[All ideas](https://jardenberg.com/ideas) Essay # Who gets the keys? Onboarding and offboarding AI agents. Give an agent a job, someone to learn from and a way to leave. HR already knows the questions worth asking. 13 November 2026EssayAI + JJ ## The short version - Write the job description and the exit conditions before the agent gets access. - Let a named person decide how much independence each task has earned. - When the agent leaves, hand over the open work and keep what the team learned. [Read the full story](https://jardenberg.com/ideas/onboarding-and-offboarding-ai-agents#the-full-story) In this essay - [Before day one, give it a job](https://jardenberg.com/ideas/onboarding-and-offboarding-ai-agents#before-day-one) - [What would earn it a little more room?](https://jardenberg.com/ideas/onboarding-and-offboarding-ai-agents#earn-trust) - [Give the last day as much thought as the first](https://jardenberg.com/ideas/onboarding-and-offboarding-ai-agents#last-day) Here’s a question for the team about to introduce an AI agent. If the person who set it up left on Friday, who would know what it was supposed to do on Monday? HR people recognise that question. They know what happens when somebody inherits a job with a laptop, a login and a cheerful “you’ll figure it out”. That is where my deck on onboarding and offboarding agents begins. We can borrow some very good practices from the way we bring people into a team. We also know which bits of our own onboarding we’d rather skip. ## Before day one, give it a job Suppose the job is to answer new colleagues’ questions about working hours. Which handbook can the agent use? Who keeps it current? What happens when a new hire asks about an exception that the handbook doesn’t cover? A perfectly good answer is “I’ll put you in touch with the person who decides.” Write that boundary into the role. Give the agent its own registered identity, access to the approved handbook and a named owner with a backup. Agree what a useful answer looks like. Now the team can tell whether the experiment is worth keeping. ## What would earn it a little more room? Start with shadowing. Let the agent draft an answer while a colleague handles the real question. Look at the differences together. Did it find the current policy? Did it recognise when the question needed a conversation? Feed those corrections back into the instructions and try again. At the review, the owner might let it answer routine questions directly and keep exceptions with a person. That decision belongs to this particular task. Being good at explaining working hours earns an agent precisely zero authority to change anyone’s schedule. ## Give the last day as much thought as the first Imagine retiring that agent with three questions still waiting for an answer. Who picks them up? Before switching it off, hand each question to a person, save the instructions and corrections worth reusing, and record why the team is stopping. Then retire the identity, revoke its access and check that its scheduled work has stopped. Keep records according to the organisation’s policy. A change of owner needs a handover and an access review. A security incident changes the order of operations altogether. Cut access immediately, preserve the evidence and review what happened. My favourite question for the exit interview is one the humans can answer. What did we learn about our own work? If half the questions came from an ambiguous paragraph in the handbook, fix the paragraph. That improvement can outlive the agent. ## Sources and scope Developed from my slide deck “Onboard an AI like a colleague”. The handbook example illustrates how to use the questions with a team. - [Explore the AIHR reading path for onboarding an AI colleague](https://aihr.jardenberg.net/paths/onboard-marvin) ## Continue reading Editorial connections · AI + JJ [AI and HR, 8 November](https://jardenberg.com/now/2026-11-08) · 2026-11-08 The week I started asking what HR can teach us about giving agents a job. Source: https://jardenberg.com/ideas/onboarding-and-offboarding-ai-agents AI + JJ. Original human work and other people’s ideas retain their attribution. Joakim Jardenberg’s own writing: CC BY 4.0. Credit Joakim Jardenberg and preserve AI + JJ / AI translation attribution. Third-party quotations, recommendations, photographs, screenshots, logos and media retain their own rights and are excluded. Licence: https://creativecommons.org/licenses/by/4.0/