AI agent vs chatbot
A chatbot answers. An agent answers and then does something. A chatbot retrieves information and puts it in front of a person. An agent takes actions in your systems on the strength of what it worked out, books the appointment, creates the record, routes the inquiry to the right owner with the details already filled in.
That difference is the whole cost gap and the whole risk gap. A chatbot that is wrong has given a bad answer. An agent that is wrong has changed something. Everything else, evaluation, confidence thresholds, approval steps, logging, follows from that one distinction.
Buy the chatbot if the job is answering questions off your own material. Buy the agent when the answer is only half the work and the other half is currently done by a person retyping it somewhere.
The difference in one table
| Chatbot | AI agent | |
|---|---|---|
| What it does | Retrieves and phrases an answer | Decides, then acts in your systems |
| Reaches into | A body of content you gave it | Your CRM, calendar, ticketing, billing |
| Worst case when wrong | A misleading answer a person can correct | A wrong record, a wrong booking, a wrong route |
| Needs an evaluation set | Advisable | Non-negotiable |
| Needs a confidence threshold | Useful | Required, with a handoff path |
| Needs approval steps | No | Yes, for anything that spends money or is hard to reverse |
| Typical build | Days | 10 business days for a single agent |
| Our price | $3,500 chat agent · $5,000 voice agent in setup, then pay as you go at $175 a change or a plan from $700 a month. Both act, both include evaluation | |
We do not sell a pure chatbot as a product, and it is worth saying why: a bot that only answers and cannot do anything tends to disappoint quickly. The question customers ask is rarely just informational. "Do you do X" is followed by "can you book me in". A bot that stops at the first half creates a handoff rather than removing one.
Where the old word still applies
"Chatbot" also covers the decision-tree widget that has been on websites for a decade: fixed buttons, scripted branches, no language understanding at all. Those are not obsolete. They are fast, free of running cost, entirely predictable, and genuinely better for narrow jobs like "track my order" or "open a return".
If your inquiries fall into five predictable types, a decision tree beats anything with a model behind it, cheaper, faster, and incapable of inventing an answer. The case for language understanding starts when customers phrase things in ways you did not anticipate, which is most of the time for anything more complex than order tracking.
Write down the twenty most recent inquiries. If sixteen fit five templates, build a decision tree. If they are all slightly different, you need something that understands language. This takes an hour and settles the argument with evidence rather than preference.
What the extra money actually buys
The engineering of connecting a model to a chat widget is not the expensive part. These four are, and they are the difference between a demo and something you can leave running.
- An evaluation set. Real historical cases with known correct answers, measured before launch and after every change. Without it, nobody can tell you how often it is wrong; you learn it from customers.
- Confidence thresholds and a handoff. The agent must be able to decline. Anything below the threshold goes to a person with the conversation attached, rather than being guessed at.
- Structured output. The agent returns typed fields your systems can act on, not a paragraph. This is what makes the action reliable rather than a parsing exercise.
- Logged exchanges. Every conversation stored, so you can see what was said when someone disputes it, and so retraining has something to learn from.
A quote materially below ours is usually missing the first and second. Those are also the two that decide whether it survives month three. What an AI agent costs breaks the pricing down, including what you pay providers rather than us.
When neither is the answer
Both lose to a plain workflow when the decision follows rules. If inquiries route by value, territory and product (all knowable from the form) a rules-based workflow does it deterministically for $1,500 setup, with no cost per conversation and no possibility of invention.
How to decide in five minutes
| If this is true | Build |
|---|---|
| Inquiries fall into a handful of fixed types | A decision tree, not AI |
| The decision follows rules you can state | A workflow, $499–$1,500 |
| People ask in their own words, and answering is the whole job | A chat agent, answering only |
| Answering is half the job and someone then retypes it | A chat agent with actions, $3,500 setup |
| The calls are already coming and nobody picks up | A voice agent, $5,000 setup |
| Being wrong is expensive and nobody would notice | None of them. Keep a person |
What we would actually recommend
Reach for language understanding when the input is genuinely unstructured and the rules cannot be written down. Reach for rules everywhere else. The workflows pay back faster, and they cost $499 to $1,500 rather than $3,500. Do I need an AI agent or just a workflow draws the line precisely.
A chatbot that only answers wins instead when answering off your own material is the whole job. A decision tree beats both when your inquiries fall into a handful of fixed types, and none of them is right when being wrong is expensive and nobody would notice.
Not sure which one your inquiries need?
Send us twenty recent ones. We will tell you whether they want a decision tree, a workflow or an agent, and the scoping call that establishes it costs nothing.
You will talk to Fayez or Kareem. Nobody else.