Intelligent Apps
An intelligent app is a business application with artificial intelligence built directly into the workflow, so the software predicts outcomes, recommends a next step, or generates content while a person works.
Definition
An intelligent app is a business application with artificial intelligence built directly into the workflow, so the software predicts outcomes, recommends a next step, or generates content while a person works. Instead of only storing records and running fixed rules, an intelligent app learns from data and adapts as conditions change.
In Salesforce, intelligent apps are the products and features that put AI inside the CRM. They run on the Einstein and Agentforce AI portfolio: predictive scoring, generative drafting and summaries, conversational copilots, and autonomous agents. The goal is the same in every case. Surface a useful recommendation or action right where the user already does their job, grounded in trusted Salesforce data.
How Salesforce builds AI into the application
What makes an app "intelligent"
The phrase describes software that goes past static records and hard-coded rules. An intelligent app uses machine learning or generative AI to read context and respond with a prediction, a recommendation, or generated text. A lead record stops being a passive form and starts carrying a score for how likely it is to convert. A case stops being a blank text box and arrives with a suggested reply and a draft summary. Two traits separate this from older analytics. First, the app learns from historical data and improves as more data accumulates, rather than depending on a fixed formula an admin wrote once. Second, the intelligence is embedded in the moment of work. The recommendation appears on the page the user is already viewing, not in a separate dashboard they have to remember to open. Salesforce delivers this pattern across Sales, Service, Marketing, and its industry clouds, which is why "intelligent app" reads as a category rather than one product name.
Predictive Einstein: scoring and recommendations
The first generation of Salesforce intelligent apps was predictive. These features look at past records and output a number or a ranked recommendation. Einstein Lead Scoring rates how likely each lead is to convert, so reps work the best ones first. Einstein Opportunity Scoring estimates close probability on a deal. Einstein Case Classification predicts field values on incoming cases to speed up routing. Predictive Einstein learns from your own org data. It studies which leads converted and which did not, then builds a model that scores new leads on the same signals. Because the model is trained on your history, the factors it weighs reflect your business, not a generic template. Admins turn these features on, let the model build, and the scores appear on the record. For custom objects and bespoke questions, Einstein Prediction Builder lets an admin train a model on any field without writing code. This predictive layer remains widely used and sits alongside the newer generative and agentic capabilities rather than being replaced by them.
Generative Einstein: drafting and summaries
The second wave added generative AI, which produces new text instead of only scoring existing records. Generative Einstein drafts sales emails, summarizes long case threads, writes knowledge articles, and proposes service replies. The user reviews and edits the output, so the AI accelerates the work without removing human judgment. The control point for this layer is Prompt Builder. An admin designs a reusable prompt template that pulls fields and related records from Salesforce, then sends that grounded prompt to a large language model. Because the prompt is built from live CRM data, the generated text reflects the specific account, contact, or case rather than a generic guess. Salesforce calls this grounding, and it is what keeps generated content relevant. The Models API and the Prompt Template Connect API let developers call the same models from Apex or external code. Generative features need Einstein turned on, and most depend on Data Cloud for the grounding data that makes the output accurate.
Agentforce: autonomous agents
The current frontier of Salesforce intelligent apps is Agentforce, described by Salesforce as the agent-driven layer of the platform. An agent does more than answer one question. It can plan and carry out a multi-step task, such as qualifying a lead, resolving a service issue, or booking a meeting, and it can act on Salesforce data along the way. Agentforce agents are built from topics, which group the jobs an agent handles, and actions, which are the concrete operations the agent can take. An action might query records, run a flow, or call Apex. Agents can run in a chat window for customers or assist employees inside Salesforce. They can also start work proactively when a CRM record changes or a scheduled event fires, rather than waiting for someone to type a request. This shifts the model from "AI that suggests" to "AI that does," with the user setting guardrails on which actions an agent is allowed to perform.
The Atlas Reasoning Engine underneath
Agentforce agents run on the Atlas Reasoning Engine, which Salesforce calls the brain behind Agentforce. Atlas decides how to handle a request. It interprets the user's goal, retrieves the right data, picks an action, checks the result, and repeats until the task is done. Salesforce describes this as a reason, act, and observe loop. Atlas leans on retrieval augmented generation, often shortened to RAG. Rather than relying only on what a language model already knows, the engine pulls relevant records and knowledge from your org and feeds them into its reasoning. That keeps answers tied to your actual data. Atlas also exposes its reasoning steps, so admins can see why an agent chose a given action. That visibility matters for trust, because it lets a team review decisions, catch mistakes, and refine the agent's instructions. The same engine powers conversational copilot experiences, so the reasoning approach is consistent across the agentic parts of the platform.
The Einstein Trust Layer
Putting AI inside a CRM raises an obvious worry. Sensitive customer data should not leak to an outside model or get reused to train it. The Einstein Trust Layer is the security boundary that addresses this, and it wraps every generative call. It applies several protections. Dynamic grounding injects only the data the running user is allowed to see, so an agent never surfaces records that person could not access directly. Data masking detects sensitive values like names and replaces them with placeholders before the prompt leaves Salesforce. Toxicity detection scans prompts and responses for unsafe content. An audit trail records the prompt, the raw response, toxicity scores, and feedback as timestamped metadata. Zero data retention agreements with model providers stop third parties from storing or training on your data. For any regulated workload, confirm these controls are configured before you let an intelligent app touch live customer information. The Trust Layer is what makes embedding AI in enterprise data defensible rather than risky.
Choosing the right intelligent app
With predictive scoring, generative drafting, and autonomous agents all available, the practical question is which one fits a given problem. The cleanest way to decide is to start from the outcome you want, not the technology label. If you need to rank or prioritize existing records, reach for predictive Einstein, such as Lead Scoring or a custom Prediction Builder model. If you need to speed up writing or reading, such as drafting replies or summarizing long threads, generative Einstein and Prompt Builder fit best. If you need to automate a whole multi-step process with minimal human clicks, an Agentforce agent is the right tool. Many orgs run all three at once, because they solve different jobs. Watch the cost model too. Predictive features are often bundled with a Sales or Service edition, while generative and agentic usage consumes credits that scale with volume. Pair each intelligent app with a measurable result, then embed it where users already work so the AI gets used instead of ignored.
Enabling intelligent apps in your org
Most Salesforce intelligent apps share one prerequisite: Einstein must be enabled in your org, and generative or agentic features also need Data Cloud and the Einstein Trust Layer in place. Here is the high-level setup an admin follows before any AI feature will appear.
- Confirm your edition and licenses
Check that your Salesforce edition includes the AI features you want. Predictive Einstein, generative features, and Agentforce have different licensing, and some need an add-on. Verify entitlements in Setup before you promise a feature to users.
- Turn on Einstein
In Setup, open the Einstein setup area and enable Einstein generative AI. This is the master switch. Many features stay hidden in their own setup pages until this is on.
- Provision Data Cloud and the Trust Layer
Generative features ground their prompts in data, so connect Data Cloud and confirm the Einstein Trust Layer settings for masking, grounding, and audit. Do this before testing with real records.
- Enable and configure the specific app
Go to the feature's own setup page, such as Einstein Lead Scoring or an Agentforce agent, and configure it. Let predictive models build, design prompt templates, or define an agent's topics and actions.
- Surface it for users and test
Add the scoring field, prompt action, or agent to the page layouts and flows where users work. Then test with a sample record and review the Trust Layer audit trail to confirm the output looks right.
Scoring and recommendation features like Einstein Lead Scoring that learn from your org history. Lowest setup effort; let the model build.
Drafting and summarization driven by prompt templates grounded in CRM data. Needs Data Cloud and Trust Layer configured.
Autonomous agents that complete multi-step tasks using topics and actions, running on the Atlas Reasoning Engine. Highest setup and governance effort.
A point-and-click way to train a custom prediction on any field or custom object without code, for predictions Salesforce does not ship out of the box.
- Generative features usually fail silently or stay hidden until Data Cloud and the Trust Layer are fully provisioned, so set those up first.
- Predictive models need enough quality history to build. A new org or a sparsely filled object can produce weak or unavailable scores.
- Generative and agentic usage consumes credits that scale with volume. Model the cost before a broad rollout, because it can exceed seat-based pricing.
- An agent can only do what its actions allow. Scope actions and permissions tightly, since an agent runs with real access to your data.
Prefer this walkthrough as its own page? How to Intelligent Apps in Salesforce, step by step
Trust & references
Cross-checked against the following references.
Straight from the source - Salesforce's reference material on Intelligent Apps.
Hands-on resources to go deeper on Intelligent Apps.
About the Author
Dipojjal Chakrabarti is a B2C Solution Architect with 29 Salesforce certifications and over 13 years in the Salesforce ecosystem. He runs salesforcedictionary.com to help admins, developers, architects, and cert/interview candidates sharpen their fundamentals. More about Dipojjal.
Test your knowledge
Q1. What does the Intelligent Apps category in Salesforce describe?
Q2. Which engine, launched in 2024, powers the multi-step agents in the modern Intelligent Apps portfolio?
Q3. What does the Einstein Trust Layer add around every AI call in the Intelligent Apps stack?
Discussion
Loading discussion…