We build Agentforce agents and Einstein features where they fit. Where they do not, we connect Salesforce to Claude, GPT or Gemini through our MCP and RAG services. Either way: guardrails, evals and audit first.
Agentforce is Salesforce's framework for agents that run inside the platform. An agent is defined by three things. Topics are the jobs it may handle, such as "order status" or "reschedule an appointment". Instructions are the natural-language rules for each topic. Actions are what it can execute: Flows, invocable Apex, prompt templates and external APIs.
The Atlas reasoning engine reads a request, picks a topic and plans a sequence of actions. It grounds its answers in retrieved CRM and knowledge data, then completes the task or hands over to a person. You test it in Agent Builder and deploy it to a chat channel, an Experience Cloud site or an internal console.
Einstein is the older, narrower family: predictive scoring for leads and opportunities, case classification and routing, generative email drafts, conversation summaries and knowledge recommendations. Most of it is configuration rather than code, and most of it is worth switching on once the data behind it is clean.
Topics
What it may do
Scoped jobs with explicit boundaries. Three agents with four topics each are easier to test than one agent with twelve.
Actions
How it does it
Flows, invocable Apex, prompt templates, MuleSoft or REST actions. Each one is testable on its own.
Grounding
What it knows
CRM records, knowledge articles and retrievers over your data. No grounding, no reliable answers.
Where it fits
Three places native Salesforce AI earns its keep
Service
An agent that answers order, booking and account questions from knowledge and records, then hands over to a human with the transcript. Case summaries and reply drafts for the agents who take the handover.
Lead and opportunity scoring on your own history, inbound lead qualification by an agent, and email drafts that read the account before writing. Forecasting that explains itself.
Employee-facing agents for policies, pricing rules and "how do I" questions, grounded on the knowledge base your team actually maintains. Cheap to pilot, quick to measure.
When to use Agentforce, and when to call an external model
The honest answer is both, for different jobs. Agentforce is the fastest route for conversational work on CRM records. External models through our MCP and RAG services win when the data or the reasoning lives outside Salesforce.
Question
Lean towards Agentforce
Lean towards an external LLM via MCP or RAG
Where does the data live?
Mostly in Salesforce objects and knowledge
Spread across documents, databases, SharePoint, ticketing and ERP
What is the task?
Conversational service and sales actions on records
Claude, GPT, Gemini or open-weight; swap per task; run your own evals
Cost model
Usage-based pricing on top of licences
Per-token API pricing; batch and cache to control spend
Compliance boundary
Trust Layer controls inside the Salesforce boundary
Your own residency, logging and guardrails; on-premises possible
Who builds and maintains it?
Admin-led with Flow and prompt templates
Engineering-led; needs an API, evals and monitoring
Time to first value
Days to weeks for a scoped topic
Weeks for a grounded, evaluated pipeline
Our guide Salesforce and LLMs works through this decision with worked examples.
Architecture
Two paths from the same CRM event
Path A keeps everything inside Salesforce. Path B puts retrieval, tools and guardrails in a service we control, so the same logic can serve Salesforce, a portal and a support desk.
Governance
Guardrails on both paths, not just the native one
Salesforce's Einstein Trust Layer sits between prompts and models on the native path. In broad terms it grounds prompts in CRM data and masks personal data before it leaves the platform. It applies retention and toxicity controls and logs prompts and responses for audit. Those are good defaults. They are not a substitute for deciding which actions an agent may take unsupervised.
On the external path we build the equivalent controls ourselves. Personal data is redacted before the call. Anything retrieved from documents or email gets prompt-injection checks. Outputs are schema-constrained, everything is logged, and an evaluation set runs before every change.
Anything that moves money, commits a delivery date or messages a customer stays behind a human approval step until the evals say otherwise. Our AI security practice reviews both paths with the same checklist.
Written list of allowed actions per agent, with a named owner
Personal data masked or redacted before any model call
Retrieved content treated as untrusted input
Outputs validated against a schema before they touch a record
Human approval for commitments, payments and outbound messages
Evaluation set with pass criteria, run on every change
Prompt, response and action log retained for audit
Process
From candidate list to a measured agent in production
A first scoped agent, grounded and evaluated, typically takes 6–10 weeks. Switching on Einstein scoring or summaries on clean data takes 2–4 weeks.
Triage
Score candidate use cases on value, data readiness and risk. Pick two. Our AI roadmap service does this across the whole org.
Ground
Audit the knowledge, records and documents the agent will rely on. Fix gaps and stale content before writing a prompt.
Build
Topics, instructions and actions in Agentforce, or MCP tools and retrieval on the external path. Guardrails in both.
Evaluate
Run the evaluation set and a red-team pass. Fix, re-run, record the scores.
Do we need Data Cloud before we can use Agentforce?
Not always. Several grounding features lean on Data Cloud, and Salesforce's packaging changes often. Agents scoped to knowledge articles and standard records can start without a large Data Cloud programme. We confirm the current prerequisites with your account team during discovery and design around them.
Can Agentforce call Claude, GPT or Gemini?
Agentforce actions can call external services, and Salesforce offers routes to connect models hosted elsewhere. In practice we often keep a thin MCP server outside Salesforce that holds the retrieval, tools and guardrails. Both Agentforce and Apex then call it. That keeps the logic testable and portable.
How do you know whether an agent is actually working?
An evaluation set written before the build: real questions, expected outcomes, pass criteria. In production we track containment rate, handover quality, customer satisfaction on bot conversations, action error rate and cost per conversation. We also review the transcripts the evals flag.
What does Salesforce AI cost compared with an external model?
Agentforce is priced on usage on top of your Salesforce licences, and the pricing model changes. External models are priced per token, with batching and caching available to control spend. We model both for your expected volumes before you choose. Our LLM cost calculator gives a first estimate.
Is our CRM data used to train the models?
Salesforce's Trust Layer includes retention controls and commitments with its model partners; the major API providers we work with offer comparable commercial terms. We confirm the contractual position for each path and write it into your governance pack rather than relying on marketing pages.
Ready to make AI work in production?
Tell us about your systems, your data and your constraints. We reply within one business day with a straight answer and a suggested next step.