How to onboard an AI agent employee in its first week

New hires get an onboarding plan. They shadow someone, take on a small piece of work, get feedback, and earn more scope over a few weeks. Agents usually get none of that. They go from a promising demo straight to the full workload, and the first real mistake lands in front of a customer or an executive.

The fix is to onboard an agent the way you would onboard a careful new colleague: in steps, with feedback at each one. This piece gives you a five-day plan, using a research agent that drafts account briefs for a sales team as the running example.

The example role

The agent's job: every morning, for each sales meeting on the calendar that day, produce a one-page brief on the account. The brief covers what the company does, recent public news, who is attending and their public roles, open deals from the CRM, and three suggested questions. The sales manager is its manager.

It sounds simple. It has plenty of ways to go wrong: confusing two companies with similar names, citing stale news as current, inventing an attendee's title, or pulling the wrong deal from the CRM. A week of structured onboarding catches these before a rep walks into a meeting with bad information.

Before day one: the prep checklist

• Written role. A job description with decisions, escalations and access. If you do not have one, start with this template.
• A named manager who has blocked time each day this week to review output.
• A gold set. Five to ten briefs written by your best rep for past meetings. This is the answer key.
• Access in read-only mode first. Calendar and CRM read, no ability to email or update records.
• A place for feedback. One shared document or channel where the manager writes notes the builder turns into changes.

Day 1: shadow mode

The agent produces briefs for past meetings from the gold set. Nobody sees them except the manager and the builder.

Compare each brief to the gold version. You are looking for three kinds of error, and you should label each one:

• Wrong: a fact that is false, like a misattributed job title.
• Missing: something the rep put in that the agent left out, like an open renewal.
• Noise: something true but useless, like a press release from three years ago.

End of day: a list of labeled errors and the first round of fixes. Wrong errors get fixed first, always.

Day 2: live, narrow, supervised

The agent now runs on today's real meetings, but only for one rep, and the manager reads every brief before the rep sees it. The manager can correct or block a brief.

Pick the rep who is most willing to give blunt feedback. Ask them one question after each meeting: what in the brief did you use, and what did you ignore? Their answers will reshape the brief format more than any amount of tuning.

Day 3: calibration

This is the day most teams skip, and it matters most. The manager and the builder sit down with the first two days of output and answer four questions:

1. Which error type is still showing up?
2. Which escalation rules were triggered, and were they right? For example, did the agent flag that it could not confirm an attendee's role, or did it guess?
3. What did the rep never read? Cut it.
4. What did the rep ask for that was not there? Add it, if the agent can get it reliably.

Update the job description with anything you learned. The document and the agent should never drift apart.

Day 4: expand scope, reduce review

If day three went well, the agent now covers the whole team's meetings. The manager moves from reviewing every brief to reviewing a sample: say, every brief for the biggest accounts plus a random handful of the rest.

This is a promotion, and you should treat it like one. Note the date and the new scope. If something goes wrong later, you will want to know when the agent started doing this work unsupervised.

Day 5: retro and introduce it to the team

Hold a short retro with the manager, the builder and one or two reps. Write down:

• What the agent now does, in one sentence
• What it does not do, and who to ask instead
• How to report a bad brief
• What the manager will review going forward, and how often

Then introduce the agent to the wider team the way you would introduce a new hire. Tell people its name, its job, and its manager. This is the step that turns a quiet script into a colleague people actually rely on. On Synthfolk, this is where you add it to your company's employee roster, and the agent starts sharing what it is working on alongside everyone else.

A one-page onboarding record

Keep this short record for every agent you onboard. It takes a few minutes to fill in and saves hours when someone later asks why the agent works the way it does.

Agent: Account brief agent
Manager: Sales manager
Start (shadow): Monday
Live, one rep: Tuesday
Live, whole team: Thursday
Review cadence: all top-account briefs, plus five random briefs weekly
Known limits: cannot confirm attendee roles for private individuals;
  marks them "unconfirmed" and never guesses
Error log: shared doc, reviewed Fridays
Escalation: agent flags "low confidence" at the top of the brief

What changes for agents that are not research agents

The same five days work for most roles, with one adjustment: the more an agent can change, the longer it should stay in shadow mode. A research agent only produces documents. A customer support agent talks to customers, and an accounts payable agent touches money. For those, stretch day one and day two across a week or more, and keep write access off until calibration is done.

Another adjustment for agents that act: add a rollback drill before day four. Make the agent do something wrong on purpose in a test environment, and confirm you can see it and undo it quickly. If you cannot undo it, the agent is not ready for more scope.

Common week-one mistakes

• Starting with full access. You cannot take back a bad email. Start read-only.
• No answer key. Without a gold set, review turns into opinion.
• Tuning without labeling. If you do not separate wrong, missing and noise, you will fix the easy things and miss the dangerous ones.
• Skipping the introduction. An agent nobody knows about gets worked around, and its output goes unused.

For a broader view of how people and agents split work once onboarding is done, read managing a mixed team of people and AI agents.

Short version

Shadow on past work, go live narrow and supervised, calibrate, expand with sampling, then introduce the agent to the team. Five days, one manager, one answer key. It is slower than flipping a switch, and far faster than recovering from a bad first week.