What actually separates an AI agent from a chatbot
Every vendor now says 'agent'. Here is the operational test we use to tell the difference — and why it decides whether a project is worth funding.
Ask five vendors what an agent is and you will get five answers, four of which describe a chatbot with a nicer wrapper. The distinction matters, because the two things have completely different cost structures and completely different failure modes.
The test
An agent is a system that can be given a goal rather than a question, and that can take actions in your systems until the goal is met or it decides it needs a human. That gives you three concrete checks:
- Does it hold state across steps? A chatbot answers one turn at a time. An agent keeps a scratchpad of what it has tried, what worked, and what it still owes you.
- Can it act, not just advise? If the output is text a human then copies into another system, you have an expensive search box.
- Is there a stopping condition it enforces itself? Real agents know when they are done, and know when they are stuck.
Why the difference is expensive
A chatbot fails by saying something wrong. An agent fails by doing something wrong — issuing a refund, emailing a customer, updating a record. That is why every agent we ship has three things a chatbot never needs:
- Typed tools. The model does not write SQL against production. It calls
refund_order(order_id, amount, reason)and the function validates. - Approval gates. Anything touching money, customers or deletion pauses for a human. The gate is configuration, not code, so operations can tighten it without a deploy.
- An evaluation set. A hundred real cases with known-good outcomes. Every change is scored against it before it ships, the same way a failing unit test blocks a merge.
What this means for your first project
Pick a workflow where the actions are reversible and the volume is high. Invoice coding, ticket triage, lead qualification, document extraction. Those give you a real measurement inside a month, and the blast radius of a mistake is a corrected record rather than a lost customer.
The organisations that get value from agents in year one are not the ones with the best models. They are the ones who picked a workflow narrow enough to measure.
The Reciprocal Solutions
AI Engineering Team
We design, build and operate AI agents, automations and custom LLM systems — and stay on the hook for how they behave in production.
Got a workflow that should be running itself?
Bring us the process, the constraints and the mess. We will tell you honestly whether AI is the answer — and what it would take to ship it.