Start AI Coworkers With Read-Only Access
Treat the bot like a new hire, and save full responsibility for last.

Y Combinator · The Rise Of AI CoworkersRoman Ugarte, who works on Grockbot, which makes persistent AI coworkers, says no one has a silver bullet answer on bot permissions, but the simplest starting point is to treat a bot like a person.
Hire it, give it its own computer and credentials scoped to its role and seniority, then expand access as it earns trust. The interviewer noted to Lauren, an engineer who works with bots in Slack threads, that she had gone from more than 2,000 software changes a month to more than 4,000; she agreed and described her workflow.
Start with step one, read only. Pick one small, annoying stream of questions, give the bot read access and no ability to act, and watch whether it does the work the way you expect.
Inside his company, the data science team did that with Databot. It lives in Slack, where requests already arrive, has read access to the database, and answers questions like whether a feature idea needs a quick analysis. A procurement team built a similar shared vendor workflow bot for the whole team.
Move to step two only after the answers prove reliable. That is when the bot can take action and carry a task end to end. A different design is Quartermaster, an open-source YC project, which inherits the permissions of the user it assists in each Slack channel, so a manager and an engineer would grant very different powers.
Save step three for later. That is a bot that owns an entire job responsibility and keeps watching an outcome on its own. Ugarte says the field is still quite far from that, which is why step one matters: start small, visible, and read only.
Questions people ask
What is a safe first permission to give an AI coworker?
Give it the smallest permission that still lets it be useful, such as read access to one database or inbox, and put it where colleagues can see its answers. Keep write access, spending power, and customer facing actions turned off until its answers are consistently correct.
How do I pick the first job for a bot?
Look for repeated questions that require lookup but not judgment, like data pulls, status checks, or policy answers. If success is easy to verify and mistakes are low cost, it is a good candidate for a read only bot.
When should I let the bot start taking action?
When the bot answers correctly without correction, and colleagues start relying on it without double checking every reply. Then try one narrow action with clear limits, such as drafting a ticket or updating a record, before widening its role.
