"We need a bot" — half of all automation enquiries start with that sentence. Then it turns out that one client genuinely needs a $300 bot and another needs a $900 AI agent, and confusing the two costs either money or results. At SUDUS·TEAM we build both Telegram bots and AI agents, so we have no interest in selling you one over the other.
Let us go through it honestly: how they differ technically, what each costs, where a bot beats an agent and where it loses hopelessly. And most importantly — how to work out which one you need before you have paid for it.
The technical difference in one paragraph
A Telegram bot follows a script. A developer mapped out the tree in advance: press "Catalogue" and it shows categories, press "Delivery" and it prints the delivery text. The bot does not understand language; it recognises commands and button presses. One step off the script and the user gets "Sorry, please choose from the menu".
An AI agent runs on a large language model. It understands a question however it is phrased, finds the answer in your knowledge base through RAG, can call functions (check stock, create an enquiry, book a slot) and composes the reply itself. There is no script tree — there are instructions, data and a set of tools.
Put simply: a bot is a menu in an app; an agent is a colleague who has read your documentation.
Seven differences that decide it
1. Handling unusual requests
Bot: cannot. "Do you have something similar but cheaper and available by Friday" flattens any script tree.
Agent: that is exactly what it is for. It understands the intent, asks for the missing details, offers options.
2. Build cost
Bot: from $300, more with complex scripts and integrations.
Agent: from $900 for a single-purpose agent, from $3,000 for a multi-agent system.
3. Running cost
Bot: close to nothing — $5–15 a month for hosting. No tokens, no per-conversation charge.
Agent: $140–200 a month for a small business, including model usage, infrastructure and maintenance. The detailed arithmetic is in how much an AI agent really costs.
4. Predictability
Bot: total. It says exactly what you wrote and never invents anything.
Agent: high but not absolute. It needs guardrails, RAG on your own data and quality control to rule out invention.
5. Time to launch
Bot: 10–25 days.
Agent: 14–25 days for a simple one, 35–60 for multi-agent.
6. Ability to act, not just answer
Roughly equal here, with a caveat. A bot can also create an order or book an appointment — but only along the path you wrote. An agent decides for itself which tool to call for a given situation, and can combine several actions in one conversation.
7. Maintenance and growth
Bot: adding a new topic means writing a new branch of the script — developer work.
Agent: adding a new topic means putting a document in the knowledge base. Your content manager can do that.
When a bot is enough — and there is no need to overpay
A bot covers the task completely when there is a clear, finite set of scenarios:
- Booking a service. Pick the specialist, the date, the time, confirm. No creativity required at all.
- Taking enquiries. Name, phone, a line about the problem, pass to a manager.
- Notifications and reminders. Order status, appointment reminders, broadcasts.
- A catalogue with a fixed structure. Categories, products, an "order" button.
- Internal team processes. Holiday requests, shift reports, opening checklists.
The sign that a bot is enough: you can draw every possible user path on one sheet of paper. If you can, there is no point paying for an LLM.
When you genuinely need an AI agent
- The questions are many and varied. If support sees 200 different phrasings a month, no script tree will cover them.
- There is a large knowledge base. Documentation, manuals, contracts, technical specifications. An agent reads it; a bot does not.
- You need qualification, not a form. Sales, where the point is to understand the need rather than force someone through seven fields.
- Customers phrase things loosely. "Something as a present for a man who has everything, budget up to 200" — a bot is helpless here.
- Documents are involved. Read an invoice, extract the details, compare a contract against a template.
The hybrid: what we recommend most often
In practice the best ratio of cost to result comes not from choosing but from combining. The bot takes the structured part — menu, buttons, booking, payment, notifications. The AI agent joins in inside the same bot wherever the user writes free text.
It looks like this: someone opens Telegram and sees the familiar menu and buttons. They press "Book" and the script runs, fast and with zero token cost. They type "can I move it to another day, I've come down with something" and the agent picks the question up, understands the intent, finds the booking and offers free slots.
The economics are pleasant: 80–90% of enquiries are handled by the scripted side at no cost, and only 10–20% reach the model. The token bill drops five- to tenfold compared with "everything through AI", and the user notices no difference.
Three cases from practice
Case 1: a dental clinic that asked for "AI for bookings"
The clinic wanted an AI agent because the receptionist could not keep up with messages. We went through the last 300 enquiries: 268 of them were "when is there a free slot", "how much is a cleaning", "where are you". Textbook script material.
We built a bot with calendar integration. It came out four times cheaper than an AI agent and closed 89% of enquiries. The remaining 11% — unusual questions about specific diagnoses — should not have gone to a machine anyway: in a medical context the cost of an error is too high.
Case 2: an equipment shop with 4,000 products
The opposite story. Customers wrote "I need a pump for a 40-metre borehole, water with sand in it" — no menu handles that. The technical specifications lived in manufacturers' PDF catalogues.
A scripted bot is helpless here in principle. We built an AI agent with RAG over the catalogues: it understands the task, picks two or three models and explains the difference. The same pattern works well in online stores with a wide technical range, where the customer does not know the name of the thing they need.
Case 3: a service centre where the hybrid was right
Half the enquiries were "what is the status of my repair" (pure script, all it needs is an order number), half were "my phone dies while charging, is that expensive to fix?" (needs understanding).
We split them: status and booking on buttons, diagnosis and cost estimates on the agent. The token bill came out three times lower than if everything had gone through the model, because the most frequent enquiries never touch it.
How the hybrid works technically
So it does not sound abstract, here is how traffic is routed in such a system:
- The user pressed a menu button — the script runs, the model is not called at all. Zero cost.
- The user typed text — a fast classifier fires: does this look like one of the known commands or not?
- It looks like a command — the matching script is used. "I'd like to book" leads to the same calendar as the button.
- It does not — the request goes to the AI agent with access to the knowledge base and the tools.
- The agent is not confident — it hands the conversation to a human along with the whole context.
The key point is that step four should be rare. If more than 30% of traffic reaches the agent, the scripted side is underbuilt and the bill will be higher than it needs to be.
First-year cost compared
Take a real task: customer support for a small online store, about 1,200 enquiries a month.
Bot only
Build from $450 with calendar and CRM integration, running cost about $10 a month. Year one: from about $570. Covers roughly 60% of enquiries; the rest go to a person.
AI agent only
Build from $900, running cost about $140 a month. Year one: from about $2,580. Covers 85–90%.
Hybrid
Build from $1,200, running cost about $50 a month. Year one: from about $1,800. Covers the same 85–90%, because the hard questions still reach the model.
The hybrid costs more than a bot and less than a pure agent for the same result. That is why we offer it as the starting configuration to most small and mid-sized clients.
What to choose by type of business
A generalisation across the niches we have worked in most. Treat it as a pointer, not a rule — the final answer comes from analysing your real enquiries.
Salons, dental practices, booking-led services
Bot. 85–90% of enquiries are bookings, reschedules, prices, address. All covered by buttons and a calendar integration. AI is worth adding only once there are several locations and real complexity in matching a specialist.
Online stores with a simple range
Bot or hybrid. Clothing, cosmetics, homeware: the catalogue is structured, filters work, most questions are about delivery and availability. A hybrid earns its place if you get a lot of "help me pick something for my situation".
Technical goods and B2B suppliers
AI agent. Thousands of SKUs, specifications in PDFs, customers describing a problem rather than a product name. A menu cannot be built for this in principle, and this is where an agent adds the most conversion in an online store.
Services with a long sales cycle
AI agent. Construction, equipment, consulting, IT. You need to understand the task, give a sense of budget and timeline, and gather the inputs for a manager. A seven-field form at this stage simply scares people off.
Companies with a lot of internal documentation
An AI agent for internal use. Policies, orders, manuals, technical conditions. An employee asks in plain language and gets an answer with a link to the source instead of searching through folders.
Logistics, delivery, statuses
Bot. A status lookup by reference number is an API call, not a job for a language model. There is nothing here worth paying tokens for.
A five-minute self-test
- Write out the last 20 real customer enquiries — verbatim, as they were phrased.
- Mark the ones a menu button could answer.
- If that is 18 out of 20 — you need a bot.
- If it is 10–15 out of 20 — you need a hybrid.
- If it is fewer than 8 — you need a full AI agent.
The exercise takes half an hour and saves you from the most expensive mistake in automation: buying a complex solution for a simple job, or a simple one for a complex job.
Frequently asked questions
Can we start with a bot and add AI later?
Yes, and that is the right route. If the bot is designed from the start with a future AI layer in mind, adding the agent later is an extension rather than a rewrite. Tell your contractor at the outset.
Does a bot still make sense in 2026 when AI exists?
Absolutely. For structured tasks a bot is faster, cheaper and more reliable. Payments, bookings and notifications do not need a language model — and should not be paying for one.
How much does it cost to add AI to an existing bot?
From about $600, depending on the quality of the existing code. If the bot was built on a no-code platform with no access to the source, rewriting it is usually cheaper than integrating with it.
Which is safer for sensitive data?
The bot, because the data never leaves your perimeter. For an agent the question is settled by the choice of model: the major commercial providers do not train on business customers' data, and for the most sensitive scenarios we deploy local models on your own infrastructure.
In short
A bot and an AI agent are not competitors but tools for different jobs. The bot wins where there is a script; the agent wins where there is language and data. Most often the right answer to "which should we choose" is "both, each on its own stretch".
Not sure which category your task falls into? Write to us — we will work through it in 30 minutes for free and say so honestly, even if the answer is "a $300 bot will do".