In 2009, the director of a state motor vehicle agency approved a fifteen-year contract for a new licensing system. It was a sound decision by the logic of 2009. The system assumed every transaction happened at a physical counter, that identity meant a paper document, and that fraud was something a clerk could spot by eye. By 2024 her successor was trapped inside it: no digital identity, no way to plug in modern fraud detection, no path to mobile licensing without paying the vendor a fortune for changes the original contract never anticipated. The director had not made a bad decision. She had made a decision with a five-year imagination for a fifteen-year commitment. Government leaders make that mistake constantly, and AI is about to make it far more expensive.

Most of what you procure, build, and codify today will still be running long after you have moved on. Thinking in long horizons is not futurism for its own sake. It is the discipline of not trapping your successors, and of making the assumptions behind a long commitment visible while there is still time to argue about them. This lesson gives you a structured way to plan across 5-, 10-, and 25-year horizons without pretending you can predict the future.

Why Prediction Fails and Scenarios Help

Nobody can tell you the state of AI twenty-five years out, and anyone who offers to is selling something. The honest tools are not forecasts but scenarios: structured, plausible stories about different ways the future could unfold, built so that you can stress-test today's decisions against each one. The goal is not to be right about the future. It is to surface the assumptions your decisions are secretly resting on, so that a choice which only works in one future is recognized as a bet before you make it rather than after.

The technique comes from defense and energy planning, where leaders routinely commit to assets that last decades under deep uncertainty. You build a small set of distinct futures, then ask of every major decision whether it still makes sense in each. That question does something a forecast cannot: it converts a private hunch about how things will go into an explicit statement that colleagues can challenge. Most of the value arrives during the argument, not in the document produced at the end.

Thinking in Three Horizons

Different decisions have different time signatures, and treating them all with the same planning method is how agencies end up over-analyzing a routine software choice and under-analyzing a fifteen-year one. A useful shorthand is that the near horizon is about current capability and deployment, the middle horizon is about discontinuities and platform shifts, and the far horizon is about transformative capability and the institutions that would have to cope with it.

The five-year horizon: extend the present

Five years out, the landscape is a recognizable extension of today, and you can plan with reasonable confidence. Tools get better, cheaper, and more embedded in ordinary software that you did not choose to buy for its AI. The planning task here is concrete and unglamorous: workforce skills, data infrastructure, procurement flexibility, and governance that can keep pace with changes arriving through vendors rather than through your own roadmap. The motor-vehicle director's failure was a five-year-horizon failure, entirely foreseeable and entirely preventable.

The ten-year horizon: prepare for divergence

Ten years out, futures begin to genuinely diverge, and the divergence is usually driven by platform shifts rather than by incremental capability. The question is less how good models get and more what everything is built on: where computation happens, what an interface is, which layer your vendors sit in. The planning task shifts from prediction to optionality. Build capabilities and write contracts that keep choices open, and avoid commitments that only pay off in one version of events. This is where modular architecture and contracts that permit substitution earn their keep.

The twenty-five-year horizon: shape values and institutions

Twenty-five years out, the technical specifics are unknowable, but the questions are not. Which decisions should always involve a human? How does democratic accountability survive when systems are more capable at a task than the people overseeing them? What does your institution intend to protect regardless of what the technology can do? At this horizon you are not planning systems. You are planning principles, institutions, and the mechanisms by which they can be revised, which will outlast every platform decision above them.

Building Four Scenarios for Government AI

A useful scenario set is small and distinct, since a set large enough to cover everything is one nobody will use. Four is a common, workable number. Vary them along the dimensions that matter most for public institutions, which are how capable the technology becomes and how it is governed and distributed, rather than along the technical distinctions that interest specialists.

• Steady progress. Capability improves incrementally and predictably, and governance broadly keeps pace. The main task is competent, continuous adoption, and the planning emphasis falls on skills and integration.
• Rapid transformation. Capability advances far faster than expected, approaching or reaching the broad, human-level ability some call artificial general intelligence. Institutions struggle to keep up, and the planning emphasis falls on adaptability and on firm guardrails around consequential decisions.
• Fragmented and contested. Powerful systems exist but are unevenly distributed and poorly governed, concentrated among a few private holders and rival states. The planning emphasis falls on sovereignty, security, and equity of access.
• Constrained and cautious. Public backlash, regulation, or technical limits slow things sharply. The planning emphasis falls on not stranding investments that assumed the optimistic case.

Write each one as a short narrative rather than a bullet list, and give it a name people will actually say out loud in a meeting. The point of the naming is practical: once a scenario has a name, a colleague can say that a proposal only works in steady progress, and the conversation moves immediately to the assumption rather than to the personalities.

Running a Real Decision Through the Set

Now the payoff. Take an actual decision in front of you, say a ten-year cloud and data platform contract, and run it against all four. Under steady progress and rapid transformation, modularity and portability matter most, because you will want to substitute components you cannot currently name. Under fragmentation, data sovereignty and security terms dominate. Under constraint, what you want is the ability to scale down without penalty and to stop paying for capacity you no longer need.

A contract that is modular, portable, sovereign, and scalable performs acceptably in all four, which is a better position than one optimized for the future you personally consider most likely. Notice what this exercise did and did not require. It required no view about which scenario will happen, and it produced a list of contract properties you can hand to the people writing the terms. That is the practical output of scenario work: not a picture of the future, but a specification you can act on this quarter.

What Scenario Work Does Not Give You

Be precise about the limits, because scenario planning is unusually easy to oversell inside an organization that wants reassurance. A decision that survives all four of your scenarios has survived the four futures you were able to imagine. Your set is itself an assumption, built from what your team currently believes is plausible, and it will systematically omit the developments nobody in the room had heard of. Robustness across a scenario set is evidence that a decision is not fragile in the obvious ways. It is not a certificate.

The same caution applies to the properties you derive. Modular, portable, sovereign, and scalable are contract words until somebody has tested them. A portability clause that has never been exercised tells you what a vendor agreed to write, not what you would actually receive during a migration under time pressure with the incumbent's cooperation now optional. Where a scenario test tells you portability is load-bearing, the follow-on work is to prove it during the contract rather than at the end of it.

Finally, resist the version of this exercise that produces a document and no decisions. If nothing about a proposal changed as a result of running it through the scenarios, either the proposal was genuinely robust or the exercise was decorative, and those two look identical in the minutes. The honest test is whether anyone can name a term, a phasing, or a commitment that is different because the analysis happened.

Wild Cards and the AGI Question

Scenarios cover the plausible middle. Wild cards are the low-probability, high-impact events outside it: a sudden capability jump, a catastrophic security failure in a widely deployed system, a development that makes a core government function obsolete faster than the workforce supporting it can be redeployed. You cannot plan for each specifically, and pretending to produces a very long document. What you can build is institutional resilience: the capacity to sense change early, decide quickly, and adapt without breaking. Resilience reduces the damage and shortens the recovery. It does not neutralize the event, and an organization that believes otherwise has simply moved its overconfidence somewhere less visible.

Artificial general intelligence, meaning systems that match or exceed human ability across most tasks, is the wild card leaders most want certainty about and can least get. The disciplined posture is neither to dismiss it nor to reorganize everything around it. Instead, identify the decisions that would matter most if it arrived, principally which choices must always keep a named human accountable, and make those commitments now as a matter of institutional values rather than as a bet on any timeline. Commitments made this way have a useful property: they are worth having even if the capability never arrives, because the same reasoning applies to any system that becomes better at a task than its overseers.

Anchor that in the principle that risk management should scale with potential consequence, which is the organizing idea behind voluntary risk frameworks such as the NIST AI Risk Management Framework. The higher the stakes of a decision, the more human oversight and reversibility you build in, whatever the technology turns out to be capable of. That rule requires no forecast at all, which is exactly why it survives one being wrong.

Designing Governance That Outlives the Technology

The far horizon is where governance design belongs, because governance is the only thing you build that can plausibly still be functioning in twenty-five years. The systems will be replaced several times over. The question is whether the arrangements that constrain them are written so that they can survive being applied to technology their authors never saw.

Two properties matter more than any specific rule. The first is that the arrangement should be stated in terms of consequences rather than of mechanisms. A policy written about a particular technique becomes obsolete when the technique does, while a policy about decisions that affect a person's rights, benefits, or safety continues to apply to whatever makes those decisions. The second is that the arrangement must contain its own revision mechanism: a stated cycle, a named owner, and a route by which evidence from operation changes the text. Governance without a revision path does not stay stable. It stays fixed while the world moves, and is then abandoned wholesale by someone who finds it unworkable, which loses the reasoning along with the rules.

Write down why each commitment exists, not only what it requires. The successors who inherit your framework will face situations you did not anticipate, and the reasoning is what lets them extend it correctly. Requirements without recorded reasoning get either applied mechanically where they do not fit or discarded entirely, and both failures look like the framework's fault rather than the drafting's.

What Other Governments Are Betting

Three national postures from the source material illustrate different long-horizon bets. Each is described as it stood when that material was written, and positions in this area have been moving quickly, so confirm any country's current stance before relying on it. What is useful here is not the accuracy of the snapshot but the structure of the wager each represents.

A light-touch regulatory posture. The source describes the UK as having chosen not to regulate AI proactively, watching for problems instead. The implicit bet is that incremental improvement is more likely than discontinuity. That works if development stays slow and predictable; if a discontinuity arrives, whether a breakthrough capability or an economic disruption, a posture built on watching may not have the instruments ready when they are needed. Named as a scenario, this is a strategy optimized for steady progress.

An unusually long planning lens. The source describes the UAE as planning AI investment within a fifty-year strategy for diversifying its economy. A horizon of that length is uncommon in government and reflects a judgment that the technology is transformative enough to warrant sustained commitment across many political cycles. The transferable point is not the number. It is that the planning horizon was set by the character of the change rather than by the budget cycle.

Technology treated as an ecosystem rather than a sector. The source describes China as treating AI as part of a broader technology programme alongside biotechnology and quantum computing, on the reasoning that each amplifies the others. This differs from the siloed approach most governments take, in which AI strategy, biotechnology strategy, and computing strategy are written by different offices with different timelines. Whatever one makes of the specific case, the structural question it raises is fair: if capabilities compound across fields, a strategy that treats one field alone will systematically underestimate what is coming.

Read together, the three make a point that no single one makes. Each government has placed a bet about the shape of change, and each bet is legible from its institutional arrangements whether or not anyone stated it. The same is true of yours. If nobody has written down which future your current strategy assumes, that assumption still exists, and it is being made by whoever wrote the most recent procurement.

Making Horizon Thinking Routine Rather Than Heroic

The reason most agencies do this work once is that it arrives as an event. Someone runs a futures workshop, the deck circulates, and the practice ends when that person's attention moves on. Horizon thinking only changes outcomes when it is attached to the ordinary machinery through which decisions already pass, which means deciding three dull things: what triggers the stress test, who owns the scenario set, and when the set gets refreshed.

Set the trigger by the length of the commitment rather than by its cost, since the expensive decisions receive scrutiny already and the long ones frequently do not. A modest contract that will shape how a service works for a decade deserves the test more than a large one that expires shortly. Write the threshold into the approval process itself, so that the analysis is a required section of the paper rather than a favor someone asks for, and keep it short enough that a program manager can complete it without help.

Give the scenario set a named owner, because a set nobody maintains silently ages into a description of the recent past and is then quietly ignored. Refresh it on a stated cycle and additionally after anything that genuinely surprised the organization, since surprise is the clearest available evidence that the set was missing a future. When you revise, keep the old version and record what changed and why. That trail is the most useful artifact the whole practice produces: over a few cycles it shows you what your institution has repeatedly failed to anticipate, which is far more actionable than any individual forecast.

Then protect the practice from its own success. The moment a scenario exercise is used to justify a decision someone already wanted, colleagues learn that the output is decorative and will stop investing effort in it. The defense is procedural and cheap: record where a proposal failed a scenario and what was done about it, and make it acceptable to approve something that fails one, provided the failure is named. A process that only ever confirms proposals is not a test.

Long-Horizon Decision Stress-Test

Before committing to any major, long-lived decision, whether a multi-year contract, a platform, a policy, or an organizational redesign, run it through this worksheet. It is deliberately short, because a stress test long enough to be thorough is one that gets skipped on the decisions that most need it.

• Horizon. How long will this decision actually bind the agency, counting renewals and switching costs rather than the stated term? Match your scrutiny to the commitment, not to the budget cycle.
• Scenario test. Does this hold up under steady progress, rapid transformation, fragmentation, and constraint? Write down where it fails, since a decision that fails in one scenario may still be right if you know which one.
• Reversibility. If this turns out wrong in five years, how hard and expensive is it to undo, and who would have to agree? Prefer reversible over optimal.
• Optionality preserved. Does this keep future choices open, or does it lock the agency into one vendor, one architecture, or one assumption about how the work is done?
• Tested, not promised. For each property you are relying on, such as portability or the ability to scale down, has anyone exercised it, or does it exist only as a clause?
• Values protected. Whatever the technology does, does this preserve human accountability, due process, and equity of access?
• Sensing built in. Will the agency notice early if the assumptions behind this decision start to break? Name who is watching and against which signals.
• Successor test. Will the person in your seat in ten years thank you or curse you? Decide as though they are reading the file, because they will be.

Anti-Patterns to Avoid

• The five-year imagination behind a fifteen-year commitment. Applying the scrutiny appropriate to a short procurement to a decision that will constrain the agency for a decade and a half. The mismatch, not the decision, is what traps your successor.
• Scenarios as a certificate. Treating survival across your four futures as proof that a decision is safe. The set is built from what your team could imagine, and it omits by construction the developments nobody in the room had heard of.
• The untested clause. Relying on portability, exit, or scale-down rights that have never been exercised. What a vendor agreed to write and what you would actually receive during a migration are different things, and the gap is discovered at the worst moment.
• The decorative exercise. Running the scenario workshop, producing the deck, and changing no term, phasing, or commitment as a result. If nobody can name what is different because of the analysis, it did not happen.
• Organizing everything around one future. Rebuilding strategy on the assumption of imminent transformative capability, or on the assumption that nothing much will change. Both are bets sold as prudence, and both are unrecoverable if wrong.
• Waiting for the picture to clear. Deferring long-horizon decisions until the technology settles. It will not settle on your schedule, and deferral is itself a decision that hands the choice to whoever signs the next renewal.
• Policy written about a mechanism. Framing rules around a particular technique or product category, which guarantees obsolescence. Rules framed around consequences to people keep applying to whatever arrives next.
• Governance with no revision path. Publishing a framework without a stated review cycle, a named owner, or a route from operational evidence to changed text. It does not stay stable; it stays fixed until someone discards it entirely, taking the reasoning with it.
• Requirements without recorded reasoning. Leaving successors a list of obligations and no account of why each exists. They will face cases you did not anticipate and will either apply the rules mechanically where they do not fit or drop them.

Practice Prompts

• Find your longest commitment. Identify the decision currently binding your agency furthest into the future, counting renewals and switching costs. Establish what assumptions it encodes about how work will be done, and whether anyone has revisited them since it was signed.
• Write the four scenarios in your own words. Draft a short narrative for each of steady progress, rapid transformation, fragmentation, and constraint, specific to your mission rather than to AI in general. Give each a name colleagues will use in meetings.
• Stress-test a live decision. Take something awaiting approval and run it through all four scenarios and the worksheet above. Bring the result to the approval discussion and note which term changed as a result.
• Test one promised property. Pick a portability, exit, or scale-down right you are relying on and find out what exercising it would actually involve. Do this while the relationship is good.
• Draft the far-horizon commitments. Write down the decisions in your domain that must always keep a named human accountable, and the reasoning for each. Circulate it as a values statement rather than as a technology policy, and see whether it survives contact with the people who would have to honor it.

Reflection

Think about a decision your predecessors made that constrains you today, and be fair to them: reconstruct what was reasonable to believe at the time, and identify precisely where their imagination stopped short of their commitment. Then turn it around. Consider the longest-lived decision you have personally approved, and ask what assumption it rests on that would look naive to someone reading the file in fifteen years. Ask whether that assumption is written down anywhere, or whether it lives only in the heads of people who will have moved on. The uncomfortable part of this exercise is that the answer is usually available now, at low cost, and simply nobody has asked.

Glossary

• Scenario planning. A structured method for making assumptions visible and testable by writing a small set of distinct, plausible futures and evaluating decisions against each, rather than forecasting which will occur.
• Planning horizon. The period over which a decision constrains an organization, counting renewals and switching costs, which is frequently longer than the stated term.
• Discontinuity. A break in trend that invalidates plans built on extrapolation, as distinct from faster or slower movement along an expected path.
• Platform shift. A change in the underlying layer that systems are built on, which reshapes what is possible and who your dependencies are, independently of how capable any individual system becomes.
• Optionality. The deliberate preservation of future choices, purchased by accepting a less optimal position today in exchange for the ability to change course.
• Robust decision. One that performs acceptably across every scenario in your set, as opposed to optimally in the one you consider most likely. Robustness is bounded by the imagination of the set.
• Wild card. A low-probability, high-impact event outside the plausible middle that scenarios cover, addressed through institutional resilience rather than through specific plans.
• Artificial general intelligence. Systems that match or exceed human ability across most tasks. Treated here as a values question about which decisions require a human, not as a forecast.
• Reversibility. How hard and expensive it would be to undo a decision, and whose agreement that would require, which is often a better selection criterion than expected performance.
• Sensing. The named responsibility for watching specified signals that would indicate a decision's assumptions are breaking, without which early warning arrives as an incident.

Related Lessons

• Futureproofing: Preparing for Unknown AI develops the architectural and contractual practices that preserve optionality in build decisions.
• Emerging AI Technologies Assessment covers how to evaluate a new capability as it appears, which is the sensing function this lesson relies on.
• Emerging Architectures: Agentic AI and Compound Systems examines the kind of platform shift the ten-year horizon is concerned with.
• Legacy System Modernization for AI addresses what to do once a long commitment has already trapped you.
• Technical Debt Management in AI covers the accumulating cost of decisions that were reasonable when made.
• National AI Competitiveness takes up long-horizon positioning at the level of a nation rather than an agency.
• Building and Maintaining Public Trust develops the institutional asset that far-horizon commitments are meant to protect.
• AI Regulatory Design goes deeper on writing rules that stay applicable as the technology underneath them changes.

Closing

The motor vehicle director was not careless and was not unlucky. She was asked to make a fifteen-year commitment and was given, by custom and by process, the tools appropriate to a five-year one. Nothing in her organization prompted her to write down what the system assumed about identity, counters, and fraud, and so nobody noticed when every one of those assumptions stopped being true. Long-horizon work does not require you to see further than anyone else. It requires you to state what you are assuming, match your scrutiny to the length of the commitment rather than the budget cycle, and leave your successor the reasoning as well as the contract. Do that consistently and you will still be wrong about the future, which is fine, because you will be wrong in ways that can be corrected.

Key Takeaways

• Plan for the length of your commitments, not your tenure. Most of what you procure and codify outlives you, and a five-year imagination behind a fifteen-year commitment is what traps your successors.
• Use scenarios to make assumptions visible. The value is not a picture of the future but an explicit statement of what a decision is quietly betting on, which colleagues can then challenge.
• Treat the three horizons differently. Plan systems and skills at five years, preserve optionality against platform shifts at ten, and design values and institutions at twenty-five.
• Build a small set of named futures. Steady progress, rapid transformation, fragmentation, and constraint let a colleague say which scenario a proposal depends on without making it personal.
• Robust beats optimal, within limits. A decision that works across your whole set is stronger than one tuned to your favorite future, and it is still bounded by what your team could imagine.
• Promised properties are not tested properties. Portability, exit, and scale-down rights are contract words until someone has exercised them, and the difference surfaces under time pressure.
• Handle AGI as a values question rather than a timeline bet. Decide now which decisions must always keep a named human accountable, and note that the same reasoning holds for any system that outperforms its overseers.
• Scale oversight to consequence. The rule that risk management should rise with potential harm requires no forecast, which is why it survives every forecast being wrong.
• Write governance about consequences and give it a revision path. Rules framed around techniques expire; rules framed around effects on people endure, and both need a named owner and a route from evidence to changed text.

Frequently Asked Questions

How do we run scenario planning without it becoming a two-day offsite that changes nothing? Attach it to a real decision that is already scheduled for approval, rather than running it as a standalone strategy exercise. Give the group a live contract, platform choice, or policy and one afternoon, and require the output to be a list of specific changes: a term to add, a phase to shorten, a dependency to avoid. The discipline that makes it stick is holding the approval discussion afterward and asking which item on that list was adopted. Exercises with no decision attached produce decks, and decks do not survive the next reorganization.

Is it responsible to plan for artificial general intelligence at all, given that nobody knows whether or when it arrives? Plan for the decisions rather than the event. You cannot responsibly reorganize an agency around a capability with no established timeline, and you can responsibly decide now which categories of decision will always keep a named human accountable, which actions a system will never be permitted to take unsupervised, and what reversibility you require of consequential automation. Those commitments are worth making regardless, because the same argument applies whenever a system becomes better at a task than the people reviewing it, which is already true in narrow domains.

Our procurement rules make long, rigid contracts hard to avoid. What can we actually change? Usually more than it appears, because the rigidity is often in local practice rather than in the rules themselves. The recurring levers are shorter base terms with option periods rather than a single long commitment, requirements stated as outcomes so the delivery method can change, explicit rights to data and configuration in a usable form, and phased scope so that later stages are contingent on evidence from earlier ones. Where the rules genuinely bind, the remaining move is to document the assumptions in the file at the time of award, so that the successor who inherits the constraint at least inherits the reasoning.

How far ahead should we really plan when the technology changes every few months? The volatility of the technology is an argument for planning further out, not less, because it is precisely what makes long commitments dangerous. What should change with the horizon is the object of planning. At the near horizon you plan systems and skills. At the middle horizon you plan for substitution, so that a platform shift is an inconvenience rather than a rebuild. At the far horizon you plan institutions and values, which are the only things you can sensibly commit to when the technical specifics are unknowable. Planning nothing beyond the budget cycle does not avoid the uncertainty; it just leaves the choice to whoever signs the next renewal.