Salesforce AI
Last reviewed
Apr 27, 2026
Sources
18 citations
Review status
Source-backed
Revision
v5 ยท 4,089 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Apr 27, 2026
Sources
18 citations
Review status
Source-backed
Revision
v5 ยท 4,089 words
Add missing citations, update stale details, or suggest a clearer explanation.
Salesforce AI refers to the suite of artificial intelligence products, research initiatives, and platform capabilities developed by Salesforce, the San Francisco-based enterprise software company. Since acquiring the deep learning startup MetaMind in 2016, Salesforce has built one of the most extensive AI research organizations in the enterprise technology sector, producing open-source large language models, multimodal models, protein generation systems, and a full stack of AI-powered CRM tools branded under the Einstein name. The company's AI strategy spans internal research through Salesforce AI Research, product integration through Einstein AI, autonomous agent deployment through Agentforce, and startup investment through Salesforce Ventures.
Salesforce AI Research traces its origins to the April 2016 acquisition of MetaMind, a Palo Alto-based AI startup founded in July 2014 by Richard Socher, a Stanford PhD specializing in machine learning, deep learning, natural language processing, and computer vision. Salesforce acquired MetaMind for approximately $32.8 million, and Socher became the company's Chief Scientist. MetaMind's products were folded into Salesforce Einstein, while MetaMind's standalone services shut down in May 2016. Socher led Salesforce Research until his departure in 2020 to co-found You.com, an AI search engine.
In April 2021, Salesforce hired Silvio Savarese as Executive Vice President and Chief Scientist to lead the research division. Savarese, a professor of computer science at Stanford University, brought expertise in computer vision and robotics. Under his leadership, Salesforce AI Research expanded its focus to include multimodal models, large action models, and generative AI applications for enterprise use cases.
Salesforce AI Research operates from offices in Palo Alto, San Francisco, Seattle, and Bellevue, as well as an international research lab in Singapore. The Singapore team, founded in 2019, produced the highly cited BLIP series of vision-language models.
Salesforce AI Research has produced several influential research outputs spanning language modeling, vision-language understanding, code generation, protein engineering, and autonomous agent systems. The team has published work at top-tier venues including NeurIPS, ICML, ICLR, and Nature Biotechnology, and has released numerous open-source models and libraries.
Notable contributions include the BLIP and BLIP-2 vision-language models, the LAVIS library for language-vision intelligence, the CodeGen family of code generation models, the xGen series of large language models, the ProGen protein generation model, and the xLAM family of large action models.
Salesforce launched Einstein AI in September 2016 at the Dreamforce conference as its integrated intelligence layer for the Salesforce platform. Einstein brought predictive analytics, natural language processing, and machine learning capabilities directly into Salesforce's core CRM products, including Sales Cloud, Service Cloud, Marketing Cloud, and Commerce Cloud.
The initial Einstein release included several key capabilities:
| Feature | Description |
|---|---|
| Lead Scoring | Prioritizes sales leads based on their likelihood of conversion using historical data patterns |
| Opportunity Insights | Provides predictive signals on deal health and likelihood of closing |
| Sales Forecasting | Predicts future revenue to inform strategic planning |
| Next Best Action | Recommends optimal actions for sales and service representatives |
| Einstein Vision | Image recognition capabilities for visual search and product identification |
| Einstein Language | Sentiment analysis and intent classification for text data |
Einstein represented Salesforce's effort to democratize AI for business users who lacked data science expertise. Rather than requiring customers to build their own models, Einstein embedded pre-trained AI capabilities directly into existing CRM workflows.
On March 7, 2023, Salesforce announced Einstein GPT at the TrailblazerDX developer conference, describing it as "the world's first generative AI for CRM." Einstein GPT combined Salesforce's proprietary AI models with OpenAI's ChatGPT technology and real-time data from Salesforce Data Cloud to generate personalized content across sales, service, marketing, commerce, and IT interactions.
Einstein GPT capabilities included:
| Product | Capability |
|---|---|
| Einstein GPT for Sales | Auto-generates sales tasks, composes personalized emails, and creates call summaries |
| Einstein GPT for Service | Generates knowledge articles and auto-generates personalized agent chat replies |
| Einstein GPT for Marketing | Generates personalized content for customer engagement across email, mobile, web, and advertising |
| Einstein GPT for Commerce | Generates personalized product descriptions and commerce-specific recommendations |
| Einstein GPT for Developers | Generates Apex code and provides an AI chat assistant using Salesforce Research's proprietary CodeGen LLM |
Alongside the Einstein GPT announcement, Salesforce and OpenAI revealed the ChatGPT for Slack app, providing AI-powered conversation summaries, research tools, and writing assistance within the Slack workspace. Salesforce Ventures also announced a $250 million Generative AI Fund on the same day.
On June 12, 2023, Salesforce announced AI Cloud, a suite of capabilities for delivering trusted generative AI experiences across all Salesforce applications and workflows. The centerpiece of AI Cloud was the Einstein Trust Layer, a security architecture designed to address enterprise concerns about data privacy when using generative AI.
The Einstein Trust Layer included several protective features:
The Trust Layer was designed to let Salesforce customers use both proprietary Salesforce models and third-party models (from Anthropic, Cohere, Google, OpenAI, and others) while maintaining enterprise-grade security.
At Dreamforce 2023 on September 12, 2023, Salesforce introduced Einstein Copilot, a conversational AI assistant built into the user interface of every Salesforce application. Einstein Copilot allowed users to interact with their CRM data using natural language, grounded in the company's proprietary data through Salesforce Data Cloud.
Einstein Copilot could proactively suggest actions beyond the user's immediate query. For example, after a sales call, it could provide a recommended action plan; in service contexts, it could check order status or change shipping dates without requiring the user to navigate multiple screens.
Salesforce also announced Einstein Copilot Studio, a toolkit comprising three components:
| Component | Function |
|---|---|
| Prompt Builder | Enables administrators to create, customize, and A/B test AI prompts tailored to their brand |
| Skills Builder | Allows companies to control which workflows and data sources the copilot can access |
| Model Builder | Lets companies bring their own AI models (from Anthropic, Cohere, Databricks, Google Cloud, or OpenAI) or use Salesforce's proprietary LLMs |
Einstein Copilot was later rebranded and evolved into Agentforce as Salesforce shifted its AI strategy from copilot-style assistance toward fully autonomous AI agents.
Released in July 2023, xGen-7B is a series of 7-billion-parameter large language models trained with standard dense attention on sequence lengths of up to 8,192 tokens. The models were trained on up to 1.5 trillion tokens of text data. The xGen-7B family includes several variants:
| Model | Context Length | Use Case | License |
|---|---|---|---|
| xGen-7B-4K-Base | 4,096 tokens | General-purpose base model | Apache 2.0 |
| xGen-7B-8K-Base | 8,192 tokens | Long-context base model | Apache 2.0 |
| xGen-7B-8K-Inst | 8,192 tokens | Instruction-following tasks | Non-commercial research |
On standard NLP benchmarks, xGen-7B achieved comparable or better results than other open-source LLMs of similar size at the time of release, including MPT, Falcon, LLaMA, RedPajama, and OpenLLaMA. The long context window of 8K tokens was a differentiating feature, as many 7B-class models at the time supported only 2K or 4K token contexts.
xGen-MM (xGen-MultiModal), also referred to as BLIP-3, is a family of open large multimodal models released in 2024. The models extend the xGen initiative into vision-language understanding, building on the earlier BLIP and BLIP-2 architectures developed by Salesforce AI Research.
The xGen-MM framework improves upon BLIP-2 in three ways: (1) increasing the richness, scale, and diversity of training data, (2) replacing the Q-Former layers with a more scalable vision token sampler, and (3) simplifying the training process by unifying training objectives to a single loss at every stage.
The v1.0 release (May 2024) included base and instruct models built on the Phi-3 Mini backbone. The pretrained foundation model, xgen-mm-phi3-mini-base-r-v1, achieved strong performance under 5 billion parameters and demonstrated robust in-context learning capabilities.
The v1.5 release (August 2024) introduced the xGen-MM-instruct-interleave variant, which improved on both single-image and multi-image benchmarks. A key capability of xGen-MM is handling "interleaved data" combining multiple images and text, enabling complex tasks like answering questions about several images at once.
BLIP (Bootstrapping Language-Image Pre-training) is a vision-language model framework developed by Salesforce AI Research for unified visual understanding and generation tasks. The original BLIP model was released in early 2022.
BLIP-2 was released on January 30, 2023, introducing a novel pre-training strategy that bootstraps vision-language capabilities from frozen pre-trained image encoders and frozen large language models. The architecture consists of three components: a CLIP-like image encoder, a Querying Transformer (Q-Former), and a large language model. Only the lightweight Q-Former is trained, while the image encoder and LLM remain frozen.
BLIP-2 achieved state-of-the-art performance on various vision-language tasks despite having significantly fewer trainable parameters than competing methods. It outperformed Flamingo80B by 8.7% on zero-shot VQAv2 while using 54 times fewer trainable parameters. The BLIP series of papers authored by Junnan Li and collaborators accumulated over 15,000 citations combined.
CodeGen is a family of open-source models for program synthesis developed by Salesforce AI Research. The models were trained on TPU-v4 hardware provided by Google, using data and model parallelism implemented in JAX.
CodeGen 1.0 was released in March 2022 with the accompanying paper presented at ICLR 2023. It included models in four sizes (350M, 2B, 6B, and 16B parameters) and three training stages:
| Variant | Training Data | Description |
|---|---|---|
| CodeGen-NL | The Pile, natural language corpus | Base model trained on English text |
| CodeGen-Multi | BigQuery, multiple programming languages | Initialized from CodeGen-NL, further trained on code in multiple languages |
| CodeGen-Mono | BigPython, Python-only corpus | Initialized from CodeGen-Multi, further trained exclusively on Python |
CodeGen 1.0 was competitive with OpenAI Codex on code generation benchmarks at the time of release.
CodeGen2 was released in May 2023, introducing several improvements including infilling capability and support for a broader range of programming languages (C, C++, C#, Dart, Go, Java, JavaScript, Kotlin, Lua, PHP, Python, Ruby, Rust, Scala, Shell, SQL, Swift, TypeScript, and Vue). CodeGen2 came in 1B, 3.7B, 7B, and 16B parameter sizes and was trained on the permissive subset of the Stack dataset (v1.1).
CodeGen2.5 was released in July 2023 as a 7-billion-parameter model trained on the StarCoderData dataset containing 783 GB of code across 86 programming languages. Despite having only 7B parameters, CodeGen2.5 outperformed earlier CodeGen models more than twice its size, making it practical for production deployment. CodeGen2.5 became the backbone for Salesforce's internal coding assistant (CodeGenie) and the Einstein for Developers product.
xGen-Code is the latest addition to Salesforce's in-house LLM family, specifically designed for developer-centric tasks within the Salesforce ecosystem. Released in 2024 as part of the Agentforce for Developers initiative, xGen-Code powers the Dev Assistant feature and excels at complex, multi-turn interactions and dynamic chat functionality.
While CodeGen2.5 handles quick, on-the-fly code completions, xGen-Code provides deeper, more interactive development support. According to Salesforce, xGen-Code outperforms both open-source and proprietary models in accuracy and efficiency on Salesforce-specific developer use cases. Since the internal launch of CodeGenie (Salesforce's internal developer tool powered by CodeGen and xGen-Code), developers have accepted over 2 million lines of AI-generated code and submitted more than 500,000 chat questions.
xGen-Sales is a proprietary model designed by Salesforce AI Research to handle sales-specific tasks with generative capabilities. The model was created by refining advanced LLMs using human-in-the-loop reinforcement learning, drawing on datasets from diverse sources including product guides, industry best practices, and curated dialogues.
xLAM is a family of Large Action Models (LAMs) optimized for function calling, reasoning, and planning in agentic AI systems. Unlike traditional LLMs that focus on text generation, LAMs are built to execute real-world actions by calling APIs, tools, and functions.
The xLAM family includes several model variants:
| Model | Parameters | Base Architecture | Use Case |
|---|---|---|---|
| xLAM-1B-fc-r | 1.3B | DeepSeekCoder-1.3B-instruct | Lightweight function calling |
| xLAM-7B-fc-r | 7B | DeepSeek-Coder-7B-instruct-v1.5 | Function calling and tool use |
| xLAM-7b-r | 7B | General | Academic exploration with limited GPU resources |
| xLAM-8x7b-r | 8x7B (MoE) | Mixture-of-Experts | Industrial applications balancing latency and performance |
The xLAM models were trained using the APIGen pipeline, which utilizes 3,673 executable APIs spanning 21 categories. Each training example undergoes a three-step verification process: format checks, actual function executions, and semantic verification.
On the Berkeley Function Calling Leaderboard, xLAM-7B-fc achieved the second-highest score overall (as of August 2024), outperforming larger models including GPT-4 and Claude 3 Opus despite being significantly smaller and more cost-effective.
ProGen is a protein generation model that applies natural language processing techniques to amino acid sequences. Rather than generating English text, ProGen treats amino acids as words and protein sequences as sentences, learning the grammar of protein structure from evolutionary data.
The model is a 1.2-billion-parameter Transformer-based neural network trained on 280 million protein sequences spanning more than 19,000 protein families. Training data was augmented with control tags specifying protein properties such as biological function and taxonomic origin.
ProGen's results were published in Nature Biotechnology on January 26, 2023. The research, conducted in collaboration with UC San Francisco and Tierra Biosciences, demonstrated that 73% of ProGen's artificially generated proteins were functional, compared to 59% of natural proteins tested under the same conditions. Artificial proteins fine-tuned to five distinct lysozyme families showed catalytic efficiencies comparable to natural lysozymes, even when sequence identity to natural proteins was as low as 31.4%.
LAVIS (LAnguage-VISion) is an open-source deep learning library developed by Salesforce AI Research for training and evaluating language-vision models. Released in September 2022, the library provides access to over 30 pre-trained and task-specific fine-tuned model checkpoints across four foundation model families: ALBEF, BLIP, CLIP, and ALPRO. LAVIS supports a range of vision-language tasks including visual question answering, image captioning, image-text retrieval, and multimodal classification.
Salesforce unveiled Agentforce on September 12, 2024, at the Dreamforce conference. Agentforce represents a strategic shift from copilot-style AI assistants to fully autonomous AI agents that can analyze data, make decisions, and take action on tasks without requiring constant human intervention.
The platform combines three major Salesforce tools: Agent Builder (for creating and configuring agents using low-code tools), Model Builder (for selecting and deploying AI models), and Prompt Builder (for designing and testing prompts).
The initial Agentforce release included several pre-built agents:
| Agent | Function |
|---|---|
| Service Agent | Handles customer service inquiries autonomously, replacing traditional chatbots |
| Sales Development Representative | Engages with inbound leads around the clock, qualifying prospects and scheduling meetings |
| Sales Coach | Provides personalized training and feedback for sales teams using role-play scenarios |
| Campaign Optimizer | Manages marketing campaign lifecycles, optimizing targeting and spend |
| Personal Shopper | Recommends products to consumers based on browsing behavior and preferences |
| Buyer Agent | Helps B2B customers find products, place orders, and track shipments |
| Merchant Agent | Handles site merchandising tasks including product placement and promotions |
Agentforce became generally available on October 25, 2024, with pricing starting at $2 per conversation.
On December 17, 2024, Salesforce announced Agentforce 2.0, positioning it as "the first digital labor platform for enterprises." The release introduced new pre-built skills for sales development, sales coaching, marketing campaign management, and commerce merchandising, along with an enhanced Agent Builder capable of interpreting natural language instructions to auto-generate new agents.
Key additions in Agentforce 2.0 included:
Agentforce 2.0 became generally available in February 2025.
At Dreamforce 2025, Salesforce announced Agentforce 360, described as "the world's first platform designed to connect humans and AI agents in one trusted system." This release introduced several new capabilities including Agentforce Builder (a conversational workspace for building, testing, and refining agents), Agent Script (a scripting language for controlling agent behavior with the predictability of code), Agentforce Voice (a voice-first platform for natural conversations with AI agents), and Intelligent Context (a feature that grounds agents in complex, unstructured business data through the Data 360 engine).
Salesforce has funded AI development externally through Salesforce Ventures, the company's corporate venture capital arm.
| Date | Milestone | Total AI Commitment |
|---|---|---|
| March 7, 2023 | Launched $250M Generative AI Fund; initial investments in Anthropic, Cohere, You.com, and Hearth.AI | $250 million |
| June 12, 2023 | Doubled the fund to $500M; added investments in Humane and Tribble | $500 million |
| March 2024 | Led a $106M funding round in Together AI | Growing |
| Mid-2024 | Announced another $500M AI fund | $1 billion |
By mid-2024, Salesforce Ventures had invested in over two dozen AI companies, including Anthropic, Cohere, Hugging Face, Mistral AI, Runway, and Together AI. The $1 billion total commitment made Salesforce Ventures one of the largest corporate investors in the generative AI ecosystem.
Salesforce co-founder and CEO Marc Benioff has positioned AI, particularly agentic AI, as the central pillar of the company's future strategy. At Dreamforce 2024, Benioff articulated a vision of "digital labor," describing autonomous AI agents as a multitrillion-dollar market opportunity. He argued that unlike copilots and chatbots, which require continuous human guidance, AI agents can independently retrieve data, build action plans, and execute tasks.
Salesforce backed this vision with early adoption data, reporting 32,000 agent-driven conversations per week, with 83% of customer support cases resolved autonomously and 50% fewer escalations to human agents.
In 2025, Benioff described agentic AI as reshaping Salesforce's internal operations as well. The company redeployed thousands of employees internally rather than expanding headcount, with engineering hiring remaining largely flat while AI-driven productivity increased. Benioff's annual strategic planning document (the V2MOM, standing for Vision, Values, Methods, Obstacles, and Measurements) emphasized industry-specific AI agents for sectors including automotive, pharmaceuticals, and government.
Salesforce's AI-driven products have shown rapid revenue growth:
| Metric | Value (FY2026) |
|---|---|
| Total Salesforce Revenue | $41.5 billion (10% YoY growth) |
| Combined AI and Data Cloud ARR | $2.9 billion (114% YoY growth) |
| Agentforce ARR | $800 million (169% YoY growth) |
| Organic AI ARR (excluding acquisitions) | $1.8 billion (100%+ YoY growth) |
While Salesforce's overall revenue growth remained in the 10% range for fiscal year 2026, its AI initiatives experienced triple-digit growth rates and became the company's most significant growth driver.
Salesforce competes with several major technology companies in the enterprise AI space:
| Competitor | AI Product | Approach |
|---|---|---|
| Microsoft | Microsoft 365 Copilot, Copilot Studio, Dynamics 365 AI | Integrated AI across Office, Teams, and Dynamics 365; horizontal platform play with 60% of Fortune 500 adoption for Copilot |
| Oracle | Oracle AI Agent Studio, Fusion Applications AI | AI agents embedded in ERP, CRM, and HCM applications; emphasis on unified data models |
| Vertex AI, Gemini for Workspace | Cloud-native AI platform with Gemini models integrated into Google Workspace | |
| SAP | SAP Joule, Business AI | AI copilot integrated into SAP S/4HANA and SuccessFactors |
| ServiceNow | Now Assist | AI assistant for IT service management and workflow automation |
The primary rivalry is between Salesforce and Microsoft, the two largest players in CRM. Salesforce maintains a customer base of over 202,600 organizations, compared to Microsoft Dynamics 365's approximately 91,400. While Microsoft leverages its broader ecosystem (Azure, Office 365, Teams) for cross-selling AI capabilities, Salesforce argues that its purpose-built, CRM-native AI delivers deeper functionality for customer-facing workflows.
The competitive landscape shifted in 2024 and 2025 as all major enterprise vendors introduced autonomous AI agent capabilities, moving beyond basic copilot functionality. Salesforce's Agentforce, Microsoft's Copilot Studio, and Oracle's AI Agent Studio represent converging approaches to agentic AI, with differentiation increasingly hinging on data integration, trust and security features, and industry-specific customization.
| Name | Type | Year | Key Details |
|---|---|---|---|
| Einstein AI | AI platform | 2016 | Predictive analytics, lead scoring, NLP built into Salesforce CRM |
| CodeGen 1.0 | Code generation LLM | 2022 | 350M to 16B parameters; competitive with OpenAI Codex |
| BLIP | Vision-language model | 2022 | Unified vision-language understanding and generation |
| LAVIS | Open-source library | 2022 | 30+ pre-trained model checkpoints for language-vision tasks |
| BLIP-2 | Vision-language model | January 2023 | Q-Former architecture; 54x fewer trainable parameters than Flamingo80B |
| Einstein GPT | Generative AI for CRM | March 2023 | First generative AI CRM tool; integrates OpenAI and Salesforce models |
| ProGen | Protein generation model | January 2023 | 1.2B parameters; 73% of generated proteins functional |
| CodeGen2 | Code generation LLM | May 2023 | 1B to 16B parameters; infilling support; 19 programming languages |
| AI Cloud | Enterprise AI suite | June 2023 | Einstein Trust Layer for data security and compliance |
| xGen-7B | Large language model | July 2023 | 7B parameters; 8K context; 1.5T training tokens |
| CodeGen2.5 | Code generation LLM | July 2023 | 7B parameters; trained on 783 GB of code in 86 languages |
| Einstein Copilot | Conversational AI assistant | September 2023 | Natural language CRM assistant with Copilot Studio |
| xGen-MM (BLIP-3) | Multimodal model | May 2024 | Multi-image understanding; scalable vision token sampler |
| xLAM | Large action model | 2024 | Function calling; outperformed GPT-4 on Berkeley Leaderboard |
| xGen-Code | Code generation LLM | 2024 | Salesforce-specific developer use cases; powers Agentforce for Developers |
| xGen-Sales | Sales-specific LLM | 2024 | Human-in-the-loop reinforcement learning for sales tasks |
| Agentforce | Autonomous AI agents | September 2024 | Pre-built agents for service, sales, marketing, and commerce |
| Agentforce 2.0 | Digital labor platform | December 2024 | MuleSoft integration, Slack deployment, enhanced reasoning |
| Agentforce 360 | Agent platform | October 2025 | Voice, Agent Script, Intelligent Context, and Builder workspace |