Every vendor selling you something with a language model in it now calls it an agent. Most of the time the thing you are being shown is a workflow, and the distinction is not academic: it decides what happens the day a caller says something nobody wrote a branch for.
A workflow is a decision tree someone drew
A workflow runs the path you configured. Someone sat down, listed the cases, and wired an outcome to each one. The language model may write the words, but it does not choose the next step, and it cannot invent one.
- The steps are fixed at design time, and the same input always takes the same path.
- It fails loudly and predictably: an input outside the tree falls to a default, usually a voicemail or a transfer.
- It is cheap to reason about, which is exactly why most useful automation should be one.
Hinweis
A word on where the term comes from
The distinction is not marketing: it is the same one the research uses. An agent has a goal and a set of actions; a workflow has a graph. The vocabulary drifted, the difference did not.
An agent chooses its own next step
An agent is given a goal, a set of tools and permission to decide the order. Book the appointment. Look up the account. Send the confirmation. Nobody wrote the sequence, and two callers with the same request can be handled two different ways because the second one mentioned they are calling from a job site.
If you can draw the whole thing on a whiteboard before it runs, it is a workflow. If you can only draw the goal and the tools, it is an agent.
That freedom is the entire value, and the entire risk. An agent handles the case nobody anticipated, and it also takes an action nobody anticipated. Which is why the interesting question is never agent or workflow, it is which actions are you willing to let it take without asking.
Achtung
The permission question comes before the model question
Decide which actions can happen without a human, then pick the tool. Doing it the other way round is how a system ends up allowed to email a customer nobody meant to contact.
Why the difference lands hardest on the phone
A missed call is not one task. It is a chain: work out who called, work out what they wanted, decide whether it is urgent, answer them, write it down somewhere a colleague will see it, and follow up if nobody did. A workflow can do the first and the fourth. The chain in the middle is where the two part ways.
- Work out who called, from the number and whatever the CRM already knows.
- Work out what they wanted, from what they actually said.
- Decide whether it can wait until tomorrow.
- Write it somewhere a colleague will see it before they call back.
- 62%
- of calls to a small business go unanswered at least once
- 4 min
- median wait before a caller hangs up rather than hold
- 1 in 3
- callers who hang up never call back
Where each one actually sits
| The job | Workflow | Agent |
|---|---|---|
| Ring the right people in the right order | Yes | Overkill |
| Text back a missed call | Yes | Overkill |
| Route by opening hours | Yes | Overkill |
| Work out what an unclear request was about | No | Yes |
| Write the call up in the CRM's own fields | Partly | Yes |
| Decide whether a call needs a human today | No | Yes |
What an agent buys you
- It handles the request nobody wrote a branch for.
- It reaches for the right tool instead of the next step in a list.
- The work a call creates gets done while the call is still fresh.
What it costs you
- You cannot predict the exact path, so you have to bound the actions.
- It needs real permissions on real systems to be worth anything.
- A mistake is an action taken, not a branch missed.
How to tell which one you are being sold
Ask to see the configuration screen
A tree of conditions is a workflow, however good the voice is. There is nothing wrong with that, but it should be priced like one.
Ask what it does with a request it has no tool for
A real answer names the escalation and who receives it. A vague one means nobody has tried.
Ask which actions it can take without a human confirming
If the answer is none, the word agent is doing marketing work. If the answer is everything, ask what the limits are.
Ask for last week's log, not the demo
The demo is the path someone drew. The log is what actually happened to real callers.
- The escalation is named, and someone receives it.
- The list of actions it may take without a human is written down.
- There is a log of last week you can actually read.
Tipp
The one question that separates them
Ask what happens when a caller asks for something the demo did not cover. A workflow tells you which fallback it lands in. An agent tells you which tools it would reach for.
If the answer comes back as a webhook payload, that is a good sign: it means someone has thought about what leaves the system, not just what the voice sounds like.
{
"event": "call.completed",
"direction": "inbound",
"from": "+13055550142",
"duration_seconds": 96,
"summary": "Asked for a quote on a water heater swap, wants a callback before 6pm.",
"actions": [
{ "type": "crm.note.created", "target": "hubspot", "id": "note_8812" },
{ "type": "task.created", "assignee": "dispatch", "due": "2026-08-26T18:00:00Z" }
]
}What we do about it
Most of what a small team needs from its phone is a workflow, and should stay one: ring the right people, text back a missed call, route by opening hours. The part worth making agentic is the work the call creates afterwards, because that is the part nobody can draw in advance. That is the line we build to, and it is why our receptionist answers on a tree and our agents write to your CRM.
A note on the vocabulary
None of this is a reason to avoid the word. Agent is a useful word for a system that picks its own next step. It stops being useful the moment it means “anything with a model in it”, which is roughly where the market is now.
Weiterlesen
Das CRM, das sich nach jedem Anruf selbst ausfüllt
Artikel lesen
Questions people actually ask
Is an AI receptionist an agent or a workflow?
Most of them are workflows with a good voice, and that is the right design for answering a call: you want the same greeting, the same routing and the same fallback every time. What happens after the call, writing it up and following up, is where an agent earns its name.
Do I need an agent at all if my calls are simple?
Probably not for answering them. If your calls create work in other systems, a CRM note, a quote, a callback task, that is where the time goes and that is what an agent is for.
What should an agent never be allowed to do without a human?
Anything that reaches a customer under your name and anything that moves money. Draft the email, do not send it. Prepare the refund, do not issue it. Everything else is a judgement call about how much you trust the log.
How do I know it is not just a chatbot with a phone number?
Ask for the list of tools it can call and the log of what it actually called last week. A chatbot has no tools and therefore no log.
