Agent or chatbot? The distinction that actually matters.
Almost every AI product I look at right now calls itself an "agent." Half of them aren't. The distinction sounds like marketing nitpicking, but it changes what you should build, what you should expect, and how to measure whether it actually worked.
Here's the version I give clients.
A chatbot waits
A chatbot is a request-response system. You prompt it, it answers. Then it forgets you exist until you prompt it again. ChatGPT in a browser tab is a chatbot. Your customer-service widget that pops up in the corner of a website is a chatbot. Your AI-powered "assistant" inside a SaaS tool — almost always a chatbot.
Chatbots are useful. They're great for one-off questions, drafting individual emails, summarizing things you paste in. But they have three structural limits:
- They don't watch anything. Nothing happens until you initiate.
- They don't remember state. They don't know what they did yesterday, or whether the thing they recommended actually got done.
- They don't take action. They can tell you what they would do, but they don't reach into your systems and do it.
A chatbot is a smart pair of hands that needs a human to point at every task.
An agent acts
An agent is software with a job. It runs continuously (or on a trigger), watches something specific, makes decisions within a defined scope, and takes action without being told each time.
The shape looks like:
- Trigger: an email arrives, a calendar event is 30 minutes away, a deal moves to a stage, a file lands in a folder.
- Decision: classify, route, summarize, escalate, or skip.
- Action: update a record, send a reply, generate a document, notify a person, create a task.
- Memory: knows what it already did so it doesn't repeat itself.
An agent is software that holds down a seat. Not a smart assistant — a digital employee with one job and no breaks.
Why the distinction actually matters
Three reasons.
One: cost structure. A chatbot is cheap to deploy because you're really just buying access to a model. An agent costs more to build because someone has to wire it into your systems and design the decision logic. But a chatbot scales with how much your team uses it (linear). An agent scales with how much work happens (often non-linear). For repetitive operational work, the agent pays back far more.
Two: what you measure. With a chatbot, you measure adoption ("how many people are using it?"). With an agent, you measure work completed ("how many COIs did it generate this week? how many emails did it route correctly? how many hours did it free up?"). The second number is much more interesting to a CFO.
Three: what kind of problem they solve. Chatbots are good for the unpredictable work — questions you didn't know you'd have, drafts you didn't know you'd write. Agents are good for the predictable work — the same shape of task happening over and over. Most service businesses are drowning in predictable work and trying to solve it with chatbots. Then they wonder why the AI initiative didn't move the needle.
Quick gut-check
If you can describe what you want from AI as "answer my questions" or "help me write things faster" — you need a chatbot. You probably already have one. Use it.
If you can describe it as "watch X, when Y happens do Z" — you need an agent. That's a different build with different math.
The mistake I see most often: someone deploys ChatGPT Enterprise or Copilot, expects operational outcomes, and gets adoption metrics instead. They're not bad tools. They're just not the right kind of tool for that job.
What I build
Agents. Almost always. Because the work that drains a service team — intake, routing, follow-through, reporting — is structurally agent-shaped. The team doesn't need help thinking. They need the predictable parts to handle themselves.
If you're not sure whether your highest-pain problem is a chatbot problem or an agent problem, the free audit is exactly where that question gets answered.