"How much does Agentforce cost?" is really two questions. One is what Salesforce charges you to run it. The other is what it costs to build an agent that works. People usually ask the first and get surprised by the second.
Short answer: the build cost is driven by how many jobs the agent does, how much custom logic sits behind it, how clean the data it reads is, and how many systems it has to reach. A first agent with a tight scope typically takes 4 to 6 weeks. The way to keep it affordable is to pick one set of conversations, ship that well, and add from there.
The Two Costs
Salesforce's side. Agentforce licensing and usage pricing comes from Salesforce, and it has changed several times since launch. Get the current model from your account executive, and ask how usage is counted for the channel you plan to use. That number belongs in the business case before anyone builds.
The build. This is the design, configuration, code, testing and rollout. It's where the scope decisions live, and it's what the rest of this post is about.
What Drives the Build Cost
Topics. Each topic is a job the agent owns, like order status or case triage. More topics means more instructions to write, more overlap to untangle and more test cases. Topics that overlap are the most common reason an agent picks the wrong path.
Actions. Standard actions and simple Flows are quick. Custom Apex actions take real build and test time, especially when they have to return exactly what the reasoning engine needs. Count the actions early, and mark which ones need code.
Grounding data. An agent is only as good as what it reads. Current, well-structured Knowledge articles and clean CRM records make the build faster. If the articles are out of date or the records are messy, part of the project becomes fixing that, and it's better to know up front.
Integrations. If the agent needs data or actions from systems outside Salesforce, each one adds design, auth and testing. On the Agentforce project for a medical company, the agents reached external systems through MuleSoft, which is a scope item of its own.
Channel. Chat on a website, an employee agent inside Salesforce, email and voice all behave differently. Voice and email bring transcripts and messy input, so they need more matching logic and more testing than a chat agent with clean questions.
Matching the right record. This one gets underestimated. When one contact belongs to many accounts, a simple lookup picks the wrong one and the agent's work lands where nobody sees it. On the medical project the fuzzy matching went through three versions before it was right.
Testing and guardrails. Test sets in Testing Center, what the agent must never do, and the handoff to a person. Skipping this is how agents stall between the demo and production, so it's part of the cost, not an extra.
Rollout. Training, an adoption dashboard and a few weeks of hypercare while real conversations arrive. Agents get better after launch, but only if someone is watching.
Keeping a First Agent Small
The cheapest agent is the one with a clear boundary. A good first scope looks like this:
- One agent, service or employee facing
- A handful of topics that don't overlap
- Mostly standard actions, with code only where it earns its place
- Grounded on Knowledge and CRM data you already trust
- A handoff to a person for everything outside the scope
That shape typically ships in 4 to 6 weeks. Once it's live and you can see which conversations it handles well, adding topics or a second agent is a smaller, cheaper project because the foundation and the tests already exist.
Questions to Ask Before You Get a Quote
- Which conversations should the agent own, and which stay with people?
- Which actions need custom Apex, and which can use standard actions or Flow?
- Is the Knowledge the agent will read current, and who owns it?
- Which systems outside Salesforce does it need to reach?
- How will we test it before customers see a change?
- What does Salesforce charge us to run it at our expected volume?
If two quotes are far apart, compare them on topics, custom actions and integrations. That's usually where the difference is.
Want a second opinion on how your first agent should be scoped? Book a call. More on how I build them on the Agentforce page, and on the scoping models in fixed scope vs time and materials.