Blog
What is an AI employee?
An AI employee owns a job in your team. It uses your tools, remembers what it learns and asks before it acts. What that means, and what to hand one first.
Benjamin Ben HattarThe phrase gets stretched to cover almost anything with a text box. So here is the definition we build against. You give an AI employee a job. You give a chatbot a question. It sits where your team already talks, holds connections to the tools the job needs, remembers what it learned last time, and keeps working when nobody is typing to it. Before anything leaves the team, it asks.
That last clause matters most. Everything else on the list makes an AI employee useful. The asking is what makes it safe to hand one real work.
- An AI employee owns a recurring job. You check the result and skip the steps.
- A chatbot answers the question in front of it and forgets. An AI employee remembers corrections for the whole team.
- A trigger-action automation follows one fixed path. An AI employee reads the situation, and when it is unsure it asks.
- Safe handover rests on one rule: reading is free, and anything that sends, changes or spends waits for a person's yes.
- Start with work that repeats, costs little when wrong, and takes a minute to check.
What an AI employee is not
It is not a chatbot. A chatbot lives in its own window. You paste in the context, it answers, and the answer stays in that window until you copy it somewhere. Next week you paste the same context again, because it kept nothing. Useful for a draft. Useless for a job that repeats.
It is not a trigger-action automation either. Those are rules of the form "when a form comes in, add a row to the sheet". They are reliable for as long as the input keeps its shape. The day a customer replies to the form email with a question instead of an answer, the rule adds the question to the sheet and nobody hears about it. The rule cannot read a thread and notice that this one needs a person.
An AI employee sits between the two. It reads like the chatbot and repeats like the automation. It also remembers what the team told it, and it stops at the points where a human should decide.
The three side by side
| A chatbot | A trigger-action automation | Cedric, an AI employee | |
|---|---|---|---|
| Where you use it | Its own window | A builder you set up once | Your team's chat, in Slack, Discord or Telegram |
| What it can reach | What you paste in | The two apps in the rule | Every tool you connect, read-only first |
| What it remembers | This conversation | Nothing, it follows the rule | One shared memory for the team |
| Works without being asked | No | Yes, on its trigger | Yes, on a schedule you set in a sentence |
| When the input is odd | Answers anyway | Runs the rule anyway | Says what he found and asks |
| Before it sends or changes something | It cannot | It just does | He asks, every time |
What one can own in a team
A good first job is one a new colleague would get in week one. It needs the team's tools, it repeats, and a person can check the result in a minute. Here are the jobs Cedric takes on.
The weekly numbers. He reads billing, the CRM and the sheet, then posts one report in the channel every Monday. See how to put the Monday report on a schedule.
Meetings. He joins the call from the calendar invite and posts the summary when it ends. Every commitment he heard becomes a proposal waiting for approval.
The inbox. He reads what came in and drafts the replies worth sending. Each draft waits as a preview.
Chasing people. "Get a status line from these four by Thursday" turns into a job that runs for days. He asks each person, nudges the quiet ones and posts the roll-up.
Meeting prep. An hour before the call, he posts who you are meeting, what changed since last time and what to ask.
A first job
The inbox, read before you.
Hannah Berg9:14 AM
@Cedric what's still unanswered in my inbox from this week?
CedricAPP9:15 AM
Four threads. Two need you. Priya at Halcyon asked for the renewal quote on Tuesday, and the offsite venue wants a headcount by Friday. I drafted both replies from the threads. The other two are newsletters that ask for nothing. Nothing has gone out.
Look at what happened there. He read four threads without asking, because reading changes nothing. He wrote two drafts without asking, because a draft stays inside the team. Sending is the step that reaches Priya, so sending waits for Hannah's yes.
How to hand work over safely
The same steps you would take with a new hire, written down.
- 01
Connect only what the job needs
Nothing is connected until you connect it. Add the tool when a task needs it, read-only where the tool allows. You can revoke any connection from the dashboard, and the change takes effect at once. See connecting your tools.
- 02
Let him draft, and keep the send
Cedric can read anything you connected and write any draft he likes. An email, a CRM change, a refund or a post in a channel you did not ask about waits for approval. You approve, edit then approve, or hold. Silence sends nothing.
The gate only works if someone reads the request. Approving everything in two seconds keeps the rule and loses the point.
- 03
Correct it in the thread
When the answer is off, reply with the fix: "use the fiscal year, ours starts in April". The correction goes into the team's shared memory, so your colleague's next report follows it too.
- 04
Prove it by hand, then schedule it
Ask for the report three times and correct it until it needs no edits. Then say "do that every Monday at 9". The schedule approves the routine work, and any send inside it still stops and asks.
- 05
Keep approvals where the team can see them
Requests land in the channel, and anyone in that channel can approve. Put sensitive work in a channel whose members should be deciding it.
What to hand one first
Pick work that passes three tests. It repeats, so each correction pays off again next week. A wrong answer costs a re-read rather than a customer. And you already know roughly what the right answer looks like, so checking takes a minute.
The weekly numbers pass all three. So do the notes for one recurring meeting, and the "what's unanswered" sweep above. A reply to your biggest customer's complaint fails the second test. Keep that one for month two, when you have seen how he writes.
My own rule of thumb is to hand over the job you would be annoyed to receive late but never surprised by. Surprising output means the task still needs judgment, and judgment is the part you keep.
What still needs a person
Decisions stay with people. An AI employee can put the numbers, the thread and the options in front of you, and the choice between them is yours.
So does anything that depends on a relationship. He can draft the note to an unhappy customer, and whether it goes out today, and in whose name, is a person's call.
Read before you approve. Once you say yes, the email is out, exactly as if you had pressed send yourself.
FAQ
Software you give a job to. It works in your team's chat, uses the tools you connect, remembers what the team taught it, runs recurring work on a schedule, and asks before it sends, changes or spends anything.
No. A chatbot answers in its own window and forgets. An AI employee holds your tools and a shared memory, and it keeps working between messages: the weekly report, the chase that runs for three days.
Cedric will not. Anything that leaves the workspace, changes a record or costs money waits for a person's yes, and no setting turns that off. See how approvals work.
Cedric starts with 100,000 free credits, worth $50, and no card. Paid plans start at $19 a month for the whole team, with no price per seat. See pricing.
Something recurring, cheap to get wrong and quick to check: the weekly numbers, the notes for one standing meeting, or a sweep of what is unanswered in your inbox.
100,000 free credits, no card.