The short answer: if 80% of your inquiries are routine questions with predictable wording (price, hours, order status), a scripted chatbot is enough — from $15–$40 a month on a builder or from $400–$1,100 custom-built, with a 1–3 week launch. If you need to understand natural language, answer from a knowledge base, qualify leads and work with a CRM — you need an AI agent: from $1,800–$3,600 for development plus support and tokens, launching in 3–6 weeks. The budget gap is 5–10x, which is why you should choose by task, not by fashion.
Below are the tools for making that choice: a comparison of capabilities and money, a task → tool → budget matrix, a five-question decision tree and a breakdown of the mistakes that cost businesses either money or customers.
How an AI agent differs from a chatbot
The confusion starts with the fact that both live in the same little chat window. Inside, though, the mechanics are fundamentally different. The easiest way to see it is as three maturity levels:
- Scripted chatbot. Executes a pre-written script: buttons, branches, canned replies. Anything the script did not anticipate triggers "I don't understand" or silence. In return it is cheap, predictable and never makes things up.
- AI assistant. A language model connected to your knowledge base. It understands a question phrased any way at all and answers from your documents: price list, policies, FAQs. But it remains an "answerer": it takes no actions on its own.
- AI agent. The next level: it does not just answer — it acts. It steers the conversation toward a goal — qualifies the customer, clarifies details, files a lead in the CRM with the conversation context, books a time slot, and hands anything complex or heated to a human rep. An agent has a task and defined boundaries of authority, not just "chat with people".
Hence the key practical takeaway: the question "how does an AI agent differ from a chatbot" is really the question "what should happen after the customer's first message". If the answer is "show the right menu item" — a bot will do. If it is "figure out who this is, what they need, and take them to a submitted request" — that is an agent's job.
Comparison table: capabilities, money, timelines
Let's put the differences in one table. Budget ranges are based on published platform and studio pricing for 2026; we return to them in a dedicated section below.
| Criterion | Scripted chatbot | AI agent |
|---|---|---|
| Understanding requests | Buttons and keywords; one step off — "I don't understand" | Natural language in any wording, with typos and context |
| Source of answers | Texts hard-coded into the script | Your knowledge base: price list, policies, FAQs, documents |
| Actions | Branch navigation, simple forms | Lead qualification, booking, CRM records, handover to a rep with context |
| Integrations | Messengers, simple webhooks | CRM, telephony, website, messengers, internal systems |
| Launch cost | Builder from $15–$40/mo; custom from $400–$1,100 | From $1,800–$3,600; complex projects from $6,000 |
| Monthly costs | Builder subscription, minor script tweaks | Support $250–$600/mo + model tokens + knowledge base updates |
| Launch timeline | 3 days — 3 weeks | 3–6 weeks, complex builds up to 8–12 |
| Where it breaks | Unusual wording, "get me a human", frustration | An empty or stale knowledge base, tasks with no clear goal |
An important caveat about "where it breaks". A bot breaks on the customer's side — it irritates people and kills conversation conversion. An agent breaks on the business's side — if it was not given a current knowledge base and a clear task, it will politely and confidently spin its wheels. The stronger the tool, the higher the bar for how well the task is defined.
The task → tool → budget matrix
Abstract differences do not drive decisions — anchoring to your task does. Find your row:
| Task | Tool | Budget reference (2026) | Timeline |
|---|---|---|---|
| Answering routine questions: hours, address, prices, delivery | Scripted bot on a builder | $15–$40/mo | 3–7 days |
| Bookings, order intake, step-by-step payment | Custom scripted bot / Telegram bot with payments | $700–$1,800 one-off | 2–4 weeks |
| First-line processing and qualification of leads in a live conversation | AI agent | $1,800–$3,600 + support | 3–6 weeks |
| Customer support from a knowledge base: documentation, policies, plans | AI agent with a connected knowledge base | $1,800–$4,800 + support | 4–8 weeks |
| Sales: driving the conversation to a request, CRM, telephony, handover to a rep | AI agent with integrations | from $3,600 to $12,000+ | 6–12 weeks |
| Automation beyond chat: documents, reports, inquiry handling inside processes | That is AI integration into processes, starting with an audit and a pilot | priced after the audit | from 4 weeks for a pilot |
If your task lands in two rows — say, "step-by-step booking" and "lead qualification" — start with the one where money is leaking right now. Most often that is overnight and weekend inquiries: a customer who did not get an answer goes to whoever answered first.
Decision tree: 5 questions to choose in 10 minutes
Work through the questions top to bottom — the first "yes" usually determines the choice.
- Can you write out 80% of inquiries as 10–15 standard questions with ready answers?
Yes → a scripted bot is enough; start with a builder. No → keep going. - Do customers phrase the same thing in dozens of different ways, with details and follow-ups?
Yes → you need AI: no script will cover all the wording. No → a bot will cope. - Should the assistant take actions — file a lead in the CRM, book a slot, calculate an option — and not only answer?
Yes → that is an AI agent, not a Q&A assistant. No → an AI assistant over a knowledge base is enough. - Do you have a knowledge base (price list, policies, FAQs) and a person who will keep it current?
Yes → the agent will perform at full strength. No → build the base first, otherwise the agent will answer into a void — this is part of the implementation work. - Will a budget of $1,800–$3,600 pay back at your inquiry volume?
Estimate: how many leads a month are lost to slow answers and off-hours, multiply by your close rate and profit per deal. If the losses over 3–6 months exceed the implementation budget — yes. If the flow is still 5–10 inquiries a day and all routine — start with a bot and revisit the agent as you grow.
What a chatbot and an AI agent cost in 2026: honest ranges
The ranges below are compiled from published builder and studio pricing for 2026. The spread is wide because very different things are sold under the word "bot" — compare the scope of work, not the price.
- Builder (SaaS). Basic plans from $15–$40/mo, advanced tiers run into hundreds a month. You pay the subscription and assemble the flow yourself or with an integrator. Fast and cheap, but only within the platform's limits.
- Custom scripted bot. An FAQ bot from $400; a sales bot with payments and a catalog from $1,100–$1,800. The market's average custom-build ticket is around $3,500 — but that already includes integrations.
- AI agent. A basic project with a knowledge base from $1,800–$3,600; builds with CRM, telephony and voice from $6,000 up to six figures with enterprise vendors. Plus recurring costs: support and tuning at $250–$600/mo, and language-model tokens from a few dollars a month, growing with the dialogue volume.
- tokens: the model bills every dialogue — as volume grows, so does the invoice;
- knowledge base upkeep: prices and policies change, and someone has to update the agent's version of them;
- post-launch tuning: for the first 2–4 weeks the agent is polished on real conversations — a normal part of the project, but it must be in the quote up front.
At Growth Bureau we have no fixed price list for agents — deliberately: we quote per task after reviewing your scenario, together with a payback calculation. When AI clearly will not pay off, we talk clients out of it.
Common selection mistakes
Five scenarios where businesses most often lose money or customers. All from real project reviews.
- Buying an AI agent for a bot-sized task. If every question is "hours, price, address", the agent will do the job of a $15–$40/mo builder — just 5–10 times more expensively. The tell: the brief contains not a single action, only answers to questions.
- Putting a button bot on live sales. The reverse mistake, and the more expensive one: a customer with a non-standard question bounces off the buttons, types "get me a human" and leaves. Saving $1,000 on the build turns into lost deals every month.
- Comparing launch price instead of cost of ownership. A bot's cost sits almost entirely at the start; a large share of an agent's cost is recurring: support, tokens, the knowledge base. The correct comparison is 12-month budget against 12-month effect.
- Believing in a "turnkey AI agent for $200". Behind that price is either a builder template with ChatGPT glued on — no knowledge base, no integrations — or a demo that falls apart on the first real customer. The floor for a meaningful custom AI project in the 2026 market is around $1,800.
- Choosing the tool before describing the task. "We want AI like everyone else" is the road to the shelf. The working sequence is the reverse: inquiry flow → what should happen to it → which tool does that most cheaply. That is exactly why a proper implementation starts with a scenario review, not with picking a technology.
The hybrid path: start with a bot, grow into an agent
The choice is not always binary. For a growing business, a ladder is often optimal:
- Now: a scripted bot covers the top 10–15 routine questions and captures inquiries outside working hours. Launch in 1–3 weeks, minimal budget.
- In parallel: build a knowledge base (price list, policies, answers) separately from the bot and connect a CRM — so every inquiry is logged with its source.
- As the flow grows: swap the "brain" — an AI agent replaces the script, inheriting the ready knowledge base, channels and CRM. Migration becomes a component swap, not a rebuild from zero.
Signs it is time to move to an agent: the share of dialogues where the bot gives up and calls a human stays above 30–40%; reps spend more than 2–3 hours a day on chat; overnight inquiries appear with nobody to handle them. From that point the agent stops being a fashionable toy and becomes cheaper than hiring.
One last thing. Chat in a messenger is only the visible part: the same models can process documents, prepare reports and strip routine out of internal workflows. If your task is wider than customer conversations, look at AI integration into business processes — the choice there follows the same task → tool → payback logic, just at the level of the whole company.