The textbook definition
In artificial intelligence an agent is anything that perceives its environment through sensors and acts on it through actuators. The definition comes from Stuart Russell and Peter Norvig’s textbook Artificial Intelligence: A Modern Approach, used in most university AI courses. A thermostat is an agent: it senses temperature and switches the heater. A self-driving car is an agent. So is a WhatsApp bot that reads a message and sends a reply.
What makes an agent "intelligent" is that it chooses actions that move it towards a goal, rather than doing the same thing every time. The more varied the situations it can handle sensibly, the more intelligent it is.
What are the five types of intelligent agents?
This classic classification is what exam questions about "intelligent agent in artificial intelligence examples" usually expect. Each type adds one capability to the one before it.
| Type | How it decides | Classic example | Business example |
|---|---|---|---|
| Simple reflex agent | Fixed condition-action rules on what it sees right now | Thermostat; automatic door | Auto-reply that sends timings when a message contains "timing" |
| Model-based reflex agent | Rules plus an internal model of things it cannot see now | Robot vacuum that remembers which rooms are clean | Inventory alert that tracks stock between Tally syncs |
| Goal-based agent | Searches for actions that reach a goal | Route planner finding a way to the airport | Scheduler that finds a slot meeting doctor, room and patient constraints |
| Utility-based agent | Picks the action with the best score among many that reach the goal | Navigation choosing the fastest and cheapest route | Ad-budget agent shifting spend to the campaign with the best cost per lead |
| Learning agent | Improves its own behaviour from feedback | Spam filter that learns from what you mark as spam | Lead-scoring model that improves as sales marks leads won or lost |
Some courses add hierarchical agents (a manager agent directing sub-agents) and multi-agent systems (several agents working together or competing). Both are common in modern AI products.
How are modern LLM agents different?
The AI agents people talk about today are built on large language models such as GPT, Claude or Gemini. They fit the textbook definition, but the reasoning comes from the language model instead of hand-written rules or a search algorithm. That lets them work with messy inputs like a voice note in Hinglish or a scanned invoice.
A modern agent runs a loop. It reads the situation, decides which tool to use, calls it, reads the result and decides again. Tools might be "search the product catalogue", "create a lead in HubSpot", "send a WhatsApp template" or "check a calendar". In textbook terms most LLM agents are goal-based or utility-based agents, and some are learning agents when their instructions or examples are updated from feedback. When people say agentic AI, they mean systems built from these agents.
The parts of an LLM agent
- Model: the reasoning engine.
- Instructions: role, goal, rules, tone and limits.
- Tools: APIs it may call, each with a clear description.
- Knowledge: documents and data it can look up, often through retrieval (RAG).
- Memory: what happened earlier in this conversation and, sometimes, past ones.
- Guardrails: approvals, spend caps, blocked actions and a hand-over to a person.
12 examples of AI agents in business
Each line shows the agent’s goal, its main tools and the type it maps to. These are typical designs, not claims about specific clients.
- WhatsApp sales agent: answers product questions, sends a catalogue and books a call. Tools: WhatsApp Business API, CRM, calendar. Goal-based.
- AI receptionist: answers clinic or salon calls in Hindi or English and books appointments. Tools: telephony, booking software. Goal-based. See AI receptionists in India.
- Lead qualification agent: calls portal leads within minutes, asks budget and timeline, scores them. Utility-based.
- Collections agent: sends polite reminders with UPI links, stops when paid, escalates disputes. Model-based (tracks who has paid).
- Invoice-entry agent: reads supplier bills from email and drafts entries for Tally or Zoho Books for an accountant to approve. Goal-based.
- Support agent: answers order-status and return questions from Shopify data and opens tickets for the rest. Goal-based.
- Recruitment screener: reads CVs against a job description and shortlists with reasons for a recruiter to review. Utility-based.
- Daily MIS agent: pulls numbers from several systems each morning and sends a summary with exceptions highlighted. Model-based.
- Ad-performance agent: watches campaigns and flags or pauses ones that cross a cost-per-lead limit. Utility-based.
- Research agent: searches the web for prospects matching a profile and drafts first notes for a salesperson. Goal-based.
- Document-collection agent for CA firms: chases clients for missing documents before filing dates. Goal-based.
- Internal knowledge agent: answers staff questions from SOPs and policies, citing the source page. Retrieval-based; see RAG chatbots.
AI agent vs chatbot vs AI assistant
| Chatbot | AI assistant | AI agent | |
|---|---|---|---|
| Main job | Answer questions | Help one person with tasks when asked | Complete tasks towards a goal |
| Takes actions | Rarely | Sometimes, with approval | Yes, within set limits |
| Starts work by itself | No | No | Can, on a schedule or a trigger such as a new lead |
| Example | FAQ bot on a website | Copilot drafting an email for you | Agent that follows up every unpaid invoice |
How to decide if you need an AI agent
- The task repeats at least daily and follows a recognisable pattern.
- Inputs are messy (chats, calls, PDFs), so simple rules keep breaking.
- The tools involved have APIs or exports the agent can use.
- You can write down what a good outcome looks like and measure it.
- Mistakes are recoverable, or a person can approve the risky step.
- Someone in your team will own it and review it weekly.
If most boxes are unticked, a plain workflow or a better spreadsheet will serve you better than an agent.
Single agent or multi-agent system?
Start with one agent. A single agent with five well-described tools is easier to test, cheaper to run and simpler to debug than a team of agents passing work to each other. Multi-agent designs make sense when tasks are genuinely separate and large, for example one agent that researches prospects and another that writes outreach, each with its own instructions and checks. Even then, a fixed workflow often connects them better than letting agents talk freely, because you can see exactly where a job failed.
How do you build your first AI agent?
- Choose the task. One job, done often, with a clear finish line, such as "reply to new enquiries within five minutes and book a call".
- Write the playbook a new employee would get: what to say, what to ask, what never to promise, when to call the manager.
- List the tools and check each has an API or a connector in n8n, Make or Zapier.
- Pick the channel: WhatsApp, website chat, phone or email. Each has its own costs and rules.
- Build in drafting mode first, so a person approves each action for a week or two.
- Measure response time, completion rate, hand-overs and errors, then widen what it may do alone.
Most of the effort goes into steps two and five, not the AI itself. A clear playbook and honest testing decide whether the agent is useful. For cost ranges by agent type, see AI agent development cost in India.
How Silvane can help
Silvane designs and builds custom AI agents, WhatsApp agents and AI calling agents for Indian businesses, inside your own accounts. A ₹199 discovery call and AI audit for founding members (normally ₹15,000) tells you which agent to build first and what it will cost. Builds start from ₹50,000, run plans from ₹15,000 a month, with tools at cost and GST extra.
Sources: Intelligent agent (Wikipedia, citing Russell and Norvig) · Anthropic: Building effective agents