Is an agent just a chatbot with API access?

Roughly, yes — plus the control loop that decides when to call which tool, and the state that ties several calls into one task. The API access is the visible part; the discipline around permissions, retries and logging is what makes it production-grade.

Can an agent replace my support team?

It can absorb the transactional layer — status checks, changes, cancellations — which is usually 40-70% of volume. The conversations that need judgment, empathy or negotiation should land on humans, and a well-built agent knows when to hand off.

What is the biggest mistake teams make with agents?

Giving them write access before building the guardrails. Start read-only, add one carefully scoped action, watch it for a month, then expand. Teams that do the reverse learn about their missing audit logs the hard way.

// ai

AI agent vs chatbot: what actually changes when the software starts doing the work

3 min read · all guides

Every vendor demo now ends with an "agent". The word has been inflated to the point of uselessness, so here is the working distinction we use with clients: a chatbot produces text in response to a message; an agent takes actions in the world — reads your systems, writes to them, and is responsible for an outcome. A support chatbot that drafts a refund email is a chatbot. A system that checks the order, issues the refund and confirms to the customer is an agent.

The difference is not cosmetic. It changes the architecture, the failure modes, the cost per conversation and the security review. This guide walks through each one so you can tell which of the two you actually need.

The technical difference in one paragraph

A chatbot is a loop: message in, text out. It may retrieve documents first (RAG), but its output is advice. An agent adds three things: tools (APIs it can call), state (memory of what it has done across steps) and a control loop (plan, act, observe, retry). That loop is where the value lives — and where the danger lives, because each action is a write operation you used to reserve for a human. The moment your software can change data, you are running a permissions problem, not a copywriting problem.

When a chatbot is the right answer

Most of the time, honestly. If the job is answering questions from a knowledge base — opening hours, order status, policy details — a retrieval chatbot is cheaper, faster to ship and dramatically easier to keep out of trouble. It has no credentials to leak and no actions to audit. A good rule from the field: start with the most passive thing that solves the problem, because every capability you add is a capability you must monitor, rate-limit and roll back. Teams that skip this step end up paying agent prices for FAQ answers.

When you actually need an agent

You need an agent when the value is in the transaction, not the conversation: booking systems that check real availability, support flows that process refunds, operations tools that triage and route. The tell is a checklist — if a human today performs a fixed sequence of reads and writes across systems, an agent can probably do it. Budget for the boring parts from day one: scoped API keys (an agent should never hold your admin credentials), idempotent actions, a dry-run mode, and a log of every step it took. These are not optional extras; they are the difference between an agent you trust and a demo you fear.

Cost: the number vendors skip

A chatbot conversation costs fractions of a cent in model calls. An agent task — plan, five tool calls, retries — costs ten to fifty times more, and the bill grows with the length of the task, not the politeness of the customer. Worse, costs are spiky: one malformed API response can send an agent into a retry loop. Before committing, instrument a pilot and measure cost per completed task, not per token. We have seen honest agent deployments land between a few cents and a few euros per task; the spread depends almost entirely on how well the tools were designed.

Frequently asked questions

Is an agent just a chatbot with API access?: Roughly, yes — plus the control loop that decides when to call which tool, and the state that ties several calls into one task. The API access is the visible part; the discipline around permissions, retries and logging is what makes it production-grade.
Can an agent replace my support team?: It can absorb the transactional layer — status checks, changes, cancellations — which is usually 40-70% of volume. The conversations that need judgment, empathy or negotiation should land on humans, and a well-built agent knows when to hand off.
What is the biggest mistake teams make with agents?: Giving them write access before building the guardrails. Start read-only, add one carefully scoped action, watch it for a month, then expand. Teams that do the reverse learn about their missing audit logs the hard way.

// next steps

Working on something adjacent? Tell us about it — or read the services we deliver today.