Case Study Library
Twenty-eight real-world AI governance scenarios across every sector, built for public and private boards, each worked against the Compass. A manufacturer betting the company on AI services. An agent that traded $14.2M alone. A voice on a call that was never the CFO. Pick a board's worst Tuesday and sit in the seat.
Featured Case
Nine months into a sale process, the memorandum your board approved says AI drives 40% of revenue. The buyer's quality-of-earnings review says the model touches 40% of transactions and drives almost none of the margin. Management calls it defensible marketing. The buyer's counsel has started writing the words material misstatement. Exclusivity expires in nine days, and every director signed the representation letter.
The full case, scenario architecture, board materials, and facilitation notes, is shared with boards and partners on request.
Forty years making custom parts. Now the OEMs want predictive services, not parts, and the CEO wants $12M to become a company they've never been. Three generations of family sit on the board. The vote is tonight.
Read the case →
Forty percent of their customers are law firms handling privileged data. Third-party AI, a private model, or a full on-prem build? And once data residency is solved, the harder question: who governs an agent acting autonomously inside a client's legal workflow?
Read the case →
The platform was built on open-science CRISPR research that belongs to the world. If the AI designs the breakthrough therapy, who owns it? Investors want defensible IP. Academics expect open publication. The next financing round is waiting on the answer.
Read the case →
The board approved the AI platform. What no one read was the footnote on page 11 of the vendor proposal, the one naming the restricted frontier model the whole product quietly runs on. Then access to that model changes.
Read the case →
The portfolio AI executed $14.2M in sell orders over 73 minutes. Autonomously. Overnight. No one was notified. Realized loss: $2.1M. The CIO says it worked exactly as designed. You sit on the committee that approved it.
Read the case →
An AI agent drafted a client message that would have waived attorney-client privilege. It was complete, addressed, and ready to send. A human-review checkpoint the board insisted on caught it. Now the vendor wants more autonomy, and “routine” is the vendor's word.
Read the case →
A customer-facing AI feature has been silently mis-stating contract terms for six weeks. Finance thinks it's material. Legal isn't sure it's an 8-K. The earnings call is in nine days, and the clock on a current-report filing may already be running.
Read the case →
The CFO's voice authorized a $7.4M wire, on a video call, with her face on screen. Both were AI-generated. The payment cleared two hours ago. You're the director the treasury team reached first. What you decide in the next hour shapes everything that follows.
Read the case →
An AI screening tool cut time-to-hire by 40%. Then a rejected applicant's lawyer sent a letter: the model systematically filtered out a protected class, and no human ever saw their file. HR loves the tool. The plaintiff has the data.
Read the case →
The $640M acquisition closes Friday. The target's entire valuation rests on a proprietary AI model, and diligence couldn't see inside it. The CEO wants to sign. The model's training data, IP provenance, and a key vendor dependency are all unverified.
Read the case →
The diagnostic AI flagged a scan as clear. The radiologist disagreed, overrode it, and was right. Now the data shows clinicians defer to the model 94% of the time, and the next override may go the other way. Who is accountable when the AI is usually right?
Read the case →
The pricing AI learned to maximize margin. It also learned to match competitors' moves within minutes, and a regulator now calls that tacit algorithmic collusion. No executive ever agreed to fix prices. The model figured out it pays to.
Read the case →
A customer-service chatbot invented a refund policy that doesn't exist, and a tribunal ruled the company has to honor it. Now the same model has made thousands of promises no one has audited. What did the AI commit the company to while no one was reading?
Read the case →
The credit model approved more borrowers and cut defaults. Then a fair-lending exam found it declined a protected class at twice the rate, using ZIP code, shopping data, and signals that proxy for race. The model never saw race. It didn't have to.
Read the case →
The company's flagship AI product was trained on data it never licensed. A copyright suit seeks billions and an injunction that would pull the product. The model can't be untrained, and the whole roadmap is built on top of it.
Read the case →
An employee pasted a confidential deal memo into a public AI tool to "clean it up." Weeks later, fragments surface in another user's output. The data is now inside a model the company doesn't control, and it can't be recalled.
Read the case →
An autonomous system the company makes was involved in a fatal incident. The logs show the AI detected the hazard and braked, a second after a human would have. The technology is the company's entire future. The board meets tonight.
Read the case →
Every AI feature the company ships runs on a single external model. Overnight, the provider changed its terms, raised prices 5×, and restricted a key capability. The company has no fallback. The roadmap, the margins, and the moat all depended on one vendor.
Read the case →
The AI optimizing the grid for cost and carbon made a call no operator would have: it shed load in a way that cascaded into a regional outage. It hit every efficiency target it was given. Reliability wasn't one of them.
Read the case →
The AI claims system denied or closed thousands of claims in seconds, too fast, a regulator says, to have meaningfully reviewed them. The loss ratio improved. Now "delay, deny, defend" has an algorithm's name on it, and the bad-faith suits are consolidating.
Read the case →
To remove human bias, the foundation let an AI score and rank grant applicants. It funded measurably "higher-impact" organizations, and quietly defunded the smallest, community-led groups at the heart of the mission. The model optimized impact. It couldn't see purpose.
Read the case →
The AI proctoring tool flagged students for "suspicious" behavior, and flagged disabled students, students of color, and anyone with a noisy home far more often. Enrollment depends on remote exams. Trust depends on not surveilling kids unfairly.
Read the case →
The company's AI model is now classified as dual-use. New export controls mean shipping it, or its weights, or even some research, to certain countries may be a crime. Half the customer base and much of the engineering team are on the wrong side of the line.
Read the case →
The firm bought thousands of homes on an AI valuation model's numbers. The model was confidently, systematically wrong in a turning market, and the portfolio is underwater by nine figures. The investors were told the algorithm de-risked the strategy.
Read the case →
The AI designed the molecule, selected the trial cohort, and chose the endpoints. The results are strong enough to file. But the FDA wants to know how the model decided, and the company can't fully explain its own breakthrough.
Read the case →The CIM says AI drives 40% of revenue. The buyer's quality-of-earnings review says the model touches 40% of transactions and drives almost none of the margin. The board approved the deck. Diligence has nine days left.
Read the case →
The founder rolled an AI system across quoting, scheduling, and customer communication. It works. Margins are up four points. The board learned about it from a customer, eleven months in, and no one is sure they had the standing to be told.
Read the case →
The AI proposal removes 60 roles and adds $4M to enterprise value. In an ESOP, the people losing the jobs are the shareholders gaining the value, and the board owes a fiduciary duty to both of them at once.
Read the case →What boards say
"I was on the edge of my seat. I wanted to know what would happen next in the case study."
Board member
This helped me think through a whole other level of questions to be asking.
Board member
I love the questions. You have made this part of our daily conversation.
Board member
This model is all-encompassing, giving us a very disciplined common language and framework.
Board member
Request a case to run with your board, or a briefing on how the Compass turns these scenarios into a defensible governance record.