Every AI decision you make is quietly turning into one of two things. A debt you pay back with interest, or a dividend that keeps paying you back. Most leaders never find out which until it is too late to choose. Here is how you choose on purpose.
Last quarter, somewhere in your company, a smart person made a fast decision.
They found a tool. It worked. It cut a two-hour task down to ten minutes, and it did it on a Tuesday afternoon while everyone else was in meetings. No form. No approval. No conversation. Maybe a free trial, maybe a company card, and what feels like a small, quiet win.
That person moved on to the next thing. That is the part that matters. Not the tool. The moving on.
Because that decision did not end when the task got faster. It is still running. Every AI choice your company makes is a loan taken out against a version of the company that does not exist yet. And a loan matures. It becomes one of two things.
A debt. Or a dividend.
That is the whole question this series runs on, and it is the only one I want you carrying around after you look up from this. Debt, or dividend?
Why “debt” and not “mistake”
I chose the word “debt” on purpose, not because it sounds serious.
A mistake is a single event. You make it, you notice it, you fix it. Debt is different. Debt is quiet. It sits there looking like nothing while the interest compounds in a room no one is watching.
And here is the part that trips people up: you do not take on AI debt the day the tool fails. You take it on the moment someone buys or starts using a tool with no one asking what it touches, who it decides for, or what happens the day it is wrong. The bill just shows up later, with a number you do not recognize, long after the person who chose it moved on. The tool working well does not mean you dodged the debt. Working well is what hides that you took on debt at all.
The tool is not the debt. The skipped decision is the debt.
What the debt looks like
It shows up in two places, and the second one is rarely talked about, but should be.
The first is your organization, and it usually does not look like failure. A vendor nobody fully vetted, because the demo was so good, now holding your customer data. A tool bought on that same excitement that six months later almost nobody uses, because no one checked whether it solved a problem your people truly have, and no one was trained to fold it into the work. A workflow quietly making calls about real customers that was not tested against what those customers needed. None of this looks like a problem on the day it ships. It looks like progress. You pay for it later, when a customer asks a question your system answers wrong, when two teams hire for the same role because the tool told each of them something different, or when you realize you paid for a year of software your team quietly chose not to use, because it did not solve their problem.
The second place is harder to see, because it is inside the people doing the work.
When a person hands their judgment to a system and stops checking, a small muscle gets weaker. The analyst who used to catch the odd number now approves the output. The manager who used to write the hard email now edits the draft the machine wrote and calls it done. Nobody notices the erosion, because nothing breaks. The work still ships. It just ships with less thinking in it every quarter, and thinking is the one thing you were paying those people for.
That is the debt you cannot refinance. Replacing a bad vendor is painful and expensive, but you can do it. You cannot easily rebuild an instinct a whole team stopped using.
The same decision, made as a dividend
Here is the part I care about more, because I am not here to scare you off AI-enabled tools. I use them every day.
Take that same fast decision. Same person, same tool, same quick win. Change one thing. Before it spreads, it passes through governance, and I am not going to pretend that lives anywhere but with your leadership. The calls it makes, what is negotiable and what is not, what legal or regulation will not allow, what the company can stand behind, can only be made by the people who see across the whole business. In a large organization, that is a cross-functional group. In a smaller one, it is the C-suite, or the layer just below. Right-sized does not mean lightweight. It means a few deliberate calls, made on purpose, instead of a standing bureaucracy or a tool left to decide for itself. I want to be careful here, because “just add governance” is the sort of advice that sounds free but isn’t.
So, let’s look at what it costs to skip it. Air Canada put a chatbot on its website. When the bot told a grieving passenger he could claim a bereavement discount after the fact, which was not the airline’s policy, he booked on that promise. Air Canada refused to honor it, then argued before the court that the chatbot was responsible for its own words. The court disagreed and made them pay. Nobody had owned that bot. It worked, right up until the afternoon it cost them, in the ruling and in every headline that ran after.
Here is what right-sized governance would have caught, and I want to be clear I am not improvising this. There is an established pattern for governing AI without slowing it down. You anchor in an established governance standard, such as the NIST AI Risk Management Framework or the ISO 42001 management standard, and others like them. These are not IT checklists. They are enterprise and board-level standards, the same species as the financial controls your leadership already expects. Then you treat governance itself like a product: set a few firm guardrails, run light checks within the normal delivery rhythm rather than stage-gate approvals, and adjust the rules as you learn. The frameworks are thorough. The place you start is not. At its smallest, it is three plain questions:
- What AI are we using?
- What is the worst thing this one could get wrong?
- Who owns it, by name, with the authority to pause it?
That is not a six-month program. It is also not twenty minutes and a shrug. It is real work, sized to the risk, and small enough to happen at all. And it is the whole difference between the two kinds of loan.
Run the same tool through those questions first, and the shape of the risk changes. The named owner and their team pick up the real work: the checks, the monitoring, the occasional hard no. But the tool still handles the volume; the judgment stays with people, and the thing compounds in your favor rather than against you. Same software. Different questions, asked at the right moment. Debt becomes dividend on the strength of the governance most companies treat as burdensome.
The gap between the two outcomes is never the technology. It is whether humans decided on purpose, or the tool decided by default.
The tell
So how do you know which one you are making, in real time, before the bill or the dividend arrives?
You ask who decided.
If you can name the person who chose to use an AI tool and say in one sentence what they were and were not allowed to do with it, you are building a dividend. If the real answer is “it just sort of started happening,” you are carrying debt and paying interest right now. You just have not seen the statement yet.
Try it on your own company before you look up from this. Pick one place AI is in use. Ask: who said yes to this, and on what basis? If you can answer in a sentence, you are ahead of most leaders I talk to. If you cannot, you did not find a gap in your reporting. You found the debt.
There’s more than one kind
AI debt is not one thing. In the book I am writing, it comes in nine kinds, and they stack as follows:
- Urgency
- Governance
- Assumption
- Adoption
- Accountability
- Trust
- Measurement
- Decision
- Drift
Each one is a different way a fast yes turns into an expensive one.
You do not have to hold all nine in your head. You do need to know which one is charging you the most right now, and for most companies it starts where the list starts. Urgency: adopting AI before anyone agreed how to say yes to it. It is the first debt because it is the one that lets all the others in. Skip the decision about how you decide, and the other eight compound on top of that gap. That is where this series goes next, and it is the place a leader can actually start.
Why I’m doing this in public, every week
Two reasons, and I will be straight about both.
The business reason: eventually some of you will want help paying this debt down, and I would rather you already know how I think before that conversation.
The bigger reason: I wanted one subject I would still stand behind in three years, and this is it. So, every week I take one AI decision, run it through one question, and tell you which side it lands on and what to do about it. No manufactured panic. No “you are behind.” Just one decision, read early, while you can still choose.
Set your values before AI sets them for you. That is the entire job, and it is more human than technical, which is exactly why it keeps getting skipped.
The bill comes later. But so does the payoff. The only thing that decides which one shows up is a question you ask now, on purpose, before the tool answers it for you.
Every AI decision turns into one or the other. Which is this one?