Comparison

    Agentforce vs custom AI agents for your CRM

    Agentforce is the right choice when the work happens inside Salesforce, the users are Salesforce users, and you want governance, logging, and permissions handled by the platform. A custom agent — typically Claude connected to Salesforce through an MCP server — is the right choice when the work spans systems, the users live in Slack or an IDE rather than Salesforce, or you need model and prompt control that a packaged agent platform does not give you. Most teams that get value end up running both: Agentforce for in-CRM, customer-facing flows, and a custom agent for internal, cross-system work.

    Last updated July 30, 2026 · Skydog Ops

    Side by side

    DimensionAgentforceCustom agent (Claude + MCP)
    Where it runsInside Salesforce UI and channelsSlack, Teams, IDE, or your own app
    Data groundingData Cloud, records, knowledgeWhatever you connect — CRM, warehouse, docs
    PermissionsInherits Salesforce sharing modelYou scope them; needs deliberate design
    Model choiceSalesforce-managedYour choice, swappable
    Cost modelPer-conversation / flex creditsAPI token usage plus build
    Build effortConfig-heavy, low codeEngineering project, 3 – 8 weeks
    Cross-system reachLimited outside SalesforceIts main advantage
    Audit and governanceBuilt inYou build it

    Where Agentforce is the right answer

    • Customer-facing service deflection on cases, where the answer lives in Salesforce Knowledge.
    • Anything that must respect Salesforce sharing rules exactly, with no separate permission layer to reason about.
    • Organizations that need every agent action logged inside the platform for compliance review.
    • Teams without engineering capacity to own an agent in production.

    Where a custom agent is the right answer

    • The question spans systems — CRM plus billing plus the warehouse plus a doc store.
    • The users are executives or reps who live in Slack and will not open Salesforce to ask a question.
    • You need to control the model, the prompt, and the retrieval strategy, and to change them weekly.
    • The workload is read-heavy analysis where token cost is far below per-conversation pricing.

    The part that decides success, either way

    Neither option fails because of the model. They fail because the CRM data underneath is wrong — duplicate accounts, stale stages, unowned records, free-text fields where a picklist belongs.

    An agent grounded in bad data produces confident wrong answers faster than a human would, which is worse than no agent. Data remediation is not a prerequisite you can skip; it is most of the project.

    Write access is the second decision. Start read-only, add scoped writes to specific fields on specific objects with a confirmation step, and log every write with the prompt that caused it.

    How Skydog builds these

    Skydog Ops deploys both: Agentforce grounded in Data Cloud for in-CRM and service use cases, and Claude connected through a Salesforce MCP server for cross-system internal work.

    Every engagement starts with the data layer, then a read-only agent, then narrowly scoped writes. A Forward Deployed Engineer embeds with the team so the system is owned internally when we leave.

    FAQ

    Common questions

    What is Agentforce?

    Agentforce is Salesforce's platform for building and running AI agents inside Salesforce. Agents are configured with topics and actions, grounded in Salesforce records, Knowledge, and Data Cloud, and run under the org's existing permission and sharing model.

    Do you need Data Cloud for Agentforce?

    Not for simple record-based actions, but for anything requiring retrieval across unstructured content or unified customer profiles, yes in practice. Grounding quality is what separates a useful agent from a plausible-sounding one.

    Is a custom agent cheaper than Agentforce?

    For high-volume read-heavy internal use, usually yes — API token costs run well below per-conversation pricing. For low-volume customer-facing service, Agentforce is usually cheaper once you account for the engineering cost of building and maintaining your own.

    Can Claude read and write Salesforce data?

    Yes, through a Model Context Protocol server that exposes scoped Salesforce operations. Read access is straightforward. Write access should be limited to named fields and objects, gated behind confirmation, and fully logged.

    How long does it take to get an agent into production?

    A scoped, read-only agent on clean data takes 3 to 6 weeks. If the CRM data needs remediation first — which it usually does — add 4 to 8 weeks for that work before the agent is worth deploying.

    Skydog

    CRM & GTM Systems Design
    for the AI era.