In the plan for AI to take our jobs, I believe we are now entering the job shadowing phase.
I am not talking about some distant future of work. I am talking about now.
For companies already working seriously with agents, assigning a personal agent to every new employee is a next-week or next-month decision. Not a five-year roadmap. That is my view of how close the next step is, not a claim that every company has already figured it out.
You start a new job. IT sets up your laptop, your email, your access … and your personal AI agent.
That is going to start sounding as normal as getting a company email address. The interesting question is what happens after someone hands you the login.
You are onboarding the agent, too
Working with my own personal agent, personally and more recently at work, is what makes this feel much closer than people realize. You explain what you are trying to accomplish. Give it context. Correct its work. Show it what a useful result actually looks like.
Essentially, you are onboarding someone. Except your name is associated with its actions, and you are trying not to spend exponentially more time teaching than it would take to do the work yourself.
That second part matters. An agent that needs an elaborate explanation for every small task can become another job. The useful relationship develops when instructions, relevant context, and corrections can carry forward within a managed system, rather than disappearing into yesterday’s chat.
It still needs supervision. Remembering an instruction does not make that instruction correct forever, and knowing how you like a report formatted does not mean it has permission to send it to a client.
Shadow, draft, review, delegate
Now imagine this being part of every employee’s first month. While you learn the company, your agent learns how to support you inside it. You both get introduced to the processes, the systems, the expectations … and the gap between the documented process and whatever everyone actually does.
At first, it shadows. Then it drafts. You review. As it proves useful on specific tasks, you delegate more.
Take preparing for a meeting. An early assignment might be to pull together the approved account notes, relevant files, and open questions into a briefing. You check whether it found the right material and whether its summary actually reflects the sources. A missing commitment or an outdated document becomes a specific correction, not just “try harder next time.”
Later, that same workflow might include drafting a follow-up or preparing an internal update for review. Sending the message is a separate permission. So is changing a customer record. Being good at one step does not automatically qualify the agent for the next.
This progression does not require AI to become capable of doing an entire job. It requires identifying work it can handle now and building a reliable way to hand that work over.
A personal agent needs a company onboarding plan
“Here is your agent. Good luck” is not much of a plan. Neither is asking every employee to invent their own access rules, instructions, and review process.
The company can provide a starting point for the role: approved knowledge sources, useful task patterns, access limited to what the employee and agent need, and clear boundaries around actions. The employee then supplies the context that a generic role description will never capture.
For an initial assignment, I would want the handoff to answer a few practical questions:
- What is the deliverable? A briefing, a draft, a calculation, or a system update are different jobs.
- Where should the information come from? Point to the approved sources and make it possible to check them.
- What does good look like? Supply an accepted example or explicit acceptance criteria.
- What can it do without asking? Separate reading, drafting, sending, and changing records.
- Where should it stop? Define who handles missing information, conflicting instructions, and exceptions.
None of that requires a hundred-page manual before anyone gets started. It requires enough clarity for a small, useful assignment to succeed without leaving everyone guessing about who authorized what.
Trust has to apply to a task, not a personality
It is easy to start trusting an agent because it is helpful, responsive, and sounds like it understands you. But the relevant question is whether it can do this particular job with the right information and checks.
A polished briefing can still use an outdated file. A well-written email can still promise something the business never agreed to. Reviewing the tone is not the same as verifying the content.
For repeatable work, give the workflow something concrete to check: a source record, an approved policy, a calculation, a required field, or a known-good example. When an exception needs human judgment, make that handoff explicit. And if an action changes a system, keep a record of what changed and a supported way to correct it.
That is how delegation can expand without turning “it did a good job yesterday” into permission to do anything today.
Delegation becomes a first-week skill
The conversation is going to move quickly beyond “Did we give everyone access to AI?” It becomes: “What work is each employee equipped to delegate from day one?”
That changes onboarding for the human as well. Delegation, reviewing someone else’s work, setting boundaries, explaining what good looks like … those stop being skills reserved for managers. They become foundational job skills.
Employers will need to teach those skills alongside the tools. Otherwise, some employees will avoid delegating anything, while others will hand over work without realizing which decisions still belong to them.
I would start with one recurring task. Establish what an acceptable result is. Review the work against that standard. Check whether the time saved exceeds the effort spent explaining, correcting, and supervising it. Expand from there when the evidence supports it.
The job-shadowing joke is funny because there is something familiar underneath it. We already know that a new colleague needs context, feedback, and boundaries before we trust them with more responsibility. An agent needs its own version of that process.
I would not be surprised if we are much closer to a new employee asking “Where’s my agent?” than most onboarding teams are prepared for.
The laptop checklist might need another line.
