"Agent" is doing a lot of work in AI marketing right now, and it is often applied to things that are not agents at all. The distinction matters because it changes the price by roughly three times.
Here is the honest definition and the test for whether you need one. Prices are on our AI agents page if you already know you do.
The definition
A chatbot answers questions. An agent takes actions.
More precisely: an agent is given a goal and a set of tools, and it decides which tools to use and in what order. It reads a result, decides what to do next, and keeps going until the task is done or it gets stuck.
Ask someone about an order. A chatbot tells you the policy. An agent looks up the order, checks the shipping status, sees it is delayed, drafts an apology with a revised date, and — with your approval — sends it and logs the interaction.
What makes something an agent rather than a workflow
A workflow follows steps you defined. An agent chooses the steps.
That is the whole difference, and it cuts both ways. Choosing steps means handling situations you did not anticipate. It also means less predictability, which is why agents need guardrails that workflows do not.
What businesses actually use them for
- Document processing that varies. Reading whatever arrived, deciding what it is, extracting the right fields, filing it correctly.
- Research and enrichment. Given a company name, gather the details your sales team needs from several sources.
- Ticket triage. Read the ticket, check the account, classify, route, and draft a first reply.
- Multi-step lookups. Questions needing three systems checked in sequence, where the second query depends on the first answer.
- Internal knowledge assistants. Search across your documents and cite the source.
What they cost
Build: ours start at $899 for a single agent and $2,499 for a coordinated multi-agent system. US agencies typically quote $8,000–$40,000.
Running: more than a chatbot, because one task is several model calls rather than one. Typically $30–$300 a month at small-business volume. The agent cost guide breaks it down.
The three controls that make them safe
Anything that can act needs limits. Every agent should have all three:
- Scope. Access only to the specific tools its job needs, so it physically cannot touch anything else.
- Approval gates. Anything irreversible — sending to a customer, moving money, deleting — waits for a human click.
- Confidence thresholds. When unsure, it escalates rather than guessing, and every action is logged with its reasoning.
If a vendor cannot describe all three without prompting, they have not run one in production.
The test: do you need an agent, or something cheaper?
Ask what you want to happen at the end.
- Someone gets an accurate answer → you want a chatbot, at about a third of the cost.
- Data moves between systems by fixed rules → you want workflow automation, cheaper to build and run.
- Something gets decided and then done, and the steps vary → you want an agent.
The expensive mistake is buying an agent for a job a chatbot does perfectly well. We talk people down from this regularly, and the full comparison is in AI agent versus chatbot.
How to know it works before it goes live
Ask for an evaluation set: a collection of your real historical examples with known correct answers, which the agent is measured against before launch.
Without one, "it seems to work" is the entire quality assurance. With one, you see actual accuracy on your actual data — and you can re-run it later, which matters because model providers update models and behaviour drifts.
Frequently asked questions
Are AI agents reliable enough for real work?
For well-scoped tasks with guardrails and a human on anything consequential, yes — that combination is what makes them usable. For open-ended work with no supervision, not yet, and be wary of anyone claiming otherwise.
Can I use an off-the-shelf agent instead?
Sometimes, and we will say so if it fits. Off-the-shelf works when your use case matches what it was built for. It struggles when the agent needs to act inside your own systems or follow rules specific to your business. If you can configure it in an afternoon, use it.
Which models do you use?
Whichever fits the step — larger models for reasoning and writing, smaller cheaper ones for classification and extraction. Routing each step to the cheapest model that does it well cuts running costs substantially. You own the API accounts and can see every dollar.
Is our data used for training?
No. Enterprise endpoints that do not train on customer data, and your NDA signed before anything is shared. If you have compliance requirements, raise them on the first call.
Want to know whether your problem needs an agent? We will tell you honestly, including when something cheaper does the job. See agent pricing or message us on WhatsApp.
