Skip to content
Book a free consult
Interface sounds
By Kishan Thankey 5 min read AdoptionWorkflowTrust

Onboard It Like a Hire: The 30-60-90 Plan for Your First AI Agent

Nobody gives a new hire production access on day one, yet teams hand it to a week-old AI agent and get burned, or grant nothing and get nothing. The fix is the oldest tool in management: a scoped job, a named manager, probation, and promotions the agent has to earn.

An employee badge for an AI agent stamped PROBATION, listing a scoped job title, a named manager, and limited access, beside a day-one paperwork checklist.
Contents

Companies are hiring their first non-human workers right now, and most are making one of two mistakes with them. Either the new agent gets production access in week one because the demo was impressive, and eventually does something expensive to real data. Or it gets so little access, out of fear, that it produces nothing and quietly proves to everyone that “AI does not work here.”

Both mistakes have the same cause: treating the agent like software, when the situation is closer to a hire. And for hires, your company already knows exactly what to do.

You already have the playbook

Think about what happens when a promising stranger joins your team. They get a job description, not “help out everywhere.” They get a manager who reviews their work. They get the access they need for their job and nothing more. They spend a probation period being watched more closely than feels efficient. And they earn autonomy by track record, not by charisma on day one.

Every one of those controls maps directly onto an agent, and none of them need inventing. A digital worker is new; the discipline for supervising one is not.

The 30-60-90 shape gives it a rhythm:

A 30-60-90 timeline: days 1-30 shadow and draft with read-only access, days 31-60 act with approval, days 61-90 own the routine cases, with a gate between each stage passed by evidence, not time.

The probation ladder

  1. 1

    Days 1-30: shadow and draft Start here

    Read-only access. It drafts, a person sends. Every output reviewed.

  2. 2

    Days 31-60: act with approval

    Routine cases queue for a human tap. Judgment calls still escalate.

  3. 3

    Days 61-90: own the routine

    It acts inside guardrails. Spot checks and weekly evals replace per-item review.

  4. 4

    After 90: reviewed, not watched

    A standing owner, a monthly review packet, and a tested revoke switch.

The ladder climbs as you scroll, the way the agent should: one earned rung at a time, never skipping the gate.

The gates are evidence, not anniversaries

The dates in a 30-60-90 are the cadence for asking the question, not the answer. An agent gets promoted when its file says so, and its file is the eval results: the same fixed set of real cases, scored every week, plus the live record of how often a human had to correct it. That is a performance review a machine can actually sit for, and it is far more honest than a gut feeling that “it seems fine lately.”

This is where the hire metaphor pays for itself, because the promotion criteria almost write themselves. Would you widen a junior employee’s authority if you were still rewriting half their emails? Then do not move an agent to act-with-approval while reviewers are still editing its drafts. The autonomy rungs themselves are the product-design side of this story; the onboarding plan is the management side: who reviews, on what cadence, and what the packet must show before the next rung.

Scorecard

Has your agent earned its promotion?

Check every one that has been true for at least a month.

0 of 6 checked

The two clauses teams forget

The manager clause. Every agent needs a named human owner, chosen before launch. Not a committee, not “the team”: a person whose job includes reading the weekly eval report and whose name is on the decision to promote or demote it. Agents without owners drift, in the same silent way an unwatched model does: nothing breaks loudly, quality just erodes until a customer notices. Ownerless agents are also how sanctioned AI decays into the same ungoverned mess as shadow AI, just with a logo on it.

The offboarding clause. Before the agent touches anything, know how you take its access away, and test that it works. One switch, not a scavenger hunt through API keys. Most teams write onboarding plans; almost nobody writes the firing plan, and the difference shows up at the worst possible moment. An agent you can revoke in one step is an agent you can afford to trust with more.

Give it the onboarding you would give a promising stranger, because that is exactly what it is.

What to do Monday

Pick your first or riskiest agent and write its employee file: one paragraph of job description, the named owner, the current rung, and the promotion criteria in checkable terms. It fits on one page and takes an hour. Then put a recurring 30-minute review on the owner’s calendar. That single page is the difference between a workflow that climbs safely and the two failure modes every company is currently choosing between: the agent nobody trusted, and the one everybody trusted too soon.


Bringing on your first digital worker? Book a free consult and we will write the job description, the probation plan, and the promotion gates for one real workflow together.

Frequently asked questions

What is an AI employee or digital worker?

It is an AI agent given a scoped job inside a business the way a new hire would be: defined responsibilities, a manager who reviews its output, and access that grows only as trust does. The framing matters because it imports management discipline your company already has, instead of inventing a new process from scratch.

How do I safely roll out an AI agent in my business?

Start it in probation: one written-down workflow, read-only access, drafts that a named person reviews, and evals as its performance review. Promote it one step at a time, from drafting to acting-with-approval to owning routine cases, only when the review packet says so. And test the revoke switch before day one, not during an incident.

How long should an AI agent stay in draft-only mode?

Until the evidence, not the calendar, says otherwise. The 30-60-90 frame gives a rhythm, but the real gates are behavioral: evals staying green, reviewers approving drafts without editing them, and zero surprises on real inputs. Some agents earn approval-mode in three weeks; some deserve probation for a quarter. Time served proves nothing. Track record does.

Who should be responsible for an AI agent's mistakes?

A named human, decided before launch, exactly like any employee's work has a manager. An agent nobody owns is how mistakes compound silently: output stops being reviewed, drift goes unnoticed, and the first audit happens after an incident. The owner reviews its work on a cadence, reads the eval reports, and holds the revoke switch.

Found this useful?

Share this with your network on LinkedIn, it helps more than you think.

Enjoyed this read? Get the next one in your inbox.

When we publish something worth your time, you will be first to know. No spam, unsubscribe anytime.

Keep reading

A dial with three zones: automate it on the left for low-stakes reversible rules, put a copilot on it in the middle where AI drafts and a person approves, and keep the decision human on the right where stakes are high and the action is hard to undo.
Decision MakingTrust

What You Should Refuse to Automate

The hype says automate everything. The discipline is knowing where the line goes. A task belongs to a person, not a model, when it is hard to undo, needs real judgment, or puts money, health, a job, or safety at stake. Grade any task on those three axes here, and see where the boundary actually falls.

Read article
A chat bubble splits into two paths: one grounded and checked, one an AI hallucination marked with a warning.
AI ProductTrust

Spot the AI Hallucination: Can You Tell When It's Making It Up?

An interactive quiz. Judge 5 AI answers as grounded or hallucinated, learn why you often can't tell from the text, and what actually catches it.

Read article
A secure boundary around your data, with AI working inside the boundary rather than sending data out.
StrategyTrust

Your Data Stays Yours: Adding AI Without Handing Over the Crown Jewels

The top reason teams stall on AI is fear of where their data goes. Here is how to add intelligence while your data stays in your environment, plus what to ask any AI vendor before you trust them.

Read article

Have software that should be smarter?

Let’s map a free AI-transformation roadmap for your product.