Insights / AI agents
AI agents vs chatbots: what is the difference, and which does a small business need?
· 6 min read
What is the difference between an AI agent and a chatbot?
A chatbot answers questions in a conversation; an AI agent takes actions — it reads a message or an image, decides what it is, and updates a system of record. Most small businesses need a chatbot first, for questions about price, availability and scope, and an agent only where a person currently retypes information from one place into another. A chatbot build is $1,199; a single workflow automation is $1,000.
What is the practical difference?
| Chatbot | AI agent | |
|---|---|---|
| What it does | Answers questions | Changes something — a record, a ledger, a pipeline |
| Input | A visitor's question | Emails, WhatsApp messages, images, documents |
| Output | A reply | A reply plus an action |
| Main risk | Inventing an answer | Taking a wrong action |
| Guardrail | Grounded in your content, allowed to say “I don't know” | A person approves anything costly or irreversible |
| Our price | $1,199 build, $150/mo hosting | $1,000 per workflow; custom from $4,000 |
Do you always need a language model?
No. The helper on our student-housing site matches keywords against a hand-written knowledge base in English and Hinglish. It is not a language model, it costs nothing per message, and for the twenty questions students actually ask — price, distance to college, food, rules — it is enough. That site now earns 577 search clicks a quarter and is named by ChatGPT in 57 of 78 student questions; the helper did not need to be clever for that to work.
When is an agent worth it?
When someone is copying information by hand. Two agents we run, and the exact work each removes:
- An export enquiry desk: 1,824 enquiries from email and WhatsApp since April 2026, deduplicated, scored for genuineness by a model, and moved through stages — 482 contacted, 67 negotiating. The model drafts replies on request; a person sends them.
- A WhatsApp property manager for a 61-bed residence: residents send a UPI payment screenshot and a vision model reads the amount, reference and time, then matches it to the rent ledger. Meter photos are read the same way.
Neither of them is impressive in a demo. Both remove an hour of retyping a day, which is the test that matters.
How do you stop an agent from doing something wrong?
Decide in advance which actions it may take alone. Reading, classifying and drafting are safe to automate. Sending a quote, changing a price or issuing a refund is not — those stay behind a person. Anything the agent cannot match confidently is flagged, never guessed. A wrong reply costs an apology; a wrong number in a quote costs the margin on the order.
Which should you start with?
- Start with a chatbot if visitors ask the same pricing and availability questions and leave when nobody answers.
- Start with an agent if enquiries arrive in several places and nobody can say how many came in last week.
- Start with neither if you have under a handful of enquiries a week — fix the website and its search visibility first.