What is Agentforce? A plain-English guide for 2026

Key takeaways

• Agentforce is Salesforce's platform for AI agents that don't just answer questions but take actions inside your Salesforce data and processes.
• It works through subagents (formerly topics), instructions and actions, planned by Salesforce's Atlas Reasoning Engine and grounded in your own CRM and Data 360 data.
• You'll also hear AIforce and SIforce: they're about using Salesforce from tools like Claude and Slack, with Agentforce agents underneath.
• Pricing is moving to usage. Today it's Flex Credits per action; Salesforce plans per-call metering for registered agents, with 30 days' notice (reported target: November 2026). Plan for it now.
• Start with one boring, frequent, low-risk job. Agents should earn promotions, just like people.

Ask ten people what Agentforce is and you'll get ten answers: a chatbot, a new licence, Einstein Copilot with a new name. It started close to that last one, but it's grown into something much bigger, and most explanations online haven't caught up.

I run Techila Global Services, a Salesforce Summit Partner, and we implement Agentforce for customers. So here's the plain-English version, from the field.

What Agentforce actually is

Agentforce is Salesforce's platform for building AI agents that can take actions on your business data, not just chat about it.

A chatbot answers a question. An Agentforce agent can look up the customer, check the order, update the case, book the follow-up and hand over to a person when it should. It does that inside Salesforce, using your records, your permissions and your processes.

Salesforce launched Agentforce at Dreamforce in 2024. At Dreamforce 2025 it became Agentforce 360: one platform that brings together agents, your data (Data 360), your Customer 360 apps and Slack.

My take: Stop thinking of Agentforce as "AI in Salesforce". Think of it as a new kind of employee that works inside your CRM. Everything about getting it right follows from that.

How an Agentforce agent works

An agent is built from topics, instructions and actions, and Salesforce's Atlas Reasoning Engine decides which to use for each request.

Here's the flow, based on Salesforce's own description of the Atlas Reasoning Engine:

1. A request comes in: a customer message, an employee question, or a trigger like a new case.
2. Atlas maps it to a topic. A topic is a job area, such as "order status" or "returns", with its own instructions and rules. If nothing fits, the request is treated as off-topic.
3. The topic's instructions and actions are added to the prompt, and the model plans the next step.
4. The agent runs an action. An action is a specific thing it can do: update a record, run a flow, call an API, summarise a case.
5. It checks the result and repeats until the job is done, or it escalates to a human.

Building block	What it means in practice
Topic	A job area the agent handles, with clear boundaries
Instructions	The rules, written like a job description
Actions	What the agent is allowed to do in your systems
Data grounding	Your CRM and Data 360 data, so answers are about your customers, not the internet
Guardrails	Salesforce's trust controls, permissions and human handover

I explain agents to business leaders the same way I'd describe a new hire: a brain (the model), a job description (instructions), tools (actions), knowledge (your data), a trigger, and guardrails. If you can onboard a person, you can design an agent.

What it looks like in a real business

The best first agents handle frequent, well-defined requests where a fast, consistent answer matters more than judgement.

At Techila we implemented Agentforce for Armakuni, a UK cloud engineering and DevOps consultancy, to handle customer support. The agent answers technical FAQs instantly, routes cases intelligently, and escalates to a human with the full context when real expertise is needed.

That pattern is where I'd start almost anyone:

• Service: answer common questions, check status, route and summarise cases.
• Sales: qualify inbound leads, research the account, draft follow-ups.
• Operations: chase missing information, update records, prepare daily summaries.

What these have in common: they happen many times a day, the rules are clear, and a mistake is easy to catch.

My take: The agent is rarely the hard part. The data, the permissions and the testing decide whether you can trust it. Budget your time accordingly.

A quick note on the names

You'll also hear "AIforce" and "SIforce". They're about where you use Salesforce from, not a replacement for Agentforce.

At Dreamforce 2026 Salesforce announced AIforce, which lets you use Salesforce from inside tools like Claude and Slack, and in October it was renamed SIforce. Agentforce agents still do the work underneath. I've written a separate guide on what the SIforce rename actually changes.

What Agentforce costs

Agentforce pricing is usage-based, and it's about to become more so, which means your agent design directly affects your bill.

According to Salesforce's pricing page, most Agentforce usage is now paid for with Flex Credits: each action an agent takes uses 20 credits (about $0.10 at list price), voice actions use more, and an interaction that doesn't need an action isn't charged. Your contract may differ, so check your own terms.

The bigger change is coming. Salesforce's help article Understand AIforce Impact describes metering every successful call a registered agent makes to a production org over MCP or the API as a "Headless Platform Interaction", with 30 days' notice before it starts (reported target: November 2026). The rate hasn't been published yet.

What I'd do this month if you run Salesforce:

1. List every agent and integration that touches Salesforce, including the ones IT doesn't know about.
2. Estimate how many calls each makes a day. That number becomes a line on your bill.
3. Design agents to batch their work: one call per task, not one per step.

How to get started without wasting a year

Pick one boring job, give the agent read-only access first, keep a human approving, and measure hours saved.

1. Choose the first job carefully. Frequent, rule-based, digital, low-risk. "Answer order-status questions" is good. "Negotiate renewals" is not, yet.
2. Write the job description. If a new hire couldn't follow it, an agent can't either.
3. Fix the data the agent will rely on. Duplicates and stale records turn into wrong answers at scale.
4. Start read-only, with a human in the loop. The agent drafts; a person approves.
5. Promote it slowly. Once it's been right for a few weeks, let it act on its own in that one area.
6. Measure it: hours saved, response time, first-contact resolution.

I run my own business this way. My AI "COO", "CRO" and "CFO" agents sort my inbox, draft proposals and check the books every morning, and nothing is sent, paid or changed without my approval.

Frequently asked questions

What is Agentforce in simple terms?

Agentforce is Salesforce's platform for AI agents that can understand a request, plan the steps and take actions in your Salesforce data, such as updating records or routing cases, with people in charge.

Is Agentforce the same as SIforce?

No. Agentforce is the platform for building agents. SIforce (announced as AIforce at Dreamforce 2026) is about using Salesforce from wherever you work, including inside Claude and Slack, with Agentforce agents doing the work underneath.

How much does Agentforce cost?

Most usage is paid with Flex Credits: at list price, each agent action uses 20 credits, about $0.10. Salesforce plans per-call metering for registered agents, with 30 days' notice (reported target: November 2026). Check your own contract and the current pricing page.

What's a good first Agentforce use case?

Something frequent, rule-based and low-risk, such as answering common service questions, checking order status or routing cases, where a mistake is easy to catch and the time saved is easy to measure.

Do I need Data 360 (formerly Data Cloud) for Agentforce?

Agents are only as good as the data they're grounded in. Many first use cases work on core CRM data; Data 360 (formerly Data Cloud) matters when the agent needs a unified view across systems.