Agentforce vs Custom AI Agent: Everyone Draws the Line Wrong

Two Kinds of AI Agent, and One Very Common Mistake

The draft on my desk last week said Agentforce lives inside Salesforce, and a custom AI agent handles everything outside it.

Clean line. Easy to say in a meeting. Wrong.

I sent it back. Not because the conclusion was bad, but because the reason underneath it was rotten, and a bad reason will eventually walk you into a bad decision. So we went and checked. What came back changed the comparison, and it made the answer a lot more useful than “one is inside, one is outside.”

Agentforce Is Not Trapped Inside Salesforce

This is the part most people have wrong, including some who sell it.

Agentforce has been able to call external systems for years, through several paths that are generally available today:

• External Services. Turns an OpenAPI spec into invocable actions. Long standing, GA.
• Apex and Flow HTTP callouts. These have been hitting outside endpoints for years. GA.
• MuleSoft. Exposes anything you have already integrated. GA.

Then there is the newer path, and this is where you need to be careful. Agentforce has a native MCP client, which lets an agent talk to any MCP compliant server without custom code. Track its status:

• Pilot in July 2025.
• Beta in January 2026.
• Still Beta as of Summer ’26.

Do not scope a load bearing integration on it yet. If a vendor tells you MCP makes the integration question disappear, ask them what “Beta” means in their contract. The GA answer for external actions today is still External Services and callouts.

So no, Agentforce is not confined to your org. Which kills the easy version of the comparison and forces a better one.

The Real Line Is Gravity, Not Reach

Here is the distinction that survived the fact check. Both kinds of agent can reach outside Salesforce. They reach differently.

So the question was never “which one can leave the building.” It is how much does the work weigh, and who is holding it.

• Agentforce operates on records. For many people, under your sharing model, at volume.
• A custom agent operates on context. Around one person, across whatever that person touches, on work that is different every time.

Agentforce vs Custom AI Agent: What Each One Is Actually For

Strip away the marketing and it comes down to the shape of the work.

Agentforce is strongest where work is structured, repeatable, and shared:

• Lead qualification and routing
• Opportunity hygiene
• Case deflection and order status
• Employee self service on approved policy

Anything that happens the same way four hundred times a week and needs a log afterward.

A custom agent is strongest where work is unstructured, one off, and personal:

• Drafting
• Research and synthesis
• Reasoning across systems that were never integrated
• The quarterly board update assembled from nine sources
• The nuanced escalation reply that a human still sends

Five questions sort almost any use case:

Left column answers point to Agentforce. Right column answers point to a custom agent. If you get a split result, you almost certainly have two use cases wearing one name, and you should separate them before anyone scopes anything.

My favorite edge case: summarize win and loss patterns from closed opportunity notes. The data is sitting in Salesforce, so people reflexively call it platform work. It is not. The work is unstructured analysis and the deliverable is a document. That is a custom agent job every time.

What It Costs, Because This Trips People Up

The two run on completely different money.

Agentforce is usage metered:

Agentforce scales with conversation volume. The budget risk is underestimating volume, or modeling too few actions per conversation.

A custom agent is a flat subscription:

Priced per person, plus a one time engagement to load context and wire up connectors. It scales with headcount, not workload: five requests a week or fifty, the number is the same. The budget risk is low and bounded, and the variable is adoption.

Model the Agentforce volume before you scope the rollout, not after.

Four Ways This Goes Wrong

How Agentforce and a Custom AI Agent Compound

Now the part that only makes sense once the reach myth is dead.

If the difference were inside versus outside, you would pick one and live with the gap. Because the difference is actually gravity, the two halves fit together into a loop. This is the argument, and it is specific:

1. Discover. The custom agent reads your process docs, tickets, and call transcripts, and drafts what the current state honestly is. Not what the SOP claims.
2. Specify. It writes the topic instructions, action definitions, and test cases in the exact form Agentforce needs. Writing requirements for an agent is design work, not transaction work, which is why it belongs on the custom side.
3. Deploy. Agentforce runs it at volume, in the org, under the sharing model, with the audit trail.
4. Refine. The custom agent reads the telemetry, finds where the agent is escalating, and drafts the next round.

The custom agent is how you build the Agentforce agent well. The Agentforce agent is how the work actually scales past one person. Neither replaces the other, and that is not a hedge. It is an architecture.

Where to Start

Frequently Asked Questions

Can Agentforce access systems outside Salesforce?

Yes. Agentforce reaches external systems through External Services, Apex and Flow HTTP callouts, and MuleSoft, all generally available. It also has a native MCP client, which remains in Beta as of Summer ’26. The constraint is not reach, it is that every external action must be built and governed in advance.

When should you use a custom AI agent instead of Agentforce?

Use a custom AI agent when the work is unstructured, serves one person, and produces a document rather than a record change. Drafting, research, synthesis, and cross system reasoning all sit here. Use Agentforce when the work is structured, repeats at volume, serves many users, and needs an audit trail.

How much does Agentforce cost per conversation?

Flex Credits run about $500 per 100,000, roughly ten cents per standard action. A typical conversation uses five to fifteen actions, so most land between fifty cents and a dollar fifty. Confirm current rates against your agreement before budgeting.

We Build Both, and We Will Tell You Which One You Need

We wrote the full decision framework up as a briefing, including the nine dimension comparison, the cost models, and nine worked examples with verdicts. It is free and there is no form. Download Agentforce or a Custom Agent? (PDF), or browse all of our briefings, including the Company Brain paper. For nine worked examples with verdicts, see Nine Agentforce Use Cases, Sorted.

If you want to skip the reading, bring us the list of processes you are already considering for automation. One working session sorts them into platform work and individual work, names your strongest first Agentforce use case, and stands up a custom agent for the person carrying your Salesforce agenda.

Borrow the expertise. Build the asset. Let’s talk.