A board director carries a duty most people never carry at work. When they vote to approve a special levy, hire a contractor, or settle a dispute, they are not just making a decision. They are making it on behalf of other people’s money, and the law holds them to a standard for how they get there.
In British Columbia the standard is written down. The Strata Property Act sets a council member’s standard of care: act honestly and in good faith with a view to the strata corporation’s best interests, and exercise the care, diligence and skill of a reasonably prudent person in comparable circumstances. It is a statutory, fiduciary-like duty rather than the full fiduciary duty a corporate director owes, and the specifics vary by jurisdiction. A condo board in Ontario or an HOA in California sits under a different regime. But the spine is close to universal: those who govern owe a duty of care, and the standard they are judged against is not whether they turned out to be right. It is whether they were careful.
That distinction is the part of governance almost no one outside it thinks about. You can make a wrong call, in good faith, on good information, and be fine. What exposes a board is not being wrong. It is acting carelessly, on something it cannot later stand behind.
Hold that in mind, because it is the whole reason AI is harder to bring into a boardroom than into almost anywhere else.
The problem everyone is solving for the wrong setting
The entire conversation about AI getting things wrong has been framed as an accuracy problem. The model hallucinates. It states something false with total confidence. The industry treats this as a bug to be engineered down over time, a number to push toward zero with better models and more data.
In most settings that framing is fine. If you ask a chatbot for a recipe and it gives you the wrong oven temperature, you lose a dinner. If it summarizes an article and gets a detail wrong, you are mildly misinformed. The cost of a confident wrong answer is annoyance, and the fix is to be a bit more careful next time.
A board does not live in that setting.
When a board acts on a confident wrong answer, the cost is not annoyance. A treasurer who tells the board “we approved the reserve contribution increase last year” when no such vote happened has not made a small error. They have created a record the board will rely on, spend against, and have to defend later. When an owner challenges it, and in a building of any size someone eventually does, the question is no longer “what did the AI get wrong.” It is “who relied on it, what did the board do because of it, and can the board show it acted carefully.”
A decision built on a fabrication does not just risk being wrong. It risks being unwound. The levy gets challenged. The vote gets reversed at a tribunal. The owners lose confidence, the manager wears the blame, and the corporation absorbs the cost. None of that requires a courtroom or a director’s personal ruin to hurt. It just requires a board that cannot reconstruct, when asked, why it did what it did.
That is not an accuracy problem. It is a defensibility problem. And a defensibility problem does not get solved by a model that is usually right.
”Usually right” is the trap
Here is the uncomfortable thing about a system that is usually right. It is far more dangerous than one that is often wrong.
A tool that is wrong half the time trains you to check everything. You never trust it, so it never hurts you. A tool that is right ninety-five percent of the time trains you to stop checking, and then quietly hurts you on the five percent, at the exact moment you have most let your guard down. The better it gets, the more confidently you lean on it, the worse the failure when it comes.
For a board this is the worst possible shape for a tool to have. The standard you are judged against is care. A tool that erodes your habit of checking erodes the one thing you would need to show if anyone ever asked how the decision was made. The smarter it seems, the more it invites a board to stop doing the very thing that makes its decisions defensible, without anyone noticing the habit slip.
So the goal cannot be a smarter answer. It has to be a different kind of answer.
Show your work
There is a reason “show your work” survived from grade school into every serious profession. An auditor does not just tell you the number. They tell you the number and hand you the trail that produced it. A good lawyer does not just give you the conclusion. They cite the clause. The conclusion alone is worthless in any setting where someone might later ask how you got there.
This is the bar AI has to clear before a board can responsibly lean on it. Many tools show sources. Far fewer make each material claim inspectable enough that a board can rely on it as part of a decision it could later defend. Most give you an answer: a confident paragraph, fluent and clean, with nothing underneath it you can check. You cannot tell what it read, what it ignored, or whether the thing it told you actually happened. You are asked to trust the fluency. For a board, trusting fluency is the definition of acting carelessly.
The answer a board can actually use looks different. It does not say “the board approved the levy in March.” It says that, then shows the meeting minutes, the resolution, the vote result, and the supporting correspondence, so the board can inspect the record before relying on the conclusion. The AI is not asking you to trust it. It is doing the legwork of assembling the evidence and then getting out of the way so the board can exercise the judgment that is its own to exercise.
That distinction sounds small. It is the entire difference between a tool a board can defend using and one it cannot.
The question that actually matters
So when someone tries to sell a board or a management firm an AI tool, the question that gets asked is almost always the wrong one. People ask “is it smart.” They ask how good the model is, how it compares, how often it is right.
The question a board should ask is narrower and far more revealing.
Can I defend acting on this?
Not is it impressive. Not is it usually right. Can the board stand in front of the owners, an auditor, or a tribunal and show that relying on this was a careful thing to do. If the tool cannot show its work, the honest answer is no, no matter how good the answers sound. And if it can, if every claim comes with the source attached and the trail is right there to inspect, then the AI has stopped being a thing you trust and become a thing you can verify. That is the only form in which it belongs anywhere near a decision that carries a duty.
Most of the industry is racing to make the answers smarter. The harder and more useful problem is making them defensible. Those are not the same race, and for anyone who governs under a duty of care, only one of them matters.