Five Types of AI Agent, and How to Choose the Right One for Your Business
Not every AI agent works the same way. Some answer questions, some act on their own when something changes, and some work as a team. Here are the five types Salesforce's architects describe, what each is good for, and where most businesses should start.
The short version
Salesforce's architecture team has published a guide that sorts AI agents into five types. The difference between them is what starts the agent, and how much it does on its own.
Conversational agents answer when someone asks. Proactive agents act when something happens. Ambient agents work quietly in the background. Autonomous agents take a goal and plan the work themselves. Collaborative agents are teams of specialists working on one job.
Most businesses should start with the first two. They are easier to build and measure, and they put your data and guardrails in place for the more independent types later. Below, each type is explained with our own diagram and an example from everyday work.
1. Conversational agents: they answer when asked
This is the agent most people picture: a customer or employee asks a question in plain language, by text or voice, and the agent replies. What makes it more than a chatbot is that it checks your real data before it answers, and it knows when to hand over to a person.
Example: a customer asks on WhatsApp where their order is. The agent finds the order in your CRM, checks the courier's status, and replies with a delivery date. If the parcel is lost, it passes the chat to your service team with the history attached, so the customer does not have to repeat themselves.
High volumes of similar questions, 24/7 cover, and first replies to new leads by email or chat.
Answers are only as good as your knowledge articles, so keep them current. Decide exactly when the agent should hand over, and test it.
2. Proactive agents: they act when something happens
Proactive agents are started by an event, not a person. That can be a change in your CRM, such as a deal that has gone quiet, or a signal from another system, such as a failed payment. The agent decides whether the event matters and what should happen next.
Example: a high-value customer leaves items in their online cart. The agent alerts the account manager in Slack, creates a follow-up task, and adds the customer to a reminder journey with a small offer.
Following up on stalled deals, ageing cases, abandoned carts, renewals and anything else that currently relies on someone noticing.
Too many triggers create noise, so start with a few events that clearly matter. Events from outside systems can arrive late or twice, so the agent must handle both.
3. Ambient agents: they work in the background
Ambient agents sit alongside the work people are already doing, such as a sales call, a meeting or a record on screen. They listen or watch, and step in only when they can help. Most of the time, nobody sees them.
Example: during a discovery call, the agent follows the conversation, finds the product details the customer asks about, and sends private tips to the sales rep in Slack. After the call, it writes up the notes and updates the opportunity.
Coaching reps during calls, capturing meeting notes, and guiding people through long processes in Salesforce.
Recording and listening need consent and a clear privacy notice, in line with Malaysia's Personal Data Protection Act. Too many suggestions get ignored, so tune what the agent surfaces.
4. Autonomous agents: they own a goal
An autonomous agent is given a goal and limits, not step-by-step instructions. It plans the steps, remembers what it has learned, uses your tools, checks whether each step worked, and changes the plan if it did not.
Example: a sales manager sets a target for new pipeline this quarter. The agent researches target accounts, qualifies leads, drafts outreach for the team to send, and reports progress every week.
Long-running work with clear rules, where the steps change from case to case and constant supervision would be too slow.
They take the most effort to build, and it is harder to step in halfway. Mistakes can travel further before anyone notices. Start with tight limits and a person approving key steps.
5. Collaborative agents: a team of specialists
Some jobs are too broad for one agent. In a collaborative setup, an orchestrator agent splits the job and hands each part to a specialist agent, then combines what comes back. People stay in the loop to approve the result.
The agents talk to each other through open standards. Agent2Agent (A2A) lets agents pass work between them, even across vendors, and the Model Context Protocol (MCP) lets them use tools and data in other systems.
Example: a customer complains about a wrong bill, a late delivery and a service cut-off in one message. The orchestrator asks the billing, delivery and service agents to check their part, combines the findings into one proposed fix, and a service rep approves it before it goes to the customer.
Cross-team problems with handoffs and approvals, where the knowledge sits in different departments or systems.
More agents mean more to secure, monitor and pay for. Each specialist needs clear boundaries, and the hand-offs need to be logged.
A closer look: how a messaging agent is built
Salesforce's guide also shows how each type is put together. The most common is the messaging service agent, a conversational agent on channels such as WhatsApp, web chat, SMS and LINE. Its set-up is a good picture of what every agent needs.
The guide describes three basic recipes for this pattern. The agent answers from your knowledge, takes an action such as updating a booking, or escalates to a person. Most real deployments use all three, and voice is next on the list.
Which type should you start with?
Start with the problem, not the type of agent. This is how the two usually match up.
| If you want to… | Start with |
|---|---|
| Answer common customer or staff questions, any time | A conversational agent |
| Stop things slipping through the cracks | A proactive agent |
| Help reps during calls and cut admin after them | An ambient agent |
| Hand over a whole piece of work with a clear goal | An autonomous agent, once the basics are proven |
| Solve problems that cross several teams or systems | Collaborative agents, built on the ones above |
What every agent needs, whatever the type
Trusted data. Agents answer from your customer records and knowledge, so these need to be connected and current. In Salesforce, that usually means Data 360 and well-kept knowledge articles.
Guardrails. Each agent needs a clear scope, the minimum permissions it needs, and protection for sensitive data. Salesforce's Einstein Trust Layer masks personal data before it reaches the AI model.
A clear route to a person. Decide when the agent must hand over, and make sure the person gets the full context.
Monitoring. Keep logs of what each agent did, review where it escalated or got stuck, and adjust its instructions over time.
A cost and limits plan. Agents use credits and system capacity, so set limits early. That way, one busy agent does not slow everything else down.
Where to start this quarter
Source: Salesforce Architects, Agentic Patterns and Implementation with Agentforce. Diagrams redrawn by Telcowin. Product names and capabilities as described by Salesforce and may change.
Not sure which agent to build first?
We will look at your processes and data, and help you pick the first agent worth piloting, with the guardrails in place from day one.