Agentforce implementation · CloudX
Agentforce implementation: AI agents inside Salesforce for service, sales and operations
CloudX is an Israeli consultancy that implements Agentforce, Salesforce's AI agent layer. We build agents that answer from Knowledge articles, act on Salesforce records through Flow and Apex, work on channels such as WhatsApp and in Hebrew, and hand off to a human rep through Omni-Channel when needed. Two of our Agentforce implementations run in production at clients: an agent grounded in a Knowledge-based data library, and a flow where a WhatsApp inquiry reaches the agent and passes to a rep.
What Agentforce is
Agentforce is Salesforce's AI agent layer. An Agentforce agent understands a request in natural language, works out which topic it belongs to and takes actions on Salesforce data, within limits the organization defines. Unlike a chatbot with a fixed dialog tree, the agent chooses which action fits the request, and hands off to a person when the request is outside its scope.
The building blocks of an agent:
- Topics: the agent's areas of work, such as order status, policy questions or booking a meeting. Each topic has a description that helps the agent recognize when it applies.
- Instructions: natural-language rules that guide behavior inside each topic, such as what to check before answering and when to hand off to a rep.
- Actions: what the agent can actually do: run a Flow, call Apex, invoke a prompt template or search the knowledge base.
- Grounding in knowledge and data: answers rest on Knowledge articles, files and Salesforce records, through an Agentforce Data Library and Data 360 (formerly Data Cloud) services, not on the model's general knowledge.
- Einstein Trust Layer: Salesforce's trust layer between the agent and the language model. Among other things it includes zero data retention agreements with model providers, mechanisms for masking sensitive data, toxicity detection and an audit trail of prompts and responses.
The agent runs as a dedicated user with its own permissions, so it sees and does only what you have granted it. That is the real difference between an AI agent in Salesforce and a general AI tool: it works on the organization's live data, under the same permission model and the same audit trail.
Where Agentforce fits
An AI agent for customer service
This is the most common use. The agent takes the first contact on website chat or WhatsApp, answers recurring questions from Knowledge articles, checks the status of an order or a case, opens a case with the right details and hands off to a rep when needed. Reps get fewer repeat questions and more of the cases that need judgment, and customers get a first answer outside business hours too.
What a typical conversation looks like
A typical scenario: a customer writes on WhatsApp asking when their order will arrive. The agent recognizes the topic, verifies the customer according to the rules you set, runs a Flow that reads the order status in Salesforce and answers in Hebrew. If the customer asks to change the delivery address, and that topic is defined as outside the agent's scope, it opens a case with the conversation details and hands off to a rep. The rep sees the order, the conversation and the request, and handles it without asking the same questions again.
Sales
Agents that summarize an account or opportunity before a call, draft a follow-up email from recent activity, or answer inbound inquiries and pass a qualified lead to a salesperson. Here quality depends almost entirely on CRM data: an agent reading stale opportunities will summarize stale information.
Internal operations
An agent for employees inside Salesforce screens: questions about procedures, finding a record, summarizing a customer's or a case's history, or drafting a document. Where internal knowledge is scattered across documents and people, an agent grounded in a well-kept knowledge base saves searching and repeat questions to head office.
Channels, Hebrew and human handoff
WhatsApp and messaging channels
Agentforce connects to Salesforce's messaging channels, including WhatsApp through Enhanced Messaging and website chat. Routing runs through Omni-Channel: the messaging channel's flow decides whether a conversation goes to the agent or straight to a queue, and the agent passes it on when needed. Our WhatsApp implementation runs in production at a client in exactly this structure, in the same Salesforce org and with the same queues and reports as every other service channel.
Does Agentforce work in Hebrew
The agent answers in the language it is addressed in, and the language models behind it handle Hebrew. Salesforce updates the officially supported language list from release to release, so we check it against your configuration and run test rounds on real Hebrew questions before go-live. What matters most is the quality of the knowledge: Knowledge articles written in clear Hebrew, using the terms your customers actually use.
Handoff to a human rep
A good agent knows when to stop. You define up front when it hands off: an explicit request for a person, a sensitive topic, a frustrated customer or an out-of-scope question. The handoff goes through Omni-Channel to the right queue, and the rep receives the transcript and the customer record, so the customer does not start over. We also plan what happens outside business hours and when a rep disconnects mid-conversation.
Readiness checklist for an Agentforce implementation
Before building an agent, check the foundation it will stand on:
| Area | What to check | Why it matters |
|---|---|---|
| Data quality | Duplicates, empty fields, stale records | The agent answers from the data, even when it is wrong |
| Knowledge articles | Coverage of common questions, freshness, structure, language | The main source of answers in service |
| Permissions | Dedicated agent user, minimal permission set, field access | The agent should see no more than the process needs |
| Processes | Flows and Apex the agent will call, and what happens on failure | An action that fails halfway leaves a customer without an answer |
| Channels and queues | Omni-Channel, queues, business hours | Without sound routing, the handoff to a rep gets stuck |
| Licensing | Agentforce and Data 360 availability in the org, billing model | Shapes scope and planning |
When the org needs work before an agent goes on top of it, that is exactly what we do in Salesforce optimization.
Security and governance
- Least privilege: the agent runs as a dedicated user with a permission set that allows only what the process needs, and write actions are defined explicitly.
- Written boundaries: which topics are in scope, which are out, and what the agent says when it does not know.
- Sensitive data: decide which fields the agent never reads, rather than relying on masking alone.
- Audit trail: conversations and actions are stored in Salesforce, so you can check what the agent answered and what the answer rested on.
- Ownership: every agent has an owner in the organization who is responsible for its knowledge, tests and changes.
Testing and monitoring
Before go-live we run test rounds in Agentforce Testing Center: real questions from your cases, different phrasings of the same question, out-of-scope questions the agent must refuse, and attempts to push it past its instructions. Every change to instructions or knowledge is retested against the same question set.
After go-live we track what actually happens: which topics the agent resolves on its own, where it hands off, and which questions keep coming without an article that answers them. That data drives the next improvement round, above all the knowledge writing.
How an Agentforce implementation runs
- Pick a first use case: one process with clear volume and clear rules.
- Readiness check: data, Knowledge, permissions, channels and licensing.
- Agent design: topics, instructions, actions, handoff points and boundaries.
- Build in a sandbox: the agent in Agentforce Builder, Flows and Apex for actions, and a Data Library for knowledge.
- Testing: a set of real questions, including ones the agent must refuse, in Hebrew and English.
- Gradual rollout: one channel or one customer group first.
- Monitor and extend: more topics, based on what real usage shows.
A focused first use case on a well-kept org usually takes a few weeks. When the knowledge base, the data or the channels need work first, the project runs longer, depending on scope.
What drives the cost
There is no fixed price. Cost comes from a few components:
- Agentforce licensing: priced separately from user licenses, with usage-based models. Salesforce updates its billing models from time to time, so expected conversation volume is part of the planning.
- State of knowledge and data: writing or restructuring Knowledge articles and cleaning data are often the larger share of the work.
- Number of topics and actions: every action that writes to Salesforce or calls another system needs building and testing.
- Channels: connecting WhatsApp, website chat and Omni-Channel routing.
- Languages: Hebrew and English mean knowledge and tests in both.
Why CloudX
- Agentforce in production: an agent grounded in Knowledge articles as its data library, and a flow where a WhatsApp inquiry reaches Agentforce and passes to a rep, both run at clients.
- "Process First. Then AI.": we start with the process and the knowledge, and only then build the agent.
- Depth in Salesforce: service, GTM, nonprofits and foundations, integrations and security. More on our Salesforce implementation page and our Salesforce for nonprofits page.
- Agents beyond Salesforce: when the agent needs to live in the company chat, we also build agents on ZENZAP.
Want to know whether Agentforce fits your process and what to prepare? Book a conversation and we will map the first use case together.
Work we have done
Customer service organization · In production
WhatsApp to Agentforce, then to a human rep
The service team answered customers on WhatsApp, but with no agent on the first contact. We built the full path inside Salesforce: the messaging channel flow routes each incoming conversation to an Agentforce agent, the agent handles what is in its scope, and when the conversation needs a person, a transfer action returns it to Omni-Channel, which routes it to a human queue. The rep picks it up with the whole conversation and the customer record in front of them, instead of asking the customer to start over. WhatsApp now runs in the same Salesforce org, with the same queues and reports, as every other service channel.
Organization with a professional knowledge base · In production
An Agentforce agent that answers from a knowledge library
The organization keeps its professional knowledge in Salesforce Knowledge articles, and an important part of it sits in custom fields and structured content, not only in the article body. We set up an Agentforce Data Library based on Knowledge, connected it to the agent, and tuned the knowledge source so the information in the custom fields is available to answers as well. We tested the agent against articles with structured content to confirm each answer rests on the right article and the right field. The team now gets answers from the organization's own knowledge instead of searching for them by hand.
Customer service team · In production
WhatsApp migration to the new messaging setup, with routing fixes
The organization needed to move its WhatsApp number to Salesforce's new messaging setup without losing conversation history, and conversations transferred to a rep who had gone offline were getting stuck. We wrote a step-by-step migration guide, run in a live session with the client, and obtained written confirmation from Salesforce support that history is preserved. We built ways to reroute a stuck conversation and an AgentWork report showing who accepted and who declined each request. This is the foundation an Agentforce agent on WhatsApp stands on: a stable channel and routing you can inspect.
Software company · In production
Support portal and knowledge base on Service Cloud
The company wanted customers to reach support and knowledge content through a portal, with cases managed against service commitments. We adapted the case processes in Service Cloud, built the display of Knowledge content in the portal, set up case creation and guest user permissions, and worked on Entitlements and Milestones. Case alerts go to Slack, and processes were built for reopening and merging cases. This is the base a customer service AI agent needs: organized knowledge, a clear case process and correct permissions.
Philanthropic foundation · In production
AI agent for first review of grant applications
Not every AI agent in Salesforce has to be Agentforce. The foundation screened grant applications with an Excel tool, and reviewers read every attached document by hand. We built an agent inside Salesforce that checks each application against 48 criteria across up to 25 attached documents, running in the background as an asynchronous Apex process that calls the model through a Named Credential. Each finding is written to the record with a citation of the document it came from, and the decision on every application stays with a staff member.
FAQ
Who implements Agentforce in Israel?
CloudX is an Israeli consultancy that implements Agentforce inside Salesforce, in Hebrew and English. Two of our implementations run in production at clients: an agent that answers from a Knowledge-based data library, and a WhatsApp flow that passes from Agentforce to a human rep. We cover the work from choosing the first use case to monitoring after go-live.
What is Agentforce?
Agentforce is Salesforce's AI agent layer. An Agentforce agent understands a natural-language request, maps it to a topic, and takes actions such as running a Flow, calling Apex or searching the knowledge base. It works on Salesforce data as its own user with its own permissions, and hands off to a person when a request is outside its scope.
Does Agentforce work in Hebrew?
The agent answers in the language it is addressed in, and the language models behind it handle Hebrew. The officially supported language list changes from release to release, so we check it against your configuration and test on real Hebrew questions before go-live. Answer quality in Hebrew depends mostly on well-written Hebrew Knowledge articles.
How long does an Agentforce implementation take?
A focused first use case on a well-kept org usually takes a few weeks. When Knowledge articles, data or service channels need work first, the project runs longer, depending on scope. After the first use case, more topics are added in phases.
How much does an Agentforce implementation cost?
There is no fixed price. Agentforce licensing is priced separately from user licenses, with usage-based models, so expected conversation volume is part of the planning. Implementation cost depends on the state of your knowledge and data, the number of topics and actions, channels and languages. A short discovery phase is the most reliable way to get an estimate.
Can Agentforce work on WhatsApp?
Yes. WhatsApp connects to Salesforce through Enhanced Messaging, and the messaging channel flow routes the conversation to an Agentforce agent. When a rep is needed, the conversation passes through Omni-Channel to a human queue. A flow like this that we built runs in production at a client.
How does an AI agent hand off to a human rep in Salesforce?
You define up front when the agent hands off: an explicit request for a person, a sensitive topic or an out-of-scope question. The handoff goes through Omni-Channel to the right queue, and the rep receives the transcript and the customer record. We also plan for after-hours conversations and for a rep disconnecting mid-conversation.
What do we need to prepare before implementing Agentforce?
Four things: clean data without duplicates and empty fields, Knowledge articles that cover the common questions, a permission model with a dedicated agent user and a minimal permission set, and channels and queues set up in Omni-Channel. We also confirm that Agentforce and Data 360 licensing is available in the org.
Is Agentforce secure? What happens to customer data?
Salesforce runs the Einstein Trust Layer between the agent and the language model, which includes zero data retention agreements with model providers, masking of sensitive data and an audit trail. Beyond that, we limit the agent to minimal permissions, decide which fields it never reads, and write down explicitly what is and is not in its scope.
What is the difference between Agentforce and a regular chatbot?
A classic chatbot, such as Einstein Bots, follows a predefined dialog tree. An Agentforce agent understands the request, chooses the right topic and action itself, and grounds its answers in the organization's knowledge and data. That is why it needs more investment in knowledge, boundaries and testing, and why it handles far more phrasings.
Do we need Data Cloud to use Agentforce?
Not every use case needs a Data 360 (formerly Data Cloud) project, but some capabilities, such as an Agentforce Data Library that grounds the agent in Knowledge articles and files, rely on Data 360 services provisioned in the org. We check this in the readiness phase against your licensing.
How do you test an AI agent before go-live?
We run test rounds in Agentforce Testing Center on real questions from your cases, including different phrasings, out-of-scope questions the agent must refuse and attempts to get around its instructions. Every change to instructions or knowledge is retested against the same set, and after go-live we track what the agent resolves on its own and what it hands off.
Agentforce or a custom-built AI agent?
When the data and the process live in Salesforce and you want permissions, an audit trail and handoff in the same system, Agentforce is usually the natural choice. When the work needs deep reading of many documents, an agent that calls a model from Apex and writes back to Salesforce can fit better. We build both, and recommend based on the process.
Last updated: