In short
An agent combines a language model with a set of tools, usually connections to other software through APIs. Given a goal, the model decides which tool to call, reads the result and decides the next step, repeating until the job is done or it reaches a stopping rule. A chatbot answers. An agent acts.
Chatbot, workflow or agent
| Chatbot | Answers questions in a conversation. Takes no action in other systems. |
|---|---|
| Workflow automation | Fixed steps that run the same way every time: when a form arrives, create a contact and send an email. |
| AI agent | Chooses its next step based on the situation, using tools it is allowed to call, within rules and limits. |
Where agents work well
Agents earn their keep on jobs with a clear outcome and some variation in how you get there: qualifying an enquiry and booking a call, triaging a support message and routing it, or gathering details from several systems into one summary.
If the steps never vary, a plain workflow is cheaper, faster and easier to test. Plenty of “agent” projects turn out to be a workflow with one AI step in the middle, and that is a good result.
Scoping an agent safely
Give it the smallest set of tools the job needs, and read-only access wherever writing is not required. Spending money, deleting records and messaging customers in bulk should need a person to approve.
Write down the stopping conditions: when the agent should hand over, how many attempts it gets, and what it says when it cannot help. Keep a log of every action, so you can see what it did and why.
Measuring whether it works
Decide the measure before you build: enquiries qualified, bookings made, time saved per week, or handovers that arrived with complete details. Review a sample of conversations each week for the first month. The fixes in that month are usually small changes to instructions and permissions, and they make the biggest difference.
When this matters
Relevant when a repeatable task needs judgement as well as steps. New Era AI scopes agents under its AI agents service, and Echo One is an agent with a voice: it answers calls, captures details and books or hands over.
Questions
Can an AI agent run unsupervised?
For low-risk, reversible tasks with clear rules, largely yes, with logs and regular review. Anything that commits money, makes promises to customers or affects someone significantly should keep a person in the loop.
What does an agent need access to?
Only the systems and actions its job requires, ideally through accounts created for the agent so its actions are traceable and can be switched off on their own.
Is an AI agent the same as an AI voice agent?
A voice agent is an agent whose conversation happens on the phone. It uses speech recognition and a synthetic voice around the same kind of model and tools.