How each one decides what to say
A flow chatbot is a map. Someone listed the questions customers ask most, drew a branch for each, and attached a fixed response to every node. Nothing is decided while the customer is typing. An AI agent has no map: it has a description of your business, a knowledge base, tools it is allowed to call, and a model that reads the message and decides what to use. "Anything Thursday?" and "free thursday if u still got space" land on the same intent, and nobody had to write the second one down.
What happens when someone goes off script
A flow chatbot has one honest response to input it does not recognize: fall back. "Sorry, I did not understand. Please choose an option." The customer picks something they did not want, or leaves, and the question stays unhandled until a human spots the pattern and builds the branch. An agent has no fallback because it has no branches. Off script is not the exception, it is the working condition. That flexibility has a price, and it shows up two sections down.
Who maintains the logic six months from now
A flow is a living asset with an owner. Prices change, services get added, holiday hours start, and someone opens the builder to edit nodes. Every question the market invents becomes another branch, and branches interact with each other. The usual end state is a flow nobody wants to touch because nobody remembers why node 40 exists. An agent moves maintenance from structure into content: you update a price list, a PDF, a page on your site. Less work per month, and less certainty per edit.
What each one can actually do, not just say
Both categories send text. The real question is which one takes an action outside the chat. "Agent" means the system can call tools: read live availability and write a calendar event, create a payment link for a specific amount and follow its status, check stock before promising a size 9. Chatbot platforms reach some of that through integrations, but a person still placed the trigger inside a flow. Since July 1, 2025 Meta bills the WhatsApp Business Platform per delivered template message rather than per conversation, and replies inside an open customer service window are free, so one live conversation beats a chain of templates.
Give your agent your price list and 7 days, and count how many of this week's messages a flow would have dropped. Start free, no card required.
Start freeHow each one fails when it fails
A chatbot fails loudly: it hits an input it cannot match and says so. You lose the lead, but nothing false is ever said, because nothing unwritten is ever said. An AI agent fails quietly. When it fails badly it invents a time slot that does not exist, a discount nobody approved, a policy that sounds like yours. So the serious question is not how smart the model is, but what stops it: date math computed on a server, tool calls validated in code before they run, every decision logged, a human takeover in one click. "We told it in the prompt not to" is not a safety system.
Chatbot vs AI agent, side by side
| Capability | Flow chatbot | AI agent |
|---|---|---|
| Unexpected questions | Falls back to a menu. Dead end until a human builds the branch | Answers from the knowledge it has. No branch has to exist first |
| Maintaining the logic | You edit nodes. Every new question is a branch, and branches collide | You update documents. Vaguer per edit, lighter per month |
| Predictability of the reply | Chatbot wins. Same input, same output, forever | Same intent, same substance, not the same sentence |
| Cost per conversation | Chatbot wins. Priced by contacts: ManyChat listed $14 to $139 a month in July 2026, Wati $29 to $299 | Higher. Model inference is paid on every turn |
| Booking on a real calendar | Possible through an integration fired by a node. Slot logic is yours to build | An agent with a calendar tool reads availability and writes the event |
| Collecting payment | Usually a static link in a message, with no status tracking | Creates the link for the right amount and follows it to paid |
| Written trace per case | Flow analytics: who entered, who dropped, at which node | Every message, tool call and outcome logged per conversation |
| Handing off to a human | A node that drops the chat into an inbox | Escalates on judgment: frustration, complexity, an explicit request |
| Time to first launch | Chatbot wins. A single-branch flow can be live in an afternoon | Longer: knowledge base, WhatsApp API approval, testing |
| Exact control over wording | Chatbot wins. You wrote every sentence. Nothing else can be said | You control tone and facts, not the exact sentence |
When a chatbot is the better choice
Three cases where a flow tool is not the cheap option, it is the correct one.
One question, enormous volume, one perfect answer. A delivery service drowning in "where is my order," a venue answering "are you open right now." If one deterministic reply resolves most of the traffic, an AI agent is an expensive way to repeat a sentence. Build the flow and spend the money somewhere else.
Text that is not allowed to vary. Consent language, credit terms, coverage statements: anything legal approved word by word. The value is that the paragraph is identical in every conversation and survives an audit. A model that rephrases accurately is still a liability. A node that emits approved text verbatim is exactly the right tool.
Predictable one-step interactions. Opt-in confirmations, "reply 1 to confirm, 2 to cancel," lookups against a tracking number. There is no intent to interpret, so paying to interpret it buys nothing.
Most businesses have some of all three, plus a slice of inbound that looks nothing like them. The question is what share of your messages are genuinely predictable.
Where HeySetter falls, and why
HeySetter is an AI sales agent for WhatsApp, built for the unpredictable half: leads who ask what nobody scripted and who need an appointment and a payment in the same conversation. It replies in under 30 seconds, 24/7, in English and Spanish. Scheduling is computed on our servers so the model never does date math, guardrails run as code before any tool executes, and a deal moves to Won only when the payment provider confirms the charge. One plan at $350 a month, 3,000 unique conversations included, $0.08 after that. If your WhatsApp is one predictable question at high volume, a flow tool costs less and we will tell you so.
Frequently asked questions
No. The writing is a symptom. A chatbot picks a stored response from a tree built in advance, so its ceiling is whatever a person predicted. An agent interprets the message and chooses an action, including calling a calendar or a payment tool. Same channel, different machinery.
One system owns the number at a time, so in practice you choose. What you can do is give an agent deterministic behavior where it matters: fixed templates for reminders, approved wording for legal text, reasoning only for the open part of the conversation.
Read one week of inbound. Sort the messages into ones a menu could resolve and ones that need a real answer about price, availability or fit. If the second pile is small, build a flow. If it is where your revenue lives, a menu is losing you those leads.
Architecture, not instructions. Dates and open slots calculated outside the model, tool calls validated programmatically before they execute, a knowledge base as the source of facts, logging of every decision, and a human who can take the chat over instantly. Ask any vendor which of those they have.
