I want to tell you about a tool that joined a meeting, said nothing, and wrote down everything everyone said.
It was doing its job, and what it did was covered in terms that were accepted without being read. It got connected months ago, between two other things. After that, it just showed up. Sales calls. A performance conversation. The call where a customer spent 40 minutes explaining exactly why they were unhappy and exactly what it would take to keep them.
Every one of those people was recorded. But were they asked?
That describes the allegations in a case a federal judge let go forward last month. 👀
It is also how AI trust debt starts.
A court just drew a line between your tool and you
Otter.ai makes an AI note-taker: it joins your video calls, records them, and writes up what was said.
On August 13, 2026, Judge Eumi K. Lee of the Northern District of California refused to throw out the core claims in In re Otter.AI Privacy Litigation.
Otter has not been found liable for anything, but the case gets to continue. What the court said while letting it continue is the part worth your time.
Otter argued it was an invited participant. A person on the call had put it there. The court was unconvinced, and the reason was specific: Otter “does not just participate in meetings, it retains the recordings of the conversations and uses them to improve its own machine-learning models and services.” The court's words: “an independent party who uses the recordings to train its models for its own benefit.”
The part you left blank
Trust debt is what you owe when the people your AI touches have every reason to doubt you and nothing to weigh against it.
Accountability debt was the missing record of your decision. Trust debt is the missing statement of your values. It builds every time a tool decides how it treats your customers and employees. You could have made that choice, but you did not.
The vendor wrote your values into a default setting
You inherited the house rules, at the speed of a checkout page. 💸
A second suit, Chamberlain v. Granola, filed in the same court, describes a note-taker built to be invisible. It shows up as nothing at all, and the meeting looks exactly like a meeting. The transparency features are documented, and they ship turned off. The complaint quotes the product's own marketing on the point: “Other people in the room won't know it's there.”
A company wrote a policy about whether customers get told they are being recorded, decided the answer was no, set it as the default, and sold it as a feature.
That policy is yours now, and you will be held to it.
Brian McGinnis, who co-chairs the data security and privacy practice at Barnes & Thornburg, put the exposure plainly this month: “I can't tell clients definitively that they're legal as they are.” His standard is consent from everyone on the call, and in California and a handful of other states, every individual has to give it, not just the person who turned the tool on.
Otter reports 35 million users. Fireflies, a competitor that does the same job, reports 20 million individuals across a million organizations. 📊
One of those organizations is going to find out what it stands for by reading a complaint.
You need an AI Inventory
There should be a list of every AI tool touching a customer or an employee, what it does, and who approved it.
Does your business have one? The tools did not arrive through one door. They came through free trials, through features bundled inside software you were already paying for, and through at least one integration a contractor built, documented, and handed to whoever was in the room that day.
Until that list exists, every question in this article has the same answer: you don't know.
You bought it, and the contract says so
Here is what makes this different from every other AI decision you will make.
A person who clicks agree is binding your business to terms they may not have read. Not reading first isn't defensible in any contract you have signed.
The decision exists, and one person made it for everyone. Did they have the authority to do so? That is decision debt, and it gets its own edition.
The Census Bureau’s Business Trends and Outlook Survey asked firms whether they used AI to produce goods or services between December 2025 and May 2026. These numbers describe what companies sat down and decided to adopt.
- 32% of companies with 100 to 249 employees said yes.
- Under 20% of the smallest firms did.
Often, the note-taker is missing from that count. Fireflies reports 20 people for every organization it serves, and that ratio is the story: these tools arrive one signature at a time, through a free trial an employee started without approval.
Your paid expense report already approved it
When the trial ended, the charge began. That charge cleared. A subscription that size does not stop an expense report, because an expense report is built to check an amount against a budget line, and the amount was fine. The control worked. But it was only ever looking at one number.
Does your accounting office ask what the money bought? Who does it touch? What is it allowed to do to them?
It should. That it doesn't is a gap in your controls, and the gap belongs to you.
Now you have a signed agreement, a monthly charge, and a record that went unreviewed and unapproved. One month, a charge like that escapes scrutiny. 12 months is a choice. 📥
Which means the question of whose values this tool runs on was answered in writing, months ago, by someone who wasn't you.
Disclosure costs you trust. Getting caught costs more.
The assumption runs like this: Transparency fixes it. Tell people, disclose the AI, and the trust problem goes away.
That assumption is false, and it has been tested thoroughly.
Oliver Schilke and Martin Reimann ran 13 experiments with 4,093 people and published the results in Organizational Behavior and Human Decision Processes. Their finding, stated in their own abstract: “actors who disclose their AI usage are trusted less than those who do not.”
Across supervisors, employees, analysts, creatives, and whole investment funds. 13 times.
They also tested what happens when the AI use gets discovered instead of disclosed.
The mechanism they identified is legitimacy. People punish the sense that something was done to them by a process they had no say in.
Which puts transparency in a different category. A disclosure is a receipt, and by its own mechanism, what it is worth depends on whether the people receiving it had any say in the decision behind it. If the decision was a checkbox, the disclosure carries that.
Your customers will ask before a court does
The statutes reaching into that Otter case were written for a different world. California's wiretapping law dates to 1967. The federal one dates to 1986. Illinois passed its biometric privacy act in 2008, years before anyone built a product that turns a sales call into a voiceprint.
Every one of them predates the thing you are asking it to govern. Courts are deciding what they mean one ruling at a time, and that will outlast your next contract and the one after it.
You could wait for the ruling that covers you. Your customers want an answer in your own words, not a citation.
Which makes this a values problem rather than a compliance one.
The Markkula Center for Applied Ethics at Santa Clara University publishes a decision framework, last revised in 2021, and the useful part for a business owner is that it refuses to let one lens run the whole decision. A choice must survive all 6 of them.
They were written for hard ethical calls. This is one.
The dividend: 6 questions to ask before you buy
Take these 6 lenses and point them at the tool in front of you.
- Who has a right to know this is running, and have they been told?
- Does it treat everyone it touches the same way, or does it work better for some of them than others?
- Who benefits, and who pays the cost when it gets something wrong?
- If every business in your line ran it exactly this way, would customers/employees be better off?
- Would you describe how you use it to a customer, out loud, in your own words?
- Who is depending on you to get this right without ever getting a say in it?
6 questions, answered by your leadership team together, on the record, before anyone books a demo.
Then ask the people who will use the tool, and the people it will be used on. They will live with the answer longer than anyone in that leadership meeting will.
Ask these before the demo, and they are a standard. Ask them after the complaint, and they are testimony. 📋
Your answers become your standard
The vendor has to show you how its tool meets all 3 of these:
- Who this tool touches. Everyone on the other end of it: customers on recorded calls, employees in performance conversations, applicants, patients, students. The list runs longer than most owners expect, and the ones at the bottom usually have the least say.
- What it is not allowed to do to them. In your words, before you see a feature list. It records a person only after they have been told. What it hears is deleted when we stop paying, or when the problem it was bought to solve is solved. Our customers' conversations do not train anyone's model.
- What happens to what it hears. Where it goes, how long it stays, who else gets to see it, and how a person asks for it back.
The tool meets your standard, or it does not. 📄
That is the dividend. A customer who hears it from you first renews. One who hears it from a filing asks for their money back and tells the next person who asks them for a referral.
The backward pass you have not performed
Writing this once fixes the next purchase. It does nothing about the ones already running.
So the standard is half the work. The other half is going through every tool already in your business and putting it through the same 6 questions. That half is slower; it has no launch date attached, and most owners will want to skip it. But it is the half the debt is sitting in.
Skipping it is how the debt accrues. A tool you have not examined is still your decision, whether you made it last week or last year. And on the day it surfaces, the purchase date will not be the question anyone asks you.
Both halves, or the ledger, keep growing while you read this.
Did the recording make a difference?
Go back to that customer. The one who spent 40 minutes explaining why they were unhappy and exactly what it would take to keep them.
That call was recorded. The transcript exists. Which raises a question worth more than any of the legal ones in this piece:
Did anything change because of it?
Or did the transcript get filed where nobody will open it again? You paid for that recording either way, but only one of those is worth the money.
A tool used properly is a chance to get better. But used as CYA, it is worse than not having it, because it costs money every month and it teaches your people that listening is something you log rather than something you do.
If that is what yours has become, cancel it this week.
And if it is worth keeping, the 6 questions are where you start.
Set your values before AI sets them for you.
Where does AI trust debt sit in your organization?
Start with the AI Debt Diagnostic. 9 questions, about 5 minutes, free. It scores your organization across the 9 debts and shows where to start.
Every edition of this series is at digitalbard.co/perspectives.