Agentforce vs. Custom Claude Integration: Which Should You Pick?

If you are evaluating Salesforce AI right now, you have probably hit this exact moment: Agentforce is on the roadmap, someone wants to use Claude, and nobody can explain clearly why you would pick one over the other — or whether you even have to choose.

Here is the short answer: you are asking the wrong question. Agentforce and Claude do not compete. They are designed for different layers of work. The mistake is not picking the wrong one — it is forcing everything into just one of them and wondering why it does not fit.

This post gives you the decision framework we use on every project. By the end you will know exactly which tool belongs where in your stack.

Does Claudeforce replace Agentforce?

No, and Salesforce has not said it does. Claudeforce is the name of a partnership, not a replacement product. What actually shipped is a plugin called Salesforce in Claude that lets a seller work their pipeline from inside Claude. Agentforce still owns the agent layer inside Salesforce. As of 28 August 2026 neither company has published a boundary between the two, so nobody outside Salesforce and Anthropic can tell you where one ends and the other starts.

Added 28 August 2026, two days after the announcement. We have dated the claims below because this is a beta product and the details will change.

What was actually announced

On 26 August 2026, Salesforce and Anthropic announced Claudeforce. In Salesforce’s own words, the partnership “launches with Salesforce in Claude, a Plugin with 37 prebuilt sales skills that enable sellers and agents to reason over live revenue context, automate pipeline updates, and take governed action right from Claude.”

The timeline, quoted from the press release: it “is available to select pilot customers now and expects to launch in open beta in September 2026,” and “additional prebuilt skills will begin launching in late 2026.” Several outlets reported that second date as Q3, which is not the wording either company used.

Marc Benioff’s framing was blunter: “Here, the UI is the AI.”

Two names, and they are not the same thing

Most of the coverage uses these interchangeably. They are three different things:

Claudeforce is the partnership brand. You cannot buy it or install it.
Salesforce in Claude is the product — the plugin, currently in closed pilot.
AIforce is the layer underneath, which Salesforce describes as “Salesforce’s trusted enterprise harness that brings all your business data and workflows to any agent through MCP servers, APIs, and CLI tools.”

If someone tells you their org is “running Claudeforce,” ask which of those three they mean.

Is this just MCP with a marketing name?

Largely, yes.

The data path is Salesforce’s Headless 360 server, reached over MCP. It exposes exactly four tools: Discover, Describe, Dispatch, and Dispatch (Read-Only). Authentication is OAuth 2.0 with PKCE through an External Client App. Classic Connected Apps are not supported, and you need the mcp_api scope plus refresh_token, because the standard api scope will not work.

Salesforce Hosted MCP Servers went generally available in April 2026 for every Enterprise Edition org and above. Most of the launch coverage did not mention this. The plumbing has been shipping for four months, so if your org is on API v67.0 or later and Enterprise Edition or above, the connection this announcement is built on is already available to you.

The 37 “skills” are Agent Skills: SKILL.md markdown files with YAML frontmatter, under an open Apache-2.0 specification that Salesforce has itself published a tutorial for. Not compiled code, and not a black box. So what you get over a raw MCP connection is a well-written set of markdown instructions and centralised governance. That is worth something. It is also something you could write yourself rather than waiting for late 2026.

What it changes about the decision in this post

Most of it still stands. Two parts do not.

The framework further down still holds. Agentforce is still the right answer for AI that surfaces inside the Salesforce UI for a business user, maintained by an admin. Custom Claude work is still the right answer for background pipelines, cross-system orchestration and precise structured output.

The first change is a third option. Previously, “use Claude with Salesforce data” meant a developer building an integration. Now there is a supported path where a knowledge worker points Claude at the org and works from the Claude side, with no build. That third column does not compete with the other two. It competes with the seller opening Salesforce at all.

The second change is what your permissions model is now for. Every transaction through this path executes as the authenticated user and is attributed to that user in the audit trail. There are no service accounts and no autonomous credentials. Salesforce documents that CRUD, field-level security, sharing rules, profile permissions and permission sets all apply.

That cuts both ways. An agent inherits exactly what its user can already do. A profile that has carried View All Data since 2019 was survivable when a human was clicking through screens and knew not to go looking. It is a different thing when a model is reading across everything that profile can reach. Most orgs have never been audited for that, because until recently there was nothing to audit for.

What nobody has answered yet

As of 28 August 2026:

No price has been published by either company. Three separate cost lines are already visible, though: paid Claude seats for each seller, Salesforce headless/API consumption billed on API calls that vary by edition and licence, and an Enterprise Edition floor. Salesforce’s own CRO conceded you cannot buy this on one piece of paper yet.
No boundary against Agentforce has been published. Five commenters asked the question under Salesforce’s own launch videos. There was no reply.
Whether validation rules, Apex triggers, record-triggered Flows and duplicate rules fire on a write through this path is not documented anywhere. We checked Salesforce’s MCP security guidance directly. It is not addressed.
Prompt injection is not mentioned in that guidance either. Neither is untrusted content or exfiltration. A CRM is full of text that strangers wrote: web-to-lead submissions, inbound email, attachments, meeting notes. If you are evaluating this, put that question to your account team.
Anthropic has published nothing. No blog post, no landing page, no documentation, no video. We checked on 28 August and anthropic.com/news/claudeforce returns a 404. Every Anthropic quote in circulation, including Dario Amodei’s, exists only inside Salesforce’s press release.
There is no named external pilot customer, and the demo footage in the official videos carries a disclaimer that it is simulated with fictional data.

Most of that is what a two-day-old announcement looks like. The security questions are different — those will still be open in six months unless someone asks them out loud.

What to do before your renewal

If your Einstein or Agentforce renewal lands in the next two quarters, three things are worth doing now, and none of them require this product to exist:

Find out what edition and API version your org is actually on. Enterprise Edition or above, API v67.0 or later. If you are below either, that is the real prerequisite conversation, not the plugin.
Pull the list of users with View All Data and Modify All Data, and check how many of them got it from a profile rather than a deliberately granted permission set. Do this whether or not you ever adopt any of this. Most orgs have not looked in years.
Do not let anyone bundle this into an Agentforce renewal as a justification. They are separately licensed, the second one has no published price, and the open beta had not started when this was written.

The rest of this post is the decision framework, unchanged. It was written in May and it still applies.

	Agentforce	Custom Claude
Best for	Inside the Salesforce UI	Background processing, cross-system pipelines
Data leaves Salesforce?	No — Einstein Trust Layer	Yes — calls Anthropic API externally
Who maintains it?	Salesforce admin	Developer
Iteration speed	Seasonal release cadence	Hours
Cost model	Per-user license (~$75/user/mo)	Usage-based — pennies per job
The sweet spot	Conversational, UI-native, regulated	Complex reasoning, multi-system, batch

What Agentforce is

Agentforce is Salesforce’s native AI agent platform. It lives inside the Salesforce platform, uses your Salesforce data without leaving the Salesforce security boundary, and is configured through the Salesforce UI — Agent Builder, Prompt Builder, Flow.

Agentforce is genuinely good at:

Conversational interfaces for sales reps and service agents inside the Salesforce UI
Internal Q&A over Knowledge articles and Case history
Next Best Action recommendations surfaced in Lightning Experience
Structured workflows that need to stay inside Salesforce’s compliance boundary (healthcare, financial services, regulated industries)
Admin-friendly automation that does not require a developer to configure or maintain

Agentforce is less good at:

Complex logic that does not fit Agent Builder’s visual model
Integrations with systems outside Salesforce
Long-running or background processing — batch jobs, scheduled pipelines
Anything requiring precise output format control at scale
Use cases where you need rapid iteration — the release cadence for Agentforce features follows Salesforce’s seasonal releases

What custom Claude integration is

A custom Claude integration means calling the Anthropic API from outside Salesforce — typically from a Cloudflare Worker, AWS Lambda, or a scheduled job — and reading from and writing to Salesforce via the REST API or the Salesforce MCP server.

Claude is good at:

Complex reasoning tasks: document classification, contract analysis, multi-step research
Pipelines that span Salesforce and other systems — CRM, ERP, email, Slack, billing in one workflow
Background automation that runs unattended
Precise, structured output at scale: JSON extraction, scored records, classified categories
Rapid iteration — you can ship a new behavior in hours, not a release window
Use cases where you need exact control over what the model sees and how it responds

Claude is less good at:

Anything that needs to surface inside the Salesforce UI without custom LWC development
Regulated industries where data cannot leave the Salesforce trust layer
Admin-maintained automation — changes require a developer

The decision framework

Here are the four questions we ask on every project:

Does it need to surface inside Salesforce? If a sales rep needs to see AI output inline on a record, in a chat interface, or as a Next Best Action — Agentforce. If it is a background job, a document pipeline, or a system integration — Claude.

Does data need to stay in Salesforce? Compliance requirements that prevent sending data to external APIs point toward Agentforce and the Einstein Trust Layer. Mid-market B2B without regulated data? Claude.

How often will the logic change? If a business admin needs to update the behavior quarterly without involving a developer — Agentforce Prompt Builder. If the logic is stable and complex — Claude.

How precise does the output need to be? Agentforce handles general conversational tasks well. Claude is better when you need exact output: extracting specific fields, scoring against defined criteria, or producing structured data that writes directly to another system.

Real-world examples

Use Agentforce for:

A chat interface in Service Cloud where agents can ask “what is the history on this account?”
Next Best Action cards that suggest which leads to call today based on engagement signals
A Knowledge assistant in the Lightning Experience sidebar
Automated case summarization visible to service agents before they open a record

Use Claude for:

Ingesting vendor contracts (PDFs), extracting key terms and dates, writing structured data to Salesforce Contract records
Triaging inbound emails, classifying by intent and urgency, routing to the right Salesforce queue with a draft reply attached
Nightly scoring of all open opportunities against custom ICP criteria, updating a custom field on each
A QA agent that generates Apex test cases from requirement documents and runs them against a scratch org

Use both:

Agentforce handles the sales-rep-facing chat experience inside Salesforce; Claude handles the nightly data enrichment and lead scoring pipeline behind it
Agentforce manages the service agent assist layer; Claude processes document uploads in the background and writes extracted data to Case records before the agent opens the case

Cost comparison

Agentforce: Requires Einstein licenses. Roughly $75/user/month for Einstein 1 (which includes Agentforce capabilities), or add-on pricing for specific agents. For 20 users, that is around $18,000/year in license cost before any implementation work. The implementation itself — configuring agents, building Prompt Builder templates, writing grounding instructions — typically runs $10,000 – $40,000 depending on complexity.

Claude API: Usage-based. Claude Haiku (fast, cheap) runs approximately $0.80 per million input tokens and $4 per million output tokens. Claude Sonnet (the balanced choice for most production work) is around $3/$15. A document processing pipeline handling 1,000 documents per month at 2,000 tokens each costs a few dollars a month at Haiku rates. For most background automation use cases, API costs are noise compared to the labor they replace.

The cost comparison depends on volume and use case. For sales-rep-facing AI inside Salesforce, Agentforce is probably the right licensing investment. For background automation, Claude API costs are usually negligible against the value generated.

The common mistake

The most common error we see is choosing based on identity rather than fit.

“We should use Agentforce because we are all-in on Salesforce” leads to forcing complex external automations into a platform not designed for them — slow to iterate, painful to debug, and limited in what it can connect to.

“We should build everything in Claude because we are engineers” leads to building things outside Salesforce that should be native — and creating friction for business users who live in Lightning every day.

The best projects use both: Agentforce for the Salesforce-native, user-facing, admin-maintainable layer; Claude for the heavy processing, cross-system orchestration, and complex reasoning behind the scenes.

Frequently asked questions

Is Agentforce the same as Einstein GPT? Agentforce evolved from Einstein GPT. Einstein GPT (2023) was Salesforce’s first conversational AI layer. Agentforce (launched 2024, significantly expanded in 2025–2026) adds the Agent Builder, Topics, and Actions framework, and deeper grounding capabilities. They share the Einstein Trust Layer but Agentforce is substantially more capable and is the current product Salesforce is investing in.

Can I use Claude inside the Salesforce UI? Not natively — Claude is not a Salesforce-certified product and does not run inside the Salesforce security boundary. The typical pattern is to call the Anthropic API from a Cloudflare Worker or AWS Lambda, read and write Salesforce data via the REST API, and optionally surface results inside Salesforce through a custom Lightning Web Component. That last step requires developer work.

Is Agentforce available on all Salesforce editions? Agentforce requires Einstein licenses. Einstein 1 Sales or Service editions (which include Agentforce capabilities) run approximately $75/user/month as of 2026, on top of your core Salesforce license cost. Availability and exact pricing vary by contract — your Salesforce account executive can confirm what your org is entitled to.

How long does it take to build an Agentforce agent vs. a Claude integration? A basic Agentforce agent scoped to one Topic with three to five Actions typically takes two to four weeks to configure, test, and deploy properly. A Claude integration handling a specific background task — document triage, lead scoring, email classification — typically takes one to three weeks to build and deploy. Both timelines assume clear, locked requirements at the start.

Should I start with Agentforce or Claude if I have never done Salesforce AI before? Start with Claude if your use case involves processing documents, running background pipelines, or integrating multiple systems. Start with Agentforce if your team needs AI assistance directly inside the Salesforce UI and your admins want to configure and maintain it without a developer. If you are not sure, book a 30-minute call — most projects have a clear answer within the first conversation.

Building Salesforce AI and not sure which layer to use for your specific use case? Book a 30-minute call. We will look at your requirements and tell you which tool fits — and whether the answer is one, the other, or both.