Blog
Automation or AI agent: how to tell which a task needs
Some tasks need a trigger and an action. Others need an agent that reads and decides. How to tell them apart, with examples, and how Cedric runs both.
Benjamin Ben HattarTrigger-and-action automation has been around for years, and it works. When a form is submitted, add a row to a sheet. When an invoice is paid, post in the finance channel. You write the rule once and it runs forever.
It breaks in a predictable place. The rule needs the input to arrive in the same shape every time, and it needs every branch written in advance. Real work often fails both tests. A lead writes three paragraphs instead of filling the form. A client replies "paid Friday" to a reminder. The rule has no branch for that, so a person steps in, and the automation quietly becomes a to-do list.
An AI agent takes a different job. It reads the input, decides what it means, and picks the next step. That costs more per run and needs guardrails, so it is worth knowing when you need one and when a plain rule is the better tool.
- Use a fixed rule when the input always has the same shape and every branch can be written down.
- Use an agent when someone has to read the input, compare sources, or choose the next step.
- Most real processes mix both. The rule carries the data and the agent handles the exceptions.
- An agent that acts needs an approval step before it sends, changes or spends.
- Cedric runs both kinds: work on a schedule, runs started by an event, and workflows that wait on people.
When a fixed rule is enough
Keep it a rule when the data is structured, the action is the same every time, and a wrong run is cheap to notice. Copying a paid invoice into the ledger sheet. Posting a new signup in a channel. Creating a calendar hold when a booking comes in. Nothing needs to be read or weighed.
Rules are also predictable in a way an agent is not. The same input gives the same output every time, which matters for anything auditors look at.
When the task needs an agent
Four signs that a task has outgrown a rule.
The input is text a person wrote. An inbound email, a meeting, a support thread. Something has to read it to know what it is.
The answer lives in two systems. Does the pipeline match what billing says? Is this lead already in the CRM under another name? A rule moves data between tools. Checking one against the other takes judgment.
The exception is the job. A monthly reconciliation that matches 209 of 212 payouts is only interesting for the three it could not match. Deciding what those three mean is the work.
The next step depends on what happened. If the client has paid, drop them from the list. If they asked for more time, flag it. If they did not answer, nudge in a week. Writing every branch in advance is how rule sets turn into diagrams nobody can maintain.
Examples of each
| Task | Fixed rule works? | Why | How Cedric runs it |
|---|---|---|---|
| New signup posted in a channel | Yes | Same shape, same action, every time | A run started by the event, if you want the signup looked up too |
| Form submission added to a sheet | Mostly | Until duplicates and messy fields appear | Adds the row, skips duplicates, confirms in the channel |
| The weekly numbers post | Partly | Collecting is mechanical, explaining the dip is not | "Every Monday at 9", with the reason behind each move |
| Chasing overdue invoices | No | Tone and the client's replies change the next step | One draft per invoice, sent on a yes, replies read |
| Qualifying an inbound lead | No | Someone has to read what they wrote | Who they are, whether they fit, the reply drafted |
| Onboarding a client over a week | No | It waits on people who answer late or not at all | A workflow that asks, nudges and resumes |
How Cedric handles both
Cedric covers the whole range without a builder. There are three ways work starts.
On a schedule. Say "send me the numbers every Monday at 9" in the channel. He repeats back what he understood before the job exists, and you tell him whether to post every time or only when something needs attention. See recurring work and automations.
On an event. A new deal, a paid invoice or a merged pull request can start a run. Deskimo's website forms send each new lead to Cedric, who adds it to the tracker, skips duplicates and confirms in Slack. The Deskimo story has the setup.
As a workflow that waits. "Collect availability from the five interviewers, then book the loop" cannot finish in one message. Cedric asks each person, credits the replies as they come in, nudges the quiet ones and books when the last answer lands. Runs are bounded, so a stuck one stops and tells you. See workflows that wait.
In all three, the same rule holds. The schedule or the trigger approves the routine part. Any step that would send to someone outside, change a record or spend money still stops and asks.
Rule plus judgment
Every new lead, read the same way.
Hannah Berg9:14 AM
@Cedric from now on, every new inbound lead: check who they are, whether they are already in the CRM, and draft the reply. Post it here.
CedricAPP9:15 AM
Done. Each new inbound gets a short brief in #sales: who wrote in, what they want, whether they fit and why, and a drafted reply. Sending the reply or creating the CRM row waits for your yes.
When not to use an agent
If a fixed rule does the job and never needs a person, keep the rule. An agent costs credits on every run, in proportion to the work, and a copy from one field to another does not need reading.
Do not schedule an agent job you have not checked by hand. Run it, correct it, compare it with a week you already know, and only then say "every week". A wrong report delivered on time is worse than no report.
And keep the decision human. Ask for the analysis on a schedule. Make the call yourself.
FAQ
Automation runs a fixed rule: when this happens, do that. An AI agent reads the input, decides what it means and picks the next step. Rules suit structured data; agents suit text, exceptions and work across systems.
For work that needs reading or judgment, yes. For a simple field-to-field copy, a fixed rule is cheaper and just as good. Many teams keep both.
Yes. A new deal, a paid invoice or a merged pull request can start a run, and anything that can send a notification, like a website form, can hand Cedric the job.
No. You describe the job and when it should run in the channel. Cedric confirms what he understood before anything is created, and you change it the same way.
With Cedric, the approval gate. The schedule or trigger approves the routine work, and every step that sends, changes a record or spends waits for a person. See how approvals work.
100,000 free credits on sign-up, no card. Plans from $19 a month for the whole team.