Today I want to tell you a story about a newsroom, a contract, and two AI tools that didn't survive the year.
Politico's newsroom had been running two of them. Report-Builder, for its branded reports. Live Summaries, for live coverage.
The union contract already had something to say about that. 60 days' notice before deployment. Good-faith bargaining. Human oversight on anything touching a job.
Management skipped all three.
Here's the bit that should stop you scrolling.
The tools got it wrong. Report-Builder's reports had what the arbitrator called "glaring factual errors." Live Summaries put out what the ruling called "error-riddled unedited coverage."
Not a plot twist. Read the contract again. 👆
The oversight that would have caught this was already written down. Management just never applied it.
The arbitrator didn't hedge:
"If accuracy and accountability [are] the baseline, then AI, as used in these instances, cannot yet rival the hallmarks of human output."
And here's the part that got me. 🤯
Both tools are gone now. Permanently. The union's general counsel, Amos Laor, put it plainly: "Management must provide notice, bargain with the union, and ensure that innovation does not come at the expense of workers' rights."
That's a union talking. A whole legal apparatus built to catch exactly this.
You don't have one of those to make someone answer for it.
That's not a relief.
That's the whole point of this edition.
The test almost nobody passes
Quick one. If I sat you down right now and asked what you personally expected the AI tool your team is using to do, not the vendor's pitch, could you answer me in one sentence? 🤔
Most owners can't.
The yes happened on a Thursday. Three other things on that meeting's agenda. What "working" was supposed to look like never made it onto paper.
It's been months. Nobody's thought about it since.
Naming the debt
This is where urgency debt hands off to the next one.
Urgency debt is the pressure that says there's time to write the rules later.
Accountability debt is what you owe when later never comes. 💸
Same pressure. Different bill.
Here's what that costs you. Sit with this part.
The tool makes a mistake. They all do, same as humans, but left unchecked, the mistake grows.
A customer gets a wrong answer.
An employee gets flagged for a performance issue nobody can explain, because the model that flagged them was never told what "good performance" meant.
A decision goes out that wouldn't survive a human looking at it.
- No record of what you expected.
- No requirements anyone can point to.
- Nothing showing the reasoning behind the yes.
So the mistake becomes yours. To explain. Personally.
Your name is the one attached to the decision. That's reputation, and it doesn't care whether you're an LLC, a C-Corp, or an Inc.
The legal and financial piece does care. How your business is structured decides who gets sued and who pays. Illinois already has a law requiring notice before AI touches an employment decision, and it can reach the decision-maker directly, entity or not.
Worth checking with your own counsel exactly where you stand.
Where the real fitness test happens
I've watched plenty of executives say yes to a tool without ever finding out what the end users needed.
Finding out takes a real fit-for-purpose, fit-for-use test. Does it do the job? Does it work for the people stuck using it? (Same test from the adoption debt edition. Still the same two questions. Still gets skipped for the same reason.)
Skip it, and here's what fills the gap: a demo nobody's proven solved anything, a nod instead of a real test, a vendor telling you it's fine. That becomes the process. The one you didn't mean to build.
You're not the only one carrying this
IBM asked 2,000 CIOs and CTOs across 33 countries this year.
- 2 out of 3 said they're held accountable for AI systems they don't fully control.
- 77% said adoption is outrunning their ability to govern it. 📊
One of them is standing in front of leadership this week, explaining a decision their AI system made without them.
BCG asked 351 CEOs and 274 board members at $100M+ companies who actually owns AI outcomes.
- CEOs said more than a third of their own review rides on it.
- Boards said 27%.
Two groups. Two different answers. Same question.
A bigger checkbook doesn't fill in the missing entry.
A hire gets more paperwork than your AI does
Think about the last person you hired. Job description, interview notes, a reason you picked them over the other finalist. All on the record.
Think about the last loan your business took out. Application, financials, a use-of-funds statement, someone underwriting the whole thing before you spent a dollar. (Somewhere, a loan officer is reading this nodding so hard they need a neck brace. 😅)
Now think about the last AI tool you said yes to.
Politico's contract already had the answer
60 days' notice. Bargaining. Human oversight. Before a tool got anywhere near someone's job. Written down, in a contract. That's the part worth stealing. Not the lawsuit. The paperwork. You can build that same discipline without a union making you.
Daniel Kahneman said something close, in a 2013 Q&A: "I have given that advice, to keep score and when you make a decision, document the options that you considered but didn't choose."
Write down what you expected, and why, before you find out if it worked. Once you know the outcome, hindsight tricks you into thinking you always knew.
OK, here's the dividend. Not a hedge. 🎉
Do this, and you catch the bad tool before it costs you the budget or the trust. That's the actual payoff. Fewer expensive yeses.
The 5-Question Sign-Off
- 1️⃣ Who decided, and who was in the room?: Were the actual users in that room? If not, stop. Go find them first.
- 2️⃣ What's the actual problem this solves?: Not "AI." The real thing underneath it.
- 3️⃣ What do you expect it to do?: One sentence. Plain enough that an employee or a customer could read it and agree.
- 4️⃣ What would tell you it's wrong?: And who's watching for that.
- 5️⃣ How does your existing governance apply here?: 5 questions. Your team, in writing, signed and dated, before AI enters the building.
The record's still there if you ever need it, what you expected, who decided, what would've told you it was wrong. But that's the backup. Not the point.
McKinsey found the closest thing to proof this works. Businesses that named an owner: 2.6 on a 0-to-4 trust-maturity scale. No named owner: 1.8. Naming someone moves the number almost a full point. Not a rounding error.
Doing This Once Is Not Enough
Here's the part that doesn't resolve, though.
Writing the 5 questions once doesn't fix anything. The habit only holds if your team keeps answering them, yes after yes, long after it stops feeling new.
That part's on you now.
The next yes is already sitting on your desk
A renewal. A vendor demo. A subscription someone already started without asking.
So here's my question for you. 🤔
Whichever one it is, could you write those 5 answers right now? From memory?
Politico's people had a union to make someone answer for it.
You don't.
The paperwork's still yours to write.
If you want to know where accountability debt sits in your own org, there's a free 5-minute diagnostic, link below.
Where Does Accountability 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.