Preparing for AI Uncertainty and Strategic Change

In the autumn of 2023, researchers at AI Impacts asked 2,778 people who publish at the leading AI conferences a simple question: by when is there an even chance that unaided machines will do every task better and more cheaply than people? The aggregate answer was 2047. A year earlier, the same question had produced 2060. The people closest to the technology had moved their own forecast by 13 years in twelve months. A separate question, about when every occupation could be fully automated, moved even further, from 2164 to 2116¹.

Ask a leadership team the same question and you will hear the same spread, held with more confidence. That is the problem this chapter addresses. If the experts cannot agree on the decade, an executive team cannot responsibly write a five-year AI plan as if it knew the destination. It still has to commit money and people this year.

Prepare, do not predict

The forecast is not the only thing that moves. Capability, cost and the rules can each change faster than an annual planning cycle: in July 2026, for example, the EU moved the date for its high-risk AI rules by sixteen months, in an amendment published nine days before they were due to apply².

The answer is not to stop planning. It is to plan differently. A strategy built for uncertainty separates what can be measured from what cannot, sketches a few plausible futures, sorts every candidate move by how it performs across them, and matches the speed of each decision to the cost of reversing it. The earlier chapters of this module have already given you the instruments that watch for change: the capability thresholds of Where AI Is Going Next and the use-by-market map of AI, Regulation, Geopolitics and Global Competition. This chapter is about deciding while those instruments are still moving.

Risk is not uncertainty

In 1921 the economist Frank Knight drew the distinction that every executive needs here. Risk is a situation in which you do not know the outcome but can estimate the odds. Uncertainty is a situation in which you cannot estimate the odds with any confidence. A measurable uncertainty, Knight wrote, is so different from an unmeasurable one “that it is not in effect an uncertainty at all”³.

The distinction matters because the tools differ. Risks can be priced, tested, insured and controlled, which is what the governance and risk modules of this course taught. Uncertainties cannot be controlled away, and treating them as if they could produces false precision: a business case that promises, say, a 32 percent cost reduction in year three when nobody knows what the models, prices or rules will be in year three.

Two habits follow. First, convert uncertainty into risk wherever you cheaply can. You cannot know how good AI will be at your work in 2029, but you can measure how good it is today on a few hundred of your own cases, and measure it again every quarter. Second, for what remains unmeasurable, stop asking for a point forecast. Ask for a range, or for a probability that can be scored later, the discipline Philip Tetlock found in the best forecasters⁴.

Four levels of uncertainty

Uncertainty is not all of one kind. In a widely used article, Hugh Courtney, Jane Kirkland and Patrick Viguerie of McKinsey distinguished four levels of what remains unknown after you have done your best analysis⁵.

AI offers an example of each. At level 1, the direction is clear even if the rate is not: Epoch AI found that the price of reaching a fixed level of model performance fell between 9 and 900 times a year, depending on the task⁶. You can plan on cheaper inference without knowing exactly how much cheaper. At level 2, there are a few discrete outcomes. Until July 2026, European firms faced two futures: the high-risk rules would apply in August 2026, or they would be deferred. At level 3, the outcome lies somewhere on a range: when AI systems will complete a multi-hour task in your process reliably enough to run without a person checking each step. At level 4, even the range is unclear, which is where the survey that opened this chapter sits.

The authors warned against a binary view in which managers treat the future as either certain or unknowable. The first error produces a single forecast and a plan that breaks when it fails. The second produces paralysis, or decisions made on gut feel because “nobody can know”. Many AI choices sit in the middle levels, where disciplined analysis of a few futures is both possible and worth the effort.

Scenarios test the strategy

The best-known corporate use of scenarios began at Royal Dutch Shell. By 1972 its planners were warning management of a sharp rise in oil prices that, in the company’s own words, “may take place at any moment in the next few years”. When the embargo came in October 1973, Shell’s account is that it was able to respond more swiftly than its competitors⁷. That is the company’s history of itself and should be read as such, but the lesson Pierre Wack, who led the work, drew from it is widely accepted. Scenarios are not forecasts. Their purpose is to change the assumptions in decision-makers’ heads before events force the change⁸.

For AI, the practical method is short. Choose the two uncertainties that would change your strategy most, not the two that are most discussed. For many organizations they are how fast AI becomes reliable on their own work, and how free they will be to deploy it: the combined effect of regulation, client acceptance and labor relations. Cross them and you have four futures.

Two differences from tools you have met elsewhere in the course are worth keeping clear. A business case’s sensitivity analysis, as in Building the AI Business Case, varies the numbers inside one future. A roadmap’s signposts, as in Building the Multi-Year AI Roadmap, watch the assumptions a chosen plan rests on. Scenarios come before both. They ask whether the plan itself is the right one, by checking it against futures the organization would rather not think about.

Three kinds of move

Once you have the futures, sort every candidate move by how it performs across them. Courtney and his colleagues named three kinds⁵. No-regret moves pay off whatever happens. Options are small commitments now that secure the right to act at scale later, limiting losses if the bad futures arrive. Big bets are large commitments that pay off handsomely in some futures and lose heavily in others.

The table above is illustrative; the cells are judgments, not data. A table like it is useful because it changes the conversation. Instead of arguing about which future will happen, the team argues about how each move performs in each one, which is a question people can answer. Teams that do the exercise often find the no-regret list longer than they expected, and the big bets fewer.

In AI, no-regret moves tend to be capabilities rather than products: a test set drawn from your own work, clean and permissioned data, people trained to check AI output, and the habit of measuring cost per completed task. Options are pilots, short contracts, a small team building expertise in a new technique, or one product line priced a new way. Quick Wins, Strategic Bets and Transformation Initiatives showed how to fund such options in stages; the point here is that a move’s kind depends on the futures, not on its size. A large investment in data that pays in every quadrant is a no-regret move. A small contract with a lock-in clause can be a big bet.

Match the speed to the door

The second sorting question is how hard a decision is to undo. Jeff Bezos put it memorably in his 2015 letter to shareholders. Some decisions are “one-way doors”, consequential and nearly irreversible, and should be made “methodically, carefully, slowly”. Most are “two-way doors”: if you are wrong, “you can reopen the door and go back through”, so they should be made quickly by small groups. His warning was about the opposite error. Large organizations tend to apply the heavy one-way process to two-way decisions, and the result is “slowness, unthoughtful risk aversion, failure to experiment sufficiently”⁹.

Where Should We Start With AI? used the door to choose a first project. At the level of strategy it does more work. Under high uncertainty the value of reversibility rises, so the door test should shape contracts and designs, not only the order of projects. Many AI one-way doors can be rebuilt as two-way: an exit clause and data-return terms in a platform contract, a design that lets you swap models, as The Evolution of AI Models recommended, or a hiring pause in place of a cut to the junior roles that train tomorrow’s experts. What cannot be made reversible, such as handing over data you can never take back or a public commitment that defines your brand, deserves the slow, senior process.

Say what you know, and what would change your mind

The last discipline is how leaders talk about all this. Boards and teams tend to hear confidence as competence, so uncertain claims are often presented as facts. A better practice is to separate four things in every AI paper: what we know from measurement, what we believe and why, what we are testing and by when, and what evidence would change our mind. The fourth is the one most often missing, and the one that makes changing course look like learning rather than failure.

Two failure modes sit on either side. Hype treats a level 3 uncertainty as settled and commits early. Paralysis treats it as level 4 and waits. The middle path is to keep no-regret moves running, keep options alive, and let evidence promote an option into a bet. Where AI Is Going Next gave the capability thresholds that tell you a bet has become safe; the roadmap’s signposts tell you which plan to revisit.

Story: a voice bot and a one-way door

In June 2025 the Commonwealth Bank of Australia introduced an AI voice bot in its call centers. It identified and verified callers and handled simple requests such as balance checks. In July the bank said that 45 roles in its direct banking team would go, explaining that “by automating simple queries, our teams can focus on more complex customer queries that need empathy and experience”¹⁰. The bank’s case, as the union later described it, was that the bot had cut call volumes by 2,000 a week¹¹.

The Finance Sector Union disputed the claim. It said the bank had not provided evidence for the 2,000 figure, and on 7 August it announced that it had taken a dispute to the Fair Work Commission¹¹. Its members reported that call volumes were rising, with overtime offered and team leaders pulled onto the phones¹². On 21 August the bank reversed the decision. Its statement said that its “initial assessment that the 45 roles in our Customer Service Direct business were not required did not adequately consider all relevant business considerations and this error meant the roles were not redundant”. It had apologized to the employees and acknowledged that it “should have been more thorough in our assessment of the roles required”. The 45 could stay in their roles, be redeployed or take the redundancy package¹³. The bank did not say whether the bot had underperformed or something else in the assessment had gone wrong, and reporting at the time found it unclear¹⁴.

Read the case through this chapter. Two decisions were bundled together. The voice bot was a two-way door: a system that can be tuned, scaled back or switched off. The redundancies were a one-way door: people who have left cannot be called back at modest cost, and their knowledge of the customers leaves with them. The question that joined the two, how demand would settle once callers met the bot, was an uncertainty only a few weeks old, and it could have been turned into a measured risk by waiting for more months of call data. Put through the robustness test, the cut paid in only one future, the one in which the bot kept demand down. In the others it would cost the bank what it did cost: a dispute, an apology and a public reversal. The two-way version was within reach: run the bot as an option, hold the headcount decision behind a named threshold in call volumes, and redeploy rather than dismiss until the evidence arrived. This reading does not depend on knowing why the assessment went wrong. The lesson is in the sequence: the one-way door was opened before the evidence that would have justified it.

What this means for leaders

Uncertainty about AI will not resolve on a timetable that suits a planning cycle. The leaders who handle it well do not claim to know the future. They measure what can be measured, sketch a few futures that matter to their own business, and sort every move by how it performs across them. They start the no-regret moves at once, keep options alive and cheap, and save big bets for the moment evidence arrives. They decide two-way doors quickly, redesign one-way doors to be reversible where they can, and slow down for the rest. And they tell their boards which parts of the plan are fact, which are belief and which are experiment.

Check yourself

1. Risk and uncertainty are two words for the same thing.MythFact
2. Leading AI researchers’ forecast of human-level machine intelligence moved 13 years earlier between two annual surveys.MythFact
3. A good scenario exercise ends by choosing the most likely scenario.MythFact
4. A no-regret move is defined by being small and cheap.MythFact
5. Bezos warned that large organizations tend to treat reversible decisions as if they were irreversible.MythFact
6. Under high uncertainty, the safest choice is to wait until the technology settles.MythFact

Reflection: the bet you have not named

What comes next

Deciding well when nothing can be forecast with confidence is one half of leadership. The final chapter, Module 10 Synthesis — The Future AI Leader, draws these threads together into the posture they ask of an executive.