Blog
How to choose an AI coworker: 8 questions to ask
A buyer's checklist for an AI coworker: where it lives, what it can reach, who approves its actions, whose memory it keeps, how it is priced, where data goes.
Benjamin Ben HattarEvery demo of an AI coworker looks good. It drafts an email, summarises a thread, answers a question about last quarter. The differences show up in week three, when it has access to your CRM, three people are asking it for things, and one of them asks it to send something to a customer.
So judge it by what happens in week three. Below are the eight questions I would ask any vendor, including us, with what a good answer sounds like and how Cedric answers each one.
- Pick one that lives where your team already talks, so nobody has to open another app.
- Ask what it can reach, and whether each connection can be read-only and revoked in one place.
- Ask who approves an action before it happens. "You can review it after" is not an answer.
- Check whether memory is shared by the team or locked in one person's account.
- Price it per workspace against per seat for your real headcount, and ask where the data lives.
The checklist
Eight questions. Ask them in this order, since the early ones rule vendors out fastest.
- 01
Where does it live?
An AI coworker the team has to go and visit gets visited by one person. Look for one that joins the chat you already use, where anyone can ask in any channel and everyone sees the answer.
Cedric joins Slack, Discord or Telegram as a member of the workspace, not of one person's account. You can also start on your own in the browser.
- 02
What can it reach, and how narrowly?
Ask how many tools it connects to, then ask the better question. Can each connection be read-only? Is it scoped to what a task needs? Can you revoke it in one click? Can you see what it read?
Cedric connects to over 3,200 apps, starts with access to nothing, and asks for a connection when a task needs one. Connections can be read-only where the tool allows, revoked from the dashboard, and every read shows in the activity view. Tools that speak MCP join the same way.
- 03
Who approves an action before it happens?
This is the question that matters most. Reading and drafting are harmless. Sending an email, changing a CRM record or issuing a refund is not. Ask which actions stop for a person, whether that stop can be switched off, and what happens if nobody answers.
Cedric asks before he sends, changes or spends anything. The card names the exact change. Anyone in the channel can approve, edit or hold it, and silence sends nothing. The gate is not a setting. How approvals work walks through it.
- 04
Whose memory is it?
A personal assistant remembers what you told it. A coworker should remember what the team told it, so a correction made on Monday holds when a colleague asks on Tuesday. Ask too what stays private, and whether you can read and delete what it knows.
Cedric keeps one shared memory for the workspace. Your direct messages with him stay private, and so does what he learns about you; a fact about the company goes to the shared memory, and he says when he saves one. The Brain page lists every memory, and delete means gone.
- 05
Can it do work nobody is asking for right now?
The weekly report, the daily check, the client onboarding that takes five days. Ask whether it can run on a schedule, start on an event, and wait for people without someone keeping a chat open.
Say "every Monday at 9" in the channel and Cedric runs it. Multi-day jobs run as workflows that ask, nudge and resume when replies arrive.
- 06
Does it come to meetings?
Half of what a team decides is decided on calls. A coworker that misses them works from half the story. Ask whether it joins as a visible participant and what it does with the action items.
Cedric joins by call link or calendar invite, answers when addressed by name, and queues actions for approval after the call.
- 07
How is it priced?
Per seat means the bill grows with headcount, and teams end up rationing seats, which defeats the point of a shared coworker. Per workspace means everyone can ask. Ask too what the free tier includes and whether features are gated by plan.
Cedric has no seats. Every workspace starts with 100,000 free credits, worth $50, with no card. Plans start at $19 a month for the whole team, and no feature is locked to a plan.
- 08
Where does the data go?
Ask who runs the service and where, whether your data trains anything, how long it is kept, and how you delete it. For a European team, ask about GDPR by name.
Cedric is built in Zurich and runs on infrastructure in Europe. Your data is never sold and never used to train models, and the service is GDPR compliant. Details on the security page.
Per seat against per workspace
Plain arithmetic, assuming a per-seat tool at $20 per person per month. Cedric's cost depends on the work done, not on headcount, so the right column is the starting plan, shared by everyone.
| Team size | Per seat at $20 | Cedric |
|---|---|---|
| 3 people | $60 a month | From $19 a month, everyone included |
| 10 people | $200 a month | From $19 a month, everyone included |
| 25 people | $500 a month | From $19 a month, everyone included |
Heavier use needs a bigger plan. Plans differ only in monthly credits, never in features.
How to run the trial
Skip the demo tasks. Give it one real job your team does every week, with the real tools connected read-only, and have two or three people ask it things in the shared channel for a week.
Then look at three moments. The first time it needed a connection, did it ask or assume? The first time it proposed an action, could you see exactly what would happen? The first time someone corrected it, did the correction hold for a colleague the next day?
If you want a first job, putting the Monday report on a schedule is a good one. It is easy to check against a week you already know.
What the checklist cannot settle
No answer to these questions replaces reading the approval card before you tap yes. A gate that gets approved in two seconds every time is still a gate on paper only.
Some choices are yours regardless of the tool. Which channels it sits in decides who can approve. Whether to bring it into customer calls depends on your consent rules. And a regulated or client-owned system may need a contract check before you connect it.
FAQ
An AI agent that joins your team's chat, reads the tools you connect, and does work for anyone on the team, with one shared memory. The good ones ask before they send, change or spend anything.
It depends on two things: whether the connection can be scoped and revoked, and whether every change waits for a person. With Cedric, both are true, and read-only is a sensible start.
For a tool the whole team shares, per seat works against you, because people stop asking. Cedric is priced per workspace: one plan, everyone included.
Yes. Cedric is built in Zurich, runs on infrastructure in Europe, and never sells your data or uses it to train models. The security page and the privacy policy have the detail.
With Cedric, yes. Every workspace starts with 100,000 free credits, worth $50, and they never expire.
100,000 free credits on sign-up, no card. Plans from $19 a month for the whole team.