AI agent vs chatbot
A chatbot chooses its next sentence. An agent chooses its next action. Everything in the table follows from that — including the fact that a well-built chatbot is the right answer for a large share of what people ask agents to do.
| Dimension | Chatbot | AI agent |
|---|---|---|
| Decides | Next sentence, from a designed tree or a model | Next action: which tool to call, then what to say |
| Handles | Inputs the designer anticipated | Free text, voice notes, photos, mixed language |
| Can take actions | Only fixed ones wired to a menu item | Any tool it is granted, gated by guardrails |
| Predictability | Deterministic (rule-based) or fluent-but-unverified (LLM chat) | Probabilistic decisions inside deterministic controls |
| Cost per conversation | Near zero (rules) or one model call | Several model calls; ≈US$0.02–0.05 on WhatsApp with caching |
| Build effort | Days for a menu; weeks for a good tree | Weeks: tools, policy, golden set, evaluation |
| Failure mode | Dead ends ("I did not understand") | Confident wrong actions if ungated |
| Governance | Content review | Permissions, confirmation gates, audit log, evaluation, data-controller duties |
| Best channel fit | USSD-style menus, WhatsApp buttons/lists | WhatsApp free text, voice, email |
| Measured by | Completion of designed flows | Task success + policy compliance + handoff quality |
| Upgrade path | Add an agent behind the menu for the open-ended branch | Add menus in front to reduce cost and ambiguity |
| When it is enough | Balance checks, opening hours, order status by number | Booking from a sentence, quoting, reconciliation, triage |
Two kinds of chatbot
The word covers two different things. A rule-based chatbot follows a designed tree — the WhatsApp equivalent of a USSD menu, and familiar to every Zimbabwean who has checked an EcoCash balance. It is deterministic, cheap and auditable, and it dead-ends on anything unplanned. An LLM chat answers free text fluently from a model with no tools; it feels intelligent, and it cannot verify a single thing it says about your stock, your prices or your calendar. Neither can book an appointment from "this week, afternoons". An agent can, because it has a calendar tool and a loop.
The decision rule
Ask two questions. Do the inputs fit a menu? If yes, build the menu: it is faster, cheaper and never hallucinates. Does the job need an action taken from what the customer said? If no, an LLM chat with retrieval over your documents is enough. Only when the answer is "no menu" and "yes, act" do you need an agent — and you then need guardrails that a chatbot never did.
The hybrid most businesses should build
A menu at the front door (WhatsApp buttons: "Book", "Prices", "Track order", "Talk to us"), a rule-based flow for the structured branches, and an agent only behind the branch that reads free text — "Book". The menu reduces ambiguity and model calls; the agent handles the part menus cannot. Cost per conversation drops, containment rises, and the agent's golden set gets smaller because the menu already narrowed the intent. This is the pattern the demo assumes: the customer's first message already says "check-up".
Cost, concretely
A rule-based flow costs nothing per message inside WhatsApp's 24-hour window. An agent conversation costs a few US cents in model tokens — roughly US$0.025 for the demo booking on a mid-tier model with prompt caching (basis on the costs page). The difference is small per conversation and large in what you get: a booking instead of a "please call us on…".
Next: agent vs automation, or the definition and four-question test.