AI Agent Client Onboarding: Six Answers to Collect Before You Install Anything

Most advice about onboarding and AI is about using AI to onboard your clients: the welcome email, the intake form, the document chase. This is the other direction. A client has said yes to an AI employee, and now you have to get that employee ready to do a real job inside their business without embarrassing either of you.

That step decides whether the setup survives its first month. The install is the quick part. What goes wrong is everything the agent was never told: who approves what, which words the owner hates, which accounts it must never touch, and who reads the Monday report. None of that lives in the software. It lives in the owner's head, and your onboarding is the only chance to get it out.

I am covering the setup side of this live on Saturday, October 31 at 10:00 AM Eastern, in a free session on setting up and selling AI agent employees. This post is the part that fits on a page: the intake I would run before installing anything.

What is AI agent client onboarding?

It is the short stretch between "yes" and the first scheduled run, where you collect everything the agent needs to behave like a person who has worked there for a month. It is not onboarding software for your customers, and it is not a welcome sequence. It produces one thing: a written role the agent reads at the start of every run.

If you can write the job down, you can hand it to an agent. If the owner cannot describe the job, no amount of setup fixes that, and finding out in week one is far cheaper than finding out in week six.

What should you collect before installing anything?

Six things. Each one maps to a section of the role file the agent will read, so nothing here is paperwork for its own sake.

1. The one job, in a sentence a stranger could check

"Handle marketing" is not a job. "Draft three posts every weekday morning from the owner's last five emails and leave them in a review folder" is. Ask the owner what they would be annoyed to do themselves tomorrow, then narrow it until the output is a thing you can point at.

A good test: could someone who has never met the owner look at the result and say whether the job was done? If not, the job is still too vague to schedule.

2. The sources of truth, with who owns each

Every number and claim the agent produces needs somewhere to come from. Ask for the pricing page, the current offer, the customer list, the calendar, the sales sheet. For each one, write down the person who keeps it correct.

This is the question most setups skip, and it is why agents confidently repeat last year's price. If a source has no owner, the agent should treat it as unverified and say so, and the owner should know that is what they signed up for.

3. The approval line

Decide, in writing, what the agent may do alone and what waits for a human. Publishing, sending, spending, deleting and replying to a customer are the usual four. My default is that anything customer-facing is staged for a person and nothing irreversible runs unattended.

I wrote a full breakdown of where to draw this in what to gate and what to let run unattended. For onboarding, the point is smaller: get the owner to say the line out loud, because they will move it later, and you want the change to be a decision rather than a drift.

4. The voice and the stop list

Ask for three pieces of writing the owner is proud of and three they would never send. Then ask for the words they refuse to use. The refusals are worth more than the examples, because they are specific and checkable.

Put the stop list in the role file as plain sentences. When the agent gets something wrong, you add one more sentence rather than rewriting the prompt. That habit is the whole subject of stopping an agent from repeating a mistake.

5. The access you will and will not take

List every account the job touches, and for each one decide the least access that works. Read-only where reading is enough. A separate login rather than the owner's own wherever the platform allows it. Nothing that can move money.

Write down what you were not given as well. "No access to the bank, the domain registrar, or the owner's personal email" belongs in the role file and in the handover note, so the next person does not assume otherwise. The permission model in the Claude Code permissions documentation is a useful reference for how to express allow and deny rules rather than relying on a sentence the agent might misread.

6. The reader

Every report needs a named human who reads it and a day they read it. An agent that writes a brilliant Monday summary to nobody is a failed setup that looks like a working one.

Ask who reads the output, on what day, and what they do when it is late. Then ask who covers when that person is away. Three answers, one line each.

How do you turn the answers into a role file?

Give each of the six answers a heading in a single text file and keep it short enough to read in a minute. The job goes first, then sources, then the approval line, the voice and stop list, the access notes and the reader. Agents read this file at the start of every run, which is exactly what the project memory feature is for, described in the Claude Code memory documentation.

Two habits keep the file honest. First, every line should be something the agent can act on; delete anything that is only a mood. Second, leave a section at the bottom called corrections, and add one sentence there every time the owner pushes back. After a month it is the most valuable part of the setup, and it is the part that makes the handover to the next person possible. I go into why a text file beats anything fancier in memory between runs.

What should the first week look like?

Keep the first week boring on purpose.

• Day one: install, run the job once by hand, and read the output together with the owner. Change the role file, not the agent.

• Days two to four: run it on a schedule, but stage everything. Nothing is sent or posted by the agent.

• Day five: review the week with the owner. Move at most one thing across the approval line, and write down why.

Resist the urge to turn on a second job. A client who trusts one working employee will ask for the next one. A client who watched three half-working ones will not.

What do you hand the client at the end?

A one-page handover: the role file, the list of accounts and the access level for each, the person who reads the output, and how to ask for a change. If you are selling this as a service, the handover is also what makes the monthly part honest. I wrote about the pricing side in how to sell AI employees to clients, and the intake above is what justifies the setup fee: it is real work, and the client can see the page it produced.

FAQ

How long should client onboarding for an AI agent take?

For one job, plan on one working session of about an hour to get the six answers, and a few more days of staged runs before anything is trusted. If the answers take much longer than an hour, the job is probably too broad and should be split.

What if the client cannot describe the job?

Do not install anything yet. Ask them to walk you through the last time they did it by hand, and write down each step. A job that cannot be narrated cannot be scheduled, and it is better to find that out before you have been paid to automate it.

Should the agent use the client's own accounts?

Where the platform allows it, give the agent its own login with the least access that works. Where it does not, keep the job read-only or staged so a person performs the final action. Never take access you cannot explain in a sentence.

Do I need special software for the intake?

No. A shared document with six headings is enough. The value is in asking the questions and writing the answers where the agent will read them, not in the tool you use to collect them.

Where to go from here

If you want to set one of these up for a client, the eight open source AI Employees are on GitHub, and the free session on October 31 walks through setting them up in a desktop agent app and packaging the setup as a service. You can register on the event page, and the Agent Ops Club has a free plan if you want to look around first.