the-box skills, or Standard Actions that Einstein Copilot can leverage to execute a task. Some examples of these actions include drafting or revising sales emails, querying records to find values of certain fields, and summarizing records such as key opportunities or accounts.
In addition, users can also create Custom Actions in Copilot Builder, which are built on top of platform features such as prompt templates, auto-launched flows, or invocable Apex classes. Building these custom actions allows users to have business-specific features available at their fingertips.
With Model Builder, companies can create a predictive model from scratch from data trained in CRM or Data Cloud or bring existing models from Databricks, Amazon SageMaker, or Vertex AI. In addition, you can also use Model Builder to bring your own generative model, leveraging leading available models like Azure OpenAI, Open AI, and Amazon Bedrock.
Bringing Copilot to Life
The first iteration of Copilot is setting the foundation for the Art of the Possible. Summarizing a record or retrieving contact information with a simple click is not going to move the needle. Across various industries, companies will need to determine where Copilot could improve processes and increase lift. Furthermore, companies may benefit with partnering with vendors like Scale AI, which ensures that models are trained on high-quality and relevant data. Scale AI then fine-tunes the AI model to a customer’s specific requirements and rigorously tests the model to ensure it meets the outlined objectives.
At first glance, customers across the pharma lifecycle will have opportunities across Content Generation and Text Analytics to exploit the flexibility of Salesforce’s platform, while leveraging the powerful low-code tools available in Einstein 1 Studio, particularly Copilot Builder. Some *potential* life sciences use cases for how customers can utilize Copilot in conjunction with their CRM and other systems are included below.
Content Generation
For Commercial teams, users can generate robust content, such as building field training documents for pharma reps, personalizing content based on predictive customer behavior, or creating dynamic “talk track” for nurse educators/patient hub members to handle real-time patient queries and objections.
Medical Affairs teams could potentially use synthetic patient data for disease/outcome prediction and treatment optimization, as well as quickly summarizing research articles for MSL notes, particularly before KOL visits.
Supply Chain & Manufacturing teams focused on CAPA identification and documentation can potentially investigate and address problems or non-conformities in products, processes, and systems.
Trial Design teams can produce content that outlines the study design, inclusion/exclusion criteria, and endpoints, as well as summarize various sections for synopsis creation.
Text Analytics
On the Commercial side, companies can examine Call Center Analytics, and analyze and extract valuable insights from customer interactions or create Dynamic Org Charts, identifying SMEs most likely to respond/engage with reps.
In Supply Chain & Manufacturing, Quality teams can summarize non-conformance reports for determining causes for manufacturing deviations.
Finally, Medical Affairs teams can enhance their predictive models and gain access to unstructured insights with less effort, such as utilizing call notes from EMR to assist in predicting diagnoses, and patient support call notes to anticipate patient drop-off.
Constitutional AI and the Einstein Trust Layer
Despite the rapid pace that Salesforce and their AI partners are moving to deliver ground-breaking technology, 2 core themes underpinning this innovation are Trust and Transparency.
During TrailblazerDX, Anthropic Co-Founder Jared Kaplan highlighted that Anthropic abides by Constitutional AI, a method to train models to follow a set of principles. Thus, instead of having humans labeling data to train their AI systems, Claude (its core product) can effectively supervise itself to abide by these principles, thus providing transparency and iterating quickly to try out new constitutions.
In addition, Salesforce highlighted that most companies are struggling to deploy AI reliably and maintain trust for 4 primary reasons: Security, Privacy, Accuracy, and Reliability.
By sending data to an LLM and back into your internal systems, companies are compromising the security and privacy of their data. Furthermore, while LLMs are powerful and are improving each day, they do have unpredictable behavior, which impacts accuracy and reliability, and potentially a brand’s bottom line.
To overcome this trust gap, Salesforce launched the Einstein Trust Layer, a collection of features, process, and guardrails that protect the privacy and security of company and customer data, improve the safety and accuracy of AI results, and promote the responsible use of AI across the Salesforce ecosystem.
The Trust Layer operates in 2 ways: The Prompt Flow (what happens before the prompt hits the LLM) and the Response Flow (what happens when the LLM returns the response to the prompt).
Response Flow: Data Demasking, Toxicity Detection, Audit Trail & Feedback
The Prompt Flow
Before data hits the LLM, there are a set of processes to ensure data security and privacy.
First, Secure Data Retrieval ensures that if a user attempts to generate a response, but doesn’t have access to that data in the Salesforce org, data isn’t pulled in.
Next, Dynamic Grounding allows companies to customize prompts to set their business needs in several ways, including: (1) referencing CRM fields using Merge Fields that are replaced with actual data when the prompt runs, (2) using complex logic in a prompt by pulling in data with Flow, (3) making callouts to grab data from external data sources with API Calls, and (4) using Copilot Search to search for structured and unstructured data that is vectorized, allowing users to use only relevant data that should be used in a prompt to avoid overcrowding.
After grounding the prompt, Data Masking can be accomplished in 2 ways so that the LLM can still understand context about the data.
Pattern-Based Masking uses a combination of Regular Expression matching which replaces sensitive data (Names, Emails, Phone Numbers) with data generated using a defined pattern. An admin also can determine which PII entries can be masked in an Org.
Field-Based Masking is used to mask based on metadata, honoring existing data sensitivity levels, compliance categories, and Shield-encrypted fields. An admin can also configure what information can be masked in an Org, such as PII, HIPAA, GDPR, or PCI.
Furthermore, an admin can also allow authorized users to override data masking at Prompt Template and Model level.
Salesforce also has also a 3-pronged approach to Prompt Defense by limiting context windows to maintain model coherence, blocking malicious prompts, and providing instructions to the LLM to return an appropriate response if it experiences an error or is unsure of the validity of a response.
Finally, Salesforce has applied an additional security layer called Zero-Data Retention with 3rd party LLMs that ensures no data can be used for model training or product improvements, no data can be retained outside of Salesforce, and no human will look at the data sent to the Model. Moreover, companies can audit and verify that no data is being retained.
The Response Flow
Once the prompt has been sent and a response has been received, the response Demasks the sensitive data. Next, the response is evaluated for Toxicity using a specialized model based that is optimized for handling multiple labels, i.e. if a response was labeled as both “hateful” and “offensive”, and is also scored for safety.
Companies can then access Audit Trail in Data Cloud which stores a comprehensive log of all prompts, responses, and feedback to see how well prompts and responses are working, and feedback users have provided. As a final step, Salesforce has a Red Team that is focused on revealing model limitations and training away from harmful outputs.
Future State
Salesforce has invested heavily in ensuring that they are the AI CRM of the future. Copilot has several generic use cases, but it will specifically shine when industries take the lead in exploiting its capabilities. There is a lot of apprehension about using these technologies, but Salesforce is making positive strides to ensure that the existing trust gap closes, albeit not at the same speed as the current pace of innovation.
If you are interested in learning more about how we are helping our clients with their AI strategy, please reach out to me at eric.ryan@zs.com.
