An AI agent added 5 pages to a restaurant website. The owner wrote in to ask if she was missing something.
She'd asked for better visibility on her site. What came back was 5 new pages, each with a lot of language about one very small topic. Then she pasted the agent's own closing note into her email. 3 sentences about a near miss, a risk it had judged too high, and a safer method for a future session. Under it she wrote: "I don't know what this means, or what I need to do."
That's an onboarding failure. Not a model failure. Nobody had told her what she'd receive, what it would look like, or what to do with it. And that's the thing about AI onboarding: it's a real process now, for coaches and consultants and for their clients, and almost nobody has written it down.
TLDR
- Client onboarding used to collect logistics. AI onboarding collects the material a system runs on.
- The 3 Sources are the intake: what you say, what you write, what you sell. 10 to 20 pieces.
- The most expensive gap is not the input. It's nobody telling the client what will come back and how to act on it.
- Intake has a shelf life, and nothing in the stack knows the expiry date.
- 5 steps, below. Most of it is a conversation, not a form.
Onboarding a Client Used to Mean a Welcome Email
The old version: welcome email, intake form, calendar link, maybe a folder. It collected logistics. Who are you, what do you want, when are you free.
When the deliverable is produced by a system, that's not enough, because the system needs to know how you think and how you sound before it can make anything. A welcome email doesn't carry that. A form can't. Which is why every build I do starts with the same intake, and the intake looks nothing like a form.
I've written about why onboarding is usually the first workflow to automate. This post is about the other direction: onboarding a system to a person.
What an AI Onboarding Actually Collects
The intake is the 3 Sources, and it's deliberately small.
- What you say. 4 to 6 recordings. A sales call, a coaching call, a talk. Mixed, so the system hears the range.
- What you write. 4 to 6 pieces, at least 2 of them unpolished. Real client replies, not marketing.
- What you sell. 3 to 5 pieces of the offer. A module, a worksheet, the sales page, an onboarding email of your own.
10 to 20 pieces. If you've been in business 2 years you cleared that bar long ago. What comes out the other end is a Voice Print and a Reference Library, and for most people it's the first time their standard has been written down anywhere.
Then there's the part of intake nobody thinks of as intake: where information arrives. In one business I work in, client information reaches me through 4 systems. Doc comments, a ticketing tool, a project board, and direct email that splits across 4 addresses. Onboarding a system into that means picking 1 place, or the system spends its life routing instead of producing. That's not a model problem. There's no prompt for routing.
The Note the Owner Couldn't Read
Back to the 5 pages, because it's the clearest example I have of the gap that costs the most.
Let me defend the machine for a second. Its note was good engineering reporting. It named the near miss, named the risk, proposed a safer method, and deferred the work. It was written for the next session of the agent. It was not written for the person who owns the website.
What that gap cost: the coach read the same thread, said it didn't look right to him, and asked me to carve out 1 hour with the owner. 1 hour of a person, to translate 3 sentences a machine wrote about its own work.
I have 50+ skills built and rebuilt. Not one of them, as far as I can remember, was built to check whether the person on the receiving end can act on what it produced. That's now step 4 of my onboarding, and it's the step that was missing from every version before it.
Intake Has a Shelf Life
An owner left 6 comments on a brand document I'd built for her. 0 were about the writing. All 6 were facts.
The document called her parking easy. It hasn't been lately, and it's hurting her lunch trade. A weekly live event the document placed on the calendar doesn't happen at that service anymore. The supplier list had turned over. A service the document described as running was still on her roadmap.
Nothing came back about tone. The document read like her place. It just read like her place 4 months ago, when she'd told us all of that at intake. That information has a shelf life, and nothing in the stack knew the expiry date. The only thing in the process that knew was the owner, reading her own document.
So AI onboarding isn't a one-time event either. It needs a re-intake, and I don't yet have a clean rule for when. Right now the obvious move is a freshness check before anything ships, and I don't know what it would check the old answers against. That's the honest limit of it.
A 5-Step AI Onboarding for Coaches and Consultants
Here's the process I run now, whether I'm onboarding a client to my system or helping a coach onboard their own clients to theirs.
- 1. Collect the 3 Sources. 10 to 20 pieces. Send a list, not a form. Most people have it in a folder within an hour once they know recordings count.
- 2. Pick 1 place information arrives. One channel for feedback, one for files. If it's 4, the system will spend its life routing.
- 3. Build the Voice Print and show it back. Before any deliverable. "Here's how the system thinks you sound." The client corrects it in 10 minutes and you've saved 10 rounds of edits.
- 4. Tell them what comes back and what to do with it. In plain words. "You'll get a document. Read it line by line. Comment on facts, not tone. Pick 1 of the 3 hook options." That step would have saved the owner with the 5 pages an hour of her life.
- 5. Book the re-intake. 90 days. Parking, suppliers, offers, prices. Whatever changed, the system needs to know before the client finds it in a document.
That's the process, and it's most of what the Encode step is in practice. The technology is the easy half.
If any of this sounds like the onboarding you're running, or the one you've been on the receiving end of, it's probably worth a conversation. Book a free call, 30 minutes, and I'll tell you honestly what your intake is missing. Or run step 1 on yourself: send me 1 piece of your work and see what a system built on it produces within 48 hours.
FAQs about AI onboarding
What is AI onboarding?
The process of collecting the material an AI system needs to produce work in someone's voice and to their standard, then telling that person what will come back and how to act on it. For a coach or consultant it means gathering 10 to 20 pieces across the 3 Sources, choosing 1 channel for information, reviewing the Voice Print, and scheduling a re-intake.
What should a client provide to onboard an AI content system?
4 to 6 recordings, 4 to 6 written pieces with at least 2 unpolished, and 3 to 5 pieces from the paid offer. No podcast or content library is required. Any 1 source is enough to start, and the rest can be added as the system runs.
Why do AI deliverables confuse clients?
Because the output is often written for the system or the operator rather than for the person receiving it. An agent's closing note can be excellent engineering reporting and still leave an owner asking what she is supposed to do. Onboarding has to include a plain-language brief on what arrives and what action it needs.
How often should intake information be refreshed?
Every 90 days is my current default, and sooner for anything operational like hours, suppliers, prices, and services. Intake facts decay quietly and nothing in the system knows when. The client usually finds the stale fact in a finished document unless you ask first.