Using Salesforce AI: Agentforce, AI coding agents or custom pipelines
Last reviewed: October 15, 2026
When people talk about "Salesforce AI", most of them mean the features Salesforce ships out of the box: Agentforce for AI agents, Einstein for predictions, and the Einstein Trust Layer, which secures requests to language models. In practice, there are three ways to put AI to work in Salesforce:
- As a runtime agent: Agentforce or an external agent that acts while the org is in use.
- As an AI coding agent: tools that write Apex, Flows and configuration.
- As a custom AI pipeline: your own logic that calls a language model of your choice.
All three draw on the same data, permissions and automation. So an AI project rarely succeeds or stalls on the choice of model alone. What decides it is the Salesforce org the project runs in.
The building blocks of Salesforce AI
Salesforce brings its AI features together on the Agentforce 360 platform. Data 360 serves as the data layer for external sources and unstructured content.
- Agentforce: An agent works with instructions, defined areas of responsibility and actions. Through those actions, such as Flows, Apex or prompt templates, the agent carries out work in the org. The Atlas Reasoning Engine picks the right action. For fixed processes like a refund, Agent Script locks the order of steps in place.
- Einstein: Einstein covers the classic predictive features, such as Einstein Lead Scoring, which learn from historical data. Prompt Builder adds templates that pull field values, Flows and Apex into generated responses as context.
- Editions: The Core, Advanced and Max editions include Agentforce, Slack and an allowance of Flex Credits. Existing contracts keep their current pricing.
| Building block | What it does | What it draws from your org |
|---|---|---|
| Agentforce | Agents that understand requests and act on them | Flows, Apex, prompt templates, the user's access |
| Einstein, predictive | Scores and forecasts | Closed records, consistently maintained stages |
| Prompt Builder | Templates for generated responses | Field values, Flows, Apex logic |
| Data 360 | A data layer for external and unstructured data | Data streams, mappings, identity resolution |
| Einstein Trust Layer | Masking, toxicity detection and zero data retention | No org data. The layer only secures the requests |
The Einstein Trust Layer in detail
The Einstein Trust Layer governs nothing but the path data takes to the language model. The layer masks sensitive values, checks responses for toxicity, and relies on zero data retention agreements so that partner models don't store prompts. The Trust Layer doesn't check whether the data it passes along is correct.
Three ways to use AI in Salesforce
The three routes don't exclude each other. They often run side by side, say a service agent in a customer channel and a coding agent in the delivery team.
1. Runtime agents: native and external agents
Runtime agents act in day-to-day operations. You configure Agentforce agents in Agentforce Builder, and they use the Flows and Apex classes you already have. For permissions, it matters who is talking to the agent:
- Employee-facing agents work in the context of the logged-in person and see only what that person can see. If a user has picked up broad access such as Modify All Data over the years, the agent can use that access too.
- Agents in a customer channel run as a dedicated agent user. That user's permissions decide exactly which data and actions the conversation can reach.
An agent uses every permission available to it, exactly as instructed and as fast as your automation allows. Unlike a person, an agent doesn't pause in front of a record that looks implausible.
Third-party agents belong to this group as well. Through Salesforce Hosted MCP Servers, clients such as Claude, ChatGPT or Cursor connect to the org and act with the logged-in user's permissions. With Headless 360, Salesforce offers platform capabilities directly as an API or an MCP tool.
2. AI coding agents
Tools such as Agentforce Vibes, Claude Code or Cursor support the delivery team. These agents read metadata, draft Flows or Apex classes, and validate the changes in a sandbox through the Salesforce CLI. AI coding agents pick up the org's patterns along the way, outdated Flow patterns included.
3. Custom AI pipelines
Here your team builds AI steps straight into its own processes. A Flow or an Apex class sends a request to a language model, say to classify incoming cases or summarise contracts. You connect the model either through the platform or directly with the provider:
- Through the platform (Models API): With Bring Your Own LLM you add models to AI Models and call them through the Models API. The requests still pass through the Einstein Trust Layer.
- Direct callout: Apex callouts or the Flow HTTP Callout action call the API of OpenAI, Anthropic or Google directly through a named credential. That leaves you the most freedom over model and parameters, but you need your own rules for data protection, masking and logging.
| Route | Typical example | What it depends on in the org |
|---|---|---|
| Runtime agent for employees | A sales assistant summarises an opportunity | Exact user permissions, maintained fields |
| Runtime agent in a customer channel | A service agent answers an order-status question | A tightly scoped agent user, current Knowledge articles |
| Runtime agent over MCP | Claude reads and updates records | Traceable logs, exposed tools |
| AI coding agent | An agent drafts a Flow or an Apex class | Clean metadata, descriptions, sandbox tests |
| Custom AI pipeline via the Models API | A Flow classifies incoming cases | Prompt templates, Trust Layer settings |
| Custom AI pipeline with a direct callout | Apex sends contract text to an LLM provider's API | Named credential, your own masking rules |
What every route inherits from your org
Every route takes four areas from your org: data, data model, permissions and automation. Whatever is clearly governed there, AI can use efficiently. Whatever grew without structure, AI inherits just the same.
In Salesforce's State of Data and Analytics report, 89% of the data and analytics leaders surveyed name a strong data foundation as the most important factor for successful AI. The same group estimates that 26% of their data can't be trusted. Among the leaders whose companies already run AI in production, 89% have seen inaccurate or misleading AI outputs.
- Data: Agents answer from the records in the org, and predictive models learn from them. If your sales team maintains stages inconsistently, the AI learns exactly that inconsistency. Incomplete fields and duplicates lower the quality of every answer directly.
- Data model: Agents choose fields by name and description. When similarly named status fields carry no description, the AI has to guess.
- Permissions: Least privilege applies to AI agents even more strictly than to people. A service agent needs a permission set cut exactly to its use case.
- Automation: Agent actions call existing Flows and Apex classes. Overlapping or outdated logic quickly leads to unexpected behaviour.
From pilot to scale
A pilot shows whether a use case works. Scaling means the use case keeps working reliably with more users, across more channels and next to other agents. Every unclear corner of the org then shows up in every single request.
For automated testing, Salesforce offers the Agentforce Testing Center, which runs scenarios in batches in a sandbox. Every test run and every request consumes credits, so Flex Credits become a central budget figure as usage grows.
| Aspect | Pilot | At scale |
|---|---|---|
| Scope | One use case in one channel | Several agents, channels and teams |
| Data | Hand-picked, maintained records | Everything the user can access |
| Permissions | One test user | A dedicated, reviewed user per agent |
| Testing | Individual conversations by hand | Recurring batch tests in Testing Center |
| Cost | A modest test allowance | Planned Flex Credit consumption |
| Operations | The project team | A named owner in operations for logic and actions |
When a business process changes, instructions and actions have to change with it right away. Otherwise the agent keeps working with last quarter's process.
Where Pariveo comes in: making the foundation visible
For AI to scale past the pilot, it needs a foundation the team knows: dependable data, a clear data model, tightly scoped permissions and clean automation. Many orgs have grown for years under delivery pressure, so a precise overview is often missing.
That's where the Platform Assessment comes in. The assessment covers more than 200 criteria across 15 modules, including the data model, automation, access rights and the existing Agentforce configuration.
The executive summary rates AI readiness separately for runtime agents, AI coding agents and custom AI pipelines. The result shows where AI can already be used safely, where it will struggle and which steps to prioritise first. Because the analysis runs deterministically, you can measure progress exactly after the first round of improvements.