Your Reviewer Will Stop Reading. Design For It.

Quick answer: Human sign-off is a necessary control and an insufficient one, because review degrades through success. Once AI output is mostly right, reading it carefully stops feeling necessary and approval becomes a reflex within weeks. A control that depends on sustained attention quietly stops working while still reporting green, so the limits have to hold when nobody is reading.
Key Takeaways
- ●A person signs off on every draft is the most common AI safety claim in construction, and on its own it is the weakest.
- ●Review degrades through success, not through failure. The better the output, the faster approval becomes a reflex.
- ●Design the control for the reviewer who has stopped reading, because in three weeks that is every reviewer.
- ●Three things make sign-off hold: a threshold that stops on its own, a reversal path, and an approval that costs something to give.
- ●If a control only works while someone is paying attention, it is not a control. It is a hope with a button.
The Most Popular Answer to AI Risk
Ask any construction software company how they keep AI safe and you will get the same sentence back. A person reviews it. Nothing goes out without a human signing off.
It is a good instinct and it is where the thinking usually stops. Sign-off is one control, doing one job, and the job it does is narrower than people assume.
The problem isn't that reviewers are careless. It's that review is a control that gets weaker the better your AI gets.
Review Degrades Through Success
Consider what happens when the output is genuinely good. Week one, the PM reads every draft closely, because they do not trust it yet. Week two, they skim, because the first forty were fine. Week four, they approve from the subject line.
Nothing went wrong to cause that. The decay is caused by the system working. A reviewer who sees a long run of correct items is learning an accurate lesson about the base rate, and acting on it rationally.
This is well understood anywhere people monitor reliable automated processes, and construction has no immunity. The dangerous part is that the record still shows an approval on every item. Your dashboard reports complete oversight at exactly the moment oversight has ended.
The Three Things That Make It Hold
Call it the Tired Friday Test. Design every control so it still works when the approver is tired, at the end of the week, approving from a phone in a truck. If it only works with a fresh, attentive reader, it does not work.
1. A threshold that stops on its own. Above a dollar value, outside a date window, against an unresolved revision, the action halts without anyone needing to catch it. The machine refuses rather than asking nicely.
2. A reversal path. The question is not whether a bad approval gets through, because it will. The question is whether it is recoverable. An action you can pull back within an hour is a different risk from one that left the building.
3. An approval that costs something. Not friction for its own sake. Attach the context that makes the decision real: what it read, what changed, what it is tied to. Approving a sentence is a reflex. Approving a sentence next to the revision it was based on is a decision.
| Control | Works when the reviewer is sharp | Works when the reviewer is tired |
|---|---|---|
| Human sign-off alone | Yes | No |
| Threshold that halts | Yes | Yes |
| Reversal path | Yes | Yes |
| Context attached to the approval | Yes | Mostly |
| Training and reminders | Yes | No |
Spend Attention Where It Is Worth Something
The counterintuitive move is to approve less, not more. Every item in the queue spends a finite budget of reviewer attention, and the low-stakes items spend it fastest because there are so many of them.
Sort actions by two questions: how expensive is this to reverse, and does it commit us to somebody outside the company. Anything cheap and internal should happen on its own with a record. Anything expensive or outward-facing stops and waits.
A queue of six things that matter gets read. A queue of two hundred gets cleared.
One Thing to Try This Week
Take whatever approval queue you already run, with or without AI in it. Pull the last fifty approvals and look at the time between the item arriving and the approval landing.
You are looking for the shape of the distribution rather than the average. A healthy queue has a spread, because some items genuinely needed thought. If most of your approvals land within a few seconds of arriving, you already have a rubber stamp, and adding AI to that queue will increase its volume rather than its risk awareness.
Fix the queue before you point AI at it. Software enforces a system; it does not supply one.
Ready to take on more premium projects without adding to your overhead?
Lock in your flat-rate, unlimited-user fee with ServiceIQ today.
Start your free trialFrequently asked questions
Keep reading

AI Harness
An AI harness is the control layer around AI on a live job: what it can touch, what it may do unasked, who signs off, and what proves the decision later.

Harness vs Assistant
An assistant hands you a draft and stops. A harness is permitted to file, send and log inside limits. Here is how to tell which one you are being sold.

RFIs, Submittals, and Change Orders
Three different documents, three different jobs, and most disputes trace back to one being late, vague, or missing. Here's what each one actually does.

You Can't Ban Your Way Out of Shadow AI
A dispatch from AGC Tech 2026, where every IT Director in the room had the same idea for stopping unapproved AI use, and the data says it's the wrong one.
Ready to take on more premium projects without adding to your overhead?
Lock in your flat-rate, unlimited-user fee with ServiceIQ today.
Start your free trial