The Trust Boundary
Every job you hand to AI has a line in it: what the model may do on its own, and what you check before it counts. This reading shows you where that line sits, how to decide where to draw it, and how it moves over time.
Time to read: about 8 minutes
The line exists whether you draw it on purpose or not, and drawing it on purpose is the skill.
Where the line is
One real handoff, taken apart.
Where to draw it
One question that sorts almost everything, and what skipping it costs.
How it moves
How trust is earned, and how it slips.
Thirteen short pages, one idea each. Use Next below, or the arrow keys on a laptop.
"I wrote none of these."
At ForwardFest in August 2026, the room watched a planning call get recorded. Ninety minutes later, three things were waiting: a recap email in the drafts folder, follow-up tasks in the task list, and notes for the file.
"I wrote none of these. I reviewed all three and hit send. That took four minutes."
Jay Fontanini, ForwardFest 2026
The system that made them could draft, but it could not send, delete, or change the permanent record. Jay did those parts himself.
That split, between what it could do and what stayed with him, is a trust boundary. Every handoff has one.
One boundary, two sides
The line has two sides, and you own both of them.
Does it suggest or act? Suggesting means it drafts and waits. Acting means it sends, files, or commits on its own.
A permission you set.
Do you review it or let it stand? Reviewing means you check it before it counts. Letting it stand means you accept it as it is.
A judgment you make.
Apply it to your own sales call. Claude hands you three things, and each one sits in a different place.
| What it made | The model | You |
|---|---|---|
| The recap email to the broker | Suggests. It waits in drafts. | Review, then send. |
| Your follow-up tasks | Acts. It adds them to your list. | Let them stand. Fix one if it's off. |
| Notes for the client file | Suggests. It drafts them. | Review before they become the record. |
Guard what leaves, not what it knows
The natural instinct is to protect yourself by telling the model less, but that usually backfires. A model that knows less about you, your clients, and your history does worse work, and you spend the savings fixing it.
The control that matters is at the exit: what leaves your hands, and who approves it before it does. In the ForwardFest example, the system knew a great deal and could send nothing.
Give it the context it needs to do good work. Put the boundary where the work goes out into the world.
What you're allowed to share is still set by your company's rules and your clients' confidentiality. Module 6 covers that. This page is about where the checkpoint sits once the context is in.
Who finds out?
One question sorts almost every handoff. If it's wrong and nobody catches it, who finds out?
You find it, and you fix it in minutes. Let it stand, and save your attention for the work that needs it.
Your follow-up task list.Embarrassing, but recoverable. Let it draft, and review before it goes out in your name. Your credibility is at stake, not the model's.
The recap email to the broker.A price, a promise, a date, a filing. Let it prepare. You validate and decide, because you're the one committing, not the model.
The renewal date or rate inside that email.The riskiest line is the shortest
A commitment isn't about how hard the work is. It can hide in one sentence of an otherwise routine email.
Four sentences, and three of them are safe to send as drafted. The highlighted one is a promise made in your name. Review the email, and decide that sentence.
Consequential work still gets delegated. Let it do the assembly and the drafting, so you arrive at the decision prepared. Keep the decision.
What skipping review costs
From the spring 2026 cohort: an executive built a summary for a C-level audience by talking it through with AI, and skipped the review to save time. The sponsor challenged it. It had pulled the wrong context and mischaracterized the status of several workstreams.
The fix took far longer than the review would have, and most of the cost was in the relationship.
Module 1's ratio said this in advance: an hour of AI work stands in for ten, and still owes about two hours of your review. Those two hours aren't overhead; they're the part of the job you kept.
It can't check its own work
Also from spring: a project used AI to plan, build, and test itself, with no human expert in the loop. The model created two yes-or-no fields where one would do. Because the same model was doing the testing, it never noticed the contradiction it had introduced. A senior developer found it weeks later.
Asking the model to review its own output is useful, and you should still do it, but it isn't the same as your review. The model that made the mistake is the one least likely to see it.
When the answer to "who finds out?" is someone outside or a commitment, the review has to be yours.
Make the work easy to review
Review goes faster when the work shows you where to look. Two spring participants wrote standing instructions that do exactly that. Add them to the Project where you delegate.
Before you proceed, disclose all assumptions you are currently working with. Flag anything you inferred that I didn't explicitly state.
Alex, spring 2026 cohort
When you include any numerical data or statistics in your output, cite the source inline, either by bolding the figure with the source name or by adding a notation. Do this throughout, not just in a references section at the end.
Vinnie, spring 2026 cohort
The first shows you what it guessed. The second shows you which numbers came from somewhere and which it made up. Both turn review from rereading everything into checking the flagged parts.
Onboard it like a new hire
Think about how you'd bring on a capable chief of staff. You don't hand over everything on day one. You work through a cycle, and each step lines up with Elements you are building.
Onboard
Memory and Boundary. Context to draw on, and the access you grant.
Scope
Boundary. A real job, with its side of the line and yours.
Review
Surface. The work comes back where you look, and you check it.
Feedback
Loop. Corrections kept where the next round will find them.
The onboarding cycle: Mike Gardon, ForwardFest 2026. The five Elements were tested against this cycle when they were named.
Trust doesn't come from onboarding it well or scoping it clearly, although both matter. It's built at step three, every time you check the work and it holds up. The feedback step is what lets the next round start from a better place. The one Element the cycle leaves out is Cadence, because a new hire shows up on their own. Your system has to be told when, and that is Module 4.
Trust is extended in steps
As the reviews keep holding up, you can move the model's side of the line, one step at a time. Here's the same sales call, four steps deep.
Read
It reads your call notes and briefs you. It changes nothing.
Draft
It drafts the recap. You edit it and send it.
Send, inside limits
It sends the meeting confirmation on its own, and nothing else.
Act on its own
It files every recap to the client record without asking.
Progressive trust: Claire Vo, Lenny's Podcast, March 29, 2026
Most of what you delegate in this module lives at step two. You don't hand a new hire your passwords on day one. You delegate something small, see how they do, and expand the role.
Trust can slip
A boundary drawn once doesn't stay true. Models are updated, your work changes, and a job that was safe to let stand in March can start producing something different in June without announcing it.
The spring cohort noticed something related. The more people used these tools, the lower their trust scores went. Nobody rated a 1 or a 10. The drop wasn't disillusionment. It was expertise: they had seen enough to know where to look.
You're never fully done checking. Module 4 turns that into a habit you can keep: when a job runs without you, you check a sample on a schedule, and you write the boundary down before you switch it on.
The whole pattern
Find the line
For each thing you hand off, ask whether it suggests or acts, and whether you review or let it stand.
Ask who finds out
Only you, someone outside, or a commitment. Review in proportion.
Make review easy
Standing instructions that flag guesses and sources.
Move the line on evidence
One step at a time, after the reviews hold up. Keep checking.
Take the task you brought from Module 2's homework. Write down both sides of its line, and answer the question: if it's wrong and nobody catches it, who finds out? That's where this week's hands-on starts.
Building With AI: The Four Ds in Practice, especially the Delegation section
The Five Elements Spine, the Boundary tab