Making AI Decisions You Can't Predict the Outcome Of

There is a specific discomfort in making AI decisions right now, and it is worth naming precisely, because most executives are experiencing it as a personal failing rather than as the structural condition it actually is. You are being asked to make commitments, budget, headcount, platform, strategy, about a technology whose capabilities, costs, and competitive implications are changing every few months. The tools you would normally use to make such a decision, forecasting, ROI projection, a five-year plan, all quietly assume a stability that does not exist here. So you either force a confident-looking plan onto a situation that does not support one, or you freeze, and both feel wrong because both are wrong.

The instinct in this position is to demand better prediction: more research, a clearer roadmap, a firmer sense of where this is going before committing. That instinct is the trap. The honest truth is that nobody knows where this is going, including the people building it, and waiting for the fog to clear is a decision to let others move while you stand still. The useful move is not to predict better. It is to decide differently, using a set of tools built specifically for the situation where prediction is not available, which turns out to be a well-developed discipline rather than something you have to invent.

Prediction is the wrong tool for this problem

There is a formal name for the condition you are in, and recognizing it changes how you approach it. Decision scientists call it deep uncertainty, distinguished from ordinary risk. Under risk, you do not know the outcome but you can assign sensible probabilities, which is the world that forecasting and expected-value math are built for. Under deep uncertainty, the probability distributions themselves are not confidently knowable, you may not even be able to identify all the relevant variables, and applying probability-based tools produces confident numbers that are worse than useless because they disguise the uncertainty rather than accounting for it.

AI strategy sits squarely in deep uncertainty. Nobody can put a reliable probability on how capable models will be in two years, what they will cost, which vendors will lead, or how competitors will deploy them. Building a five-year ROI model on those unknowables is not rigor. It is a forecast dressed as a plan, and it will be wrong in ways you cannot anticipate.

The field built to handle this, coordinated in large part by RAND, makes one reframe that resolves most of the discomfort: under deep uncertainty, you are not seeking confidence in a forecast, you are seeking confidence in a decision. Those are different goals. You cannot be confident about what AI will look like in 2029. You can be confident that a particular decision is a sound one to make today given that you cannot know. Aiming for the second, rather than despairing at the impossibility of the first, is the whole shift.

Three kinds of move, sorted by regret

The most useful framework for this comes from a 1997 Harvard Business Review article, "Strategy Under Uncertainty," by Hugh Courtney, Jane Kirkland, and Patrick Viguerie, and it has held up for nearly thirty years because it maps decisions by how they perform across futures rather than trying to pick the right future. It sorts strategic moves into three types, and the sorting is the tool.

No-regret moves pay off no matter how the future turns out. In AI, a large share of the genuinely valuable work is here, and it is the least glamorous. Getting your data into a state where it could be used is valuable whatever AI becomes, because every plausible future rewards clean, connected, accessible data. Building your organization's literacy so people can evaluate AI claims is a no-regret move, as is the decoding discipline for vendor claims. So is running small, cheap experiments that build understanding. None of these require you to predict anything. They are correct across the entire range of outcomes, which is exactly why they should come first and why it is a mistake that they usually come last, behind the big visible bet everyone wants to debate.

Options cost a little now to preserve the ability to act later, buying a significant upside in some futures for a small cost in the others. A limited pilot, a contract with favorable exit terms, a reversible architectural choice, a modest investment that could scale if a capability matures and be abandoned cheaply if it does not. Options are how you participate in a future you cannot predict without betting the company on any single version of it. In deep uncertainty, most of your AI posture should be a portfolio of these rather than one committed direction.

Big bets are large commitments that pay off handsomely in some futures and hurt in others. They are sometimes correct, but they are correct rarely, and the discipline is to make very few of them, only where the potential payoff genuinely justifies the exposure, and never by accident. The failure mode is not making big bets. It is making them without realizing that is what you are doing, sliding into an irreversible, expensive, single-future commitment because it was presented as the obvious plan rather than as the gamble it is.

The practical power of the framework is diagnostic. Take any AI decision in front of you and ask which of the three it is. Most of what matters turns out to be no-regret moves that got deprioritized because they were boring, and options that got skipped in favor of a big bet nobody labeled as one. Simply sorting your choices this way tends to redirect attention to the moves that are sound regardless of what happens.

Reversibility is the resource to protect

There is a second principle that runs alongside the three-move framework, and it is the one to build into how every AI decision gets made: preserve your ability to change course, and treat that ability as something with real value rather than as indecision.

Because you cannot predict which direction is right, the decisions that keep you able to move when you learn more are worth more than they appear on a spreadsheet, and the decisions that lock you in are more dangerous than they appear. This connects to a property covered at length in the systems decision you can't easily reverse: switching costs are real and they compound. In a fast-moving technology, that argument gains force. A choice that commits you deeply to one vendor's approach, one architecture, one bet about where the technology goes, is expensive precisely because the thing you committed to may look wrong in eighteen months, and you will be unable to move.

So the operative question for any significant AI decision is not only "is this a good bet," but "if this turns out wrong, how easily can we change course, and what did preserving that ability cost us?" Often, paying a modest premium to stay flexible, a shorter contract, a more portable architecture, a smaller initial commitment, is the highest-value thing you can do, and it is systematically undervalued because flexibility does not show up as a line item while the premium you paid for it does.

The honest part

Several qualifications keep this from becoming a rationalization for never committing to anything, which is its own failure and arguably the more common one among cautious leaders.

Optionality is not free, and a portfolio of options with no big bets in it will lose to a competitor who correctly identified the moment to commit and did. There are times when the uncertainty has resolved enough, or the strategic prize is large enough, that a real bet is the right move, and hedging everything forever is how you end up having participated in a transformation without ever having benefited from it. The framework is not "always hedge." It is "know which kind of move you are making, make mostly the robust ones, and reserve real bets for where they are justified."

Deep uncertainty also does not mean total ignorance, and treating it as license to learn nothing is a mistake. Some things about AI are becoming clear, and a leader should update as they do. The point is calibration: be confident where confidence is warranted and honest about where it is not, rather than either forcing false precision or throwing up your hands at all of it.

And there is such a thing as too much deliberation. The frameworks here can become an excuse to analyze indefinitely, and at some point you have to act on the sorting you have done. The goal of thinking clearly about uncertainty is better decisions made at a reasonable pace, not a more sophisticated way to avoid deciding.

Decide for the world you're actually in

The reframe worth carrying is that the discomfort you feel making AI decisions is not a sign you are doing it wrong. It is the correct response to genuine deep uncertainty, and the mistake is trying to make it go away by forcing a confident forecast the situation cannot support. The leaders who navigate this well are not the ones with the best predictions, because nobody has good predictions here. They are the ones who stopped trying to predict and started deciding well without prediction: loading up on the moves that pay off regardless, holding a portfolio of cheap options into the futures they cannot rule out, making big bets rarely and deliberately, and protecting their ability to change course as the valuable resource it is.

There is one question that operationalizes all of it. For any AI decision on your desk, ask: does this pay off across every plausible future, does it preserve my ability to act later, or is it a bet on one specific future? If it is the third, the follow-up is whether you meant to make a bet, and whether this is one of the few worth making. Most of the AI decisions that go badly are big bets that nobody recognized as bets. Simply seeing them for what they are is most of the protection you need.

FAQs

Q1. Why do normal planning tools fail for AI decisions?
Because they assume risk, where outcomes are unknown but probabilities are estimable, while AI strategy involves deep uncertainty, where the probabilities themselves are not knowable and you may not even know all the relevant variables. Forecasting and ROI models applied to deep uncertainty produce confident-looking numbers that disguise the uncertainty rather than accounting for it, which is worse than acknowledging you cannot predict.

Q2. What is the difference between risk and deep uncertainty?
Under risk you can assign sensible probabilities to outcomes, which is what expected-value math is built for. Under deep uncertainty the probability distributions are not confidently knowable, and using assumed distributions leads to failure. AI strategy is deep uncertainty: no one can reliably estimate model capability, cost, or competitive dynamics a few years out, so tools requiring probabilities do not apply.

Q3. What are no-regret moves in AI?
Actions that pay off no matter how the future unfolds. Getting your data clean and connected, building organizational literacy to evaluate AI claims, and running small cheap experiments all qualify, because every plausible future rewards them. They tend to be undervalued because they are unglamorous, but they should come first precisely because they require no prediction to justify.

Q4. What does it mean to treat AI investments as options?
An option costs a little now to preserve the ability to act later, delivering large upside in some futures and small cost in others. Limited pilots, contracts with favorable exit terms, reversible architecture choices, and modest investments that could scale or be abandoned cheaply are all options. In deep uncertainty, most of your AI posture should be a portfolio of these rather than a single committed direction.

Q5. When is a big bet justified?
Rarely, and only deliberately. A big bet is a large commitment that pays off in some futures and hurts in others, and it is appropriate when the potential payoff genuinely justifies the exposure and the uncertainty has resolved enough to commit. The common failure is not making big bets but making them accidentally, sliding into an irreversible commitment because it was framed as the obvious plan rather than the gamble it is.

Q6. Why is reversibility so important here?
Because you cannot predict which direction is right, so the ability to change course when you learn more has real value, and decisions that lock you in are more dangerous than they look. Switching costs compound, and in a fast-moving technology a deep commitment to one vendor or architecture may look wrong within months. Paying a modest premium to stay flexible is often the highest-value move available.

Q7. Isn't this just an argument for never committing?
No, and that is the opposite failure. Optionality is not free, and a portfolio of hedges with no bets will lose to a competitor who committed at the right moment. The framework is to make mostly robust moves, hold options into uncertain futures, and reserve real bets for where they are justified, not to avoid commitment indefinitely, which is how you participate in a transformation without ever benefiting from it.

Q8. What single question should I apply to an AI decision?
Ask whether it pays off across every plausible future, preserves your ability to act later, or bets on one specific future. If it is the third, ask whether you meant to make a bet and whether it is one of the few worth making. Most AI decisions that fail are unrecognized big bets, so identifying which type you are facing is most of the protection.