Soundings
Our Perspectives and Insights
How to Measure AI Trust: Whose Name Is on the Keys?
Remember the first time you handed your teenager the car keys and watched them back out of the driveway alone? Or the first time you let a new hire run the client meeting while you sat in the back and kept quiet? Or the first time you gave someone the company card and stopped asking for the receipts?
You remember each of those days, because trust has a date on it: the day you stopped watching every move. That date is how you measure trust, in people and in AI.
And none of it happened all at once. The teenager got a learner’s permit first, then daylight driving, then a car with no friends in it. The new hire ran the internal meeting before the client one. The card came with a limit. You handed over authority one key at a time, and the size of each key matched how much you could undo if it went wrong.
Letting It Run Means Giving Up Oversight
AI gets to the same moment. When you let it run, you stop watching every action it takes, and that’s the point. An agent you supervise on every step is just a slower way of doing the work yourself. The oversight you were doing by hand doesn’t go away, though. It moves somewhere else: into a QA line that samples what the agent does, and into decisions made ahead of time about what it’s allowed to touch.
Most companies measure AI trust with a survey. Do your people trust the tools? How confident do they feel? That’s an opinion poll. It tells you how people feel about the agent, but nothing about what they’ve actually handed it.
Two Dates That Measure Trust
Noah Shinn, founder of the personal agent Instinct, measures it differently. On a recent episode of Invest Like the Best, he described tracking how long it takes a user to hand the agent a first password, and how long before they give it a credit card to buy things on their own. He calls those proxies for trust. Three weeks in, he said, roughly 40 percent of users have shared a card.
Controls are what let you hand over the keys sooner without betting the company.
Inside a company, those two dates look like this:
- Time to first production credential: the day an agent touches a live system, not a sandbox.
- Time to first unsupervised spend: the day an agent spends money with no human approving each purchase.
A lot of governance programs are built to push those dates out as far as they can. A good assurance framework should do the opposite and bring them forward, safely. Controls are what let you hand over the keys sooner without betting the company.
Size the Key to What You Can Undo
The restrictions on that learner’s permit weren’t random. Each one capped the blast radius: how much damage a mistake could do, and how much of it could be fixed afterward.
Jack McCullough, founder of the CFO Leadership Council, gave finance leaders a good question to ask before approving any AI deployment: “What is the maximum credible loss if this system fails” and “how quickly would we know?” I’d add a third part: once you know, can you undo it?
So stop asking whether you trust the AI. Ask which keys, how big, and what happens when one gets misused.
Reversibility is what separates a key you can hand over this week from one that should wait. A draft the agent writes for a person to send can be thrown away. An email it sends can’t be unsent. A refund can be clawed back. A wire transfer mostly can’t. A configuration change with a tested rollback is a bad afternoon. A deleted customer record is a bad quarter. It’s the same agent in every case, holding very different keys.
And most companies haven’t practiced the undo. Grant Thornton’s 2026 AI Impact Survey found that nearly three in four organizations are piloting, scaling or running autonomous AI, yet only one in five has tested a response plan for when it fails.
Long before AI, I remember the first time I ran a real disaster recovery test. Not a walkthrough of the plan on paper: we pulled the plug on the machine and cut the connection on purpose, to see if everything came back the way the plan said it would. Until that day, the plan was a document. After it, the plan was something we could trust. An AI rollback is no different. Until you’ve actually run it, you don’t really have one.
So stop asking whether you trust the AI. Ask which keys, how big, and what happens when one gets misused.
Four Questions Before You Hand Over a Key
- What exactly is the key? Name the authority: which system, read or write, what spending limit. “Access to the CRM” is not a key. “Can send a campaign to up to 500 contacts” is.
- How big is the blast radius? The worst credible loss, how fast you would know, and whether it can be undone. Anything you can’t reverse gets a smaller key or a person in the loop.
- Whose name is on it? One person who decided this key gets handed over, per decision, not a committee. If the box is empty, nobody granted the key. Somebody just forgot to say no.
- What takes it back? The trigger that narrows or revokes the key: a new model version, a climbing error rate, an incident. A teenager who gets a speeding ticket loses the car for a weekend. Your agent should too.
A diagram with no name in the box is a picture, not a control.
Then do what you did with the teenager. Widen the keys as the record earns it. Screen, onboard, set a probation period, and grant more authority one step at a time.
The Bottom Line
You remember the day you handed over the car keys because somebody made that call on purpose. Most AI keys today are handed over by default: a credential pasted into a config file, a card added just for the pilot, and nobody’s name next to either one.
Map what your agents can do, and put a name in every box. A diagram with no name in the box is a picture, not a control.
A Pilot & Rutter principle.