AI agent vs chatbot: difference, price, when to choose what

Andrew Maryasov, AI consultant. 28 AI projects across 14 industries — from simple bots to multi-agent systems. This guide answers the question that comes up on discovery sessions more often than any other: “do we need a bot or an agent?”

TL;DR (in 30 seconds)

• A chatbot is a script. A limited set of rules, predictable answers. Cheap ($500–3,000), fast (2–3 weeks), scales poorly.
• An AI agent is a system that makes decisions. It has tools, memory, plans. More expensive ($4,000–25,000), takes longer (4–12 weeks), scales together with the business.
• Rule of thumb: if the number of scenarios is < 20 and they’re stable — chatbot. If > 20 or they change every month — agent.

What a chatbot is (in 2026)

A chatbot is a program that responds to messages following a pre-defined script. Modern chatbots are built on one of three approaches:

1. Rule-based

The classic approach. The user taps buttons → the bot performs an action. Like an ATM.

Platforms: ManyChat, BotHelp, Salebot, Chatfuel. Strength: 100% predictable, cheap, fast. Weakness: any deviation from the script — “didn’t understand”.

2. NLU-based (intent recognition)

The bot classifies intents and responds with a template tied to a specific intent. Like “question about price → answer A”, “question about delivery → answer B”.

Platforms: Dialogflow, Rasa, IBM Watson Assistant. Strength: a bit more flexible than rule-based, handles variations in phrasing. Weakness: requires labeling hundreds of examples; a new intent = a new training cycle.

3. LLM wrapper (the new wave of 2024–2026)

Bot = LLM (GPT-4/Claude) + system prompt. Input — the user’s message, output — a generated reply.

Platforms: Voiceflow, custom GPTs, Bot9. Strength: looks “smart”, easy to tune to a tone of voice. Weakness: no tools. Can’t add a client to the CRM, can’t check stock. Text only.

What an AI agent is

An AI agent is an LLM with tools, memory, and autonomous decision-making. It doesn’t just answer — it acts.

Key components:

1. Brain (LLM)

A large language model — Claude Sonnet/Opus, GPT-4, Gemini. Responsible for reasoning and choosing the next step.

2. Tools

Functions the agent can call: search_catalog(), book_appointment(), send_email(), query_crm(). This is the agent’s body.

3. Memory

What the agent remembers between messages. Can be:

• Short-term (the context of the current conversation).
• Long-term (previous interactions with the user — vector DB).
• Project (a knowledge base about the product, pricing, FAQs).

4. Planning

In complex scenarios the agent breaks the task into steps: “first I’ll check stock → then suggest an alternative → then book a slot”.

5. Multi-step execution (agentic loops)

The agent can perform 5–10 actions in a row without human intervention: write → check → fix → send.

Comparison table

Criterion	Chatbot	AI agent
Logic	Script	Decision-making
Tools	Limited (usually replies only)	API, databases, external services
Memory	Session only	Long-term + context
Conversation complexity	Linear	Branching, recovery
Implementation time	2–3 weeks	4–12 weeks
Price (Ukraine)	$500–3,000	$4,000–25,000+
Monthly costs	$30–150	$80–500 (LLM API + infra)
Who it fits	FAQ, simple booking	Sales, support, complex flows
Scaling	Linear (more scenarios = more work)	Improves over time
Risks	”I don’t understand” on deviations	Hallucinations, tool errors

When to choose a chatbot

If your case meets all of these conditions:

• ☑ The number of unique scenarios is < 20.
• ☑ The scenarios are stable (don’t change every month).
• ☑ No need to integrate with more than 1–2 systems.
• ☑ Budget up to $3,000.
• ☑ MVP in 2–3 weeks.

Typical cases:

1. Booking a haircut at a barbershop. Why an agent here — there are slots, services, and barbers. A clear script.
2. FAQ for a small online store. “How much is delivery”, “what are your hours”, “do you accept exchanges” — the list is fixed.
3. Lead capture on a landing page. Question → form → CRM. Straightforward.

When to choose an AI agent

If even one of these signs applies — you need an agent:

• ☑ More than 20 scenarios, or they change frequently.
• ☑ You need to make decisions (for example, product recommendations).
• ☑ Integration with 3+ systems (CRM, inventory, payment, email).
• ☑ Multi-step workflows (do reasoning → action → more reasoning).
• ☑ Persistence — the agent must remember previous interactions.
• ☑ Budget from $4,000 and ROI on a 3–12 month horizon.

Typical cases:

1. An AI stylist for a clothing store chain. Selecting products with the collection, season, and context in mind. Case study — upsells framed as a stylist’s advice, not “buy more”.
2. A sales agent for B2B SaaS. Lead qualification, scheduling demos, answering technical questions, escalation to a human.
3. Customer support for fintech. Checking transaction status in the database, refund processing, escalation on complex cases.
4. An HR agent. Screening candidates, scheduling interviews, answering questions about the company.

Real cost: how to calculate ROI

Two model calculations to show the logic (the numbers are typical for the market — plug in your own):

Model 1: a chatbot for a barbershop

Budget: ~$1,500 (off-the-shelf platform + 2 weeks of setup). Savings: −10 hours/week of admin work = ~$400/month. Payback: ~4 months.

Model 2: an AI agent for a clinic

Budget: from ~$5K (custom development + Claude API). Savings: less admin time on routine + bookings outside working hours that previously went to competitors. What the depth of working through such a task looks like — in the SmileClinic demo: there the calculation for a typical dental chain is laid out down to the hryvnia.

How to choose: a decision tree

Question 1: How many unique conversation scenarios?

• < 20 → step 2
• 20+ → AI agent

Question 2: Do you need to integrate with more than one system (CRM, database, payments)?

• Yes → AI agent
• No → step 3

Question 3: Does the agent need to make decisions (recommendations, escalations, taking orders)?

• Yes → AI agent
• No → step 4

Question 4: What’s the budget?

• < $3,000 → chatbot
• $4,000+ → AI agent

Common mistakes

1. Buying a chatbot when you need an agent

“We’ll set up ManyChat, then add ChatGPT.” It won’t work. A chatbot’s architecture doesn’t allow integrating tools, memory, planning. A year later you’ll rewrite it as an agent anyway — but now with time lost.

2. Buying an agent when a chatbot is enough

A more typical mistake among SMBs. “We want AI” → they spend $15K → then 6 months of bug-fixes. If you have 8 scenarios and simple booking — ManyChat for $50/mo will do the same thing.

3. Not counting LLM costs

An agent on Claude Sonnet 4.6 costs about $3–15 per 1,000 dialogues (depending on length). For 10,000 dialogues/mo = $30–150. If a chatbot for $50/mo covers the need — don’t overpay.

4. Not planning for human-in-the-loop

Any agent makes mistakes. In 2–5% of cases. Without escalation to a live manager, those 2–5% turn into lost clients and 1-star reviews.

FAQ

How is an AI agent different from ChatGPT?

ChatGPT is an interface to an LLM. You write → it answers. Without tools (although custom GPTs changed this somewhat).

An AI agent is a system with an LLM inside. It has tools, integrations, memory, planning. ChatGPT can become part of an agent (as the brain), but on its own it isn’t an agent.

Will AI agents replace chatbots completely?

No. A simple FAQ bot for $50/mo will remain the optimal choice for many cases. An agent is overkill for linear scenarios.

Forecast: by 2028 the line will blur — most “chatbots” will have minimal agentic capabilities (one or two tools), and agents will become simpler to deploy (low-code platforms).

How much does it cost to maintain an AI agent?

Monthly — from $80 (a light case) to $500 (high-traffic). It’s made up of: LLM API ($30–300), hosting ($10–50), CRM extensions ($20–100), monitoring ($10–50).

Plus engineering support — at least 5–10 hours/month in the first 6 months (finetuning prompts, updating tools, debug).

Can you build an agent yourself without a developer?

Since 2026 — partly yes. Platforms like Voiceflow, Vapi, Botpress let you assemble agents from blocks. Good for a prototype.

But for production-ready with real CRM/database integration — you still need a developer or a consultant. Or, at minimum, someone with a technical background who understands APIs.

Are AI agents safe?

With the right approach — yes. Basic practices:

• Input validation: don’t trust the LLM output 100%, verify it.
• Tool permissions: the agent shouldn’t have access to anything it doesn’t need.
• Audit log: all of the agent’s actions go into the logs.
• Rate limiting: preventing abuse.
• Human-in-the-loop: for critical actions (payments, deleting data).

What’s next

If you want to discuss what fits your business — a 30-minute discovery session is free.

I’m not an agency, I don’t “push” solutions. Half of those who book get the recommendation “you don’t need AI right now — do X first.” That’s fine.

See also:

• How to hire an AI consultant: a 12-question checklist — if you are in fact looking for someone to do this end-to-end.
• AI consulting in Ukraine 2026: reality and prices — about the market, cases, budgets.
• SmileClinic case study — a demo of an AI pipeline for dentistry: a full proposal package in 6 minutes.