# Why Most AI Roadmaps Are Fiction
The Planning Fallacy in Enterprise AI Strategy
---
Abstract
Context: Organizations invest substantial effort in AI roadmapping—multi-year plans specifying which AI capabilities will be built, when they'll be delivered, and what value they'll create. These roadmaps inform investment decisions, resource allocation, and executive commitments.
Problem: AI roadmaps are systematically inaccurate. They fail to predict what organizations will actually build (use cases change as discovery reveals what's possible), when they'll build it (timelines compress or extend based on unpredictable factors), and what value it will create (outcomes depend on adoption and integration, not technology delivery). The planning fallacy—systematic overconfidence in predictions—is amplified in AI by technology uncertainty, organizational learning, and rapidly shifting capabilities.
Here we argue: That traditional roadmapping approaches are fundamentally unsuited to AI. AI initiatives involve discovery (what's possible?), not just delivery (how do we build it?). Roadmaps that assume known destinations prevent the learning that determines success. Organizations need planning approaches that embrace uncertainty, enable learning, and adapt as knowledge accumulates.
Conclusion: Effective AI planning replaces detailed roadmaps with outcome objectives, investment theses, and decision frameworks that guide action while preserving optionality. Organizations should plan for learning rather than planning for delivery, treating AI strategy as portfolio management rather than project scheduling.
---
1. Introduction: The Roadmap Ritual
Every year, organizations produce AI roadmaps. These documents specify multi-year plans: which AI capabilities will be developed, which use cases will be addressed, which milestones will be achieved, and which business outcomes will result.
These roadmaps serve important organizational functions. They coordinate expectations across stakeholders. They justify investment requests. They demonstrate strategic intent. They provide executives with the comfortable feeling that AI transformation is under control.
There is only one problem: the roadmaps are fiction.
1.1 The Evidence of Roadmap Failure
Consider the evidence:
Use case volatility. Organizations rarely build what their roadmaps specify. MIT found that 60% of organizations evaluated AI tools, but only 20% reached pilot stage and just 5% reached production. The attrition occurs because roadmapped use cases prove infeasible, lower-value than expected, or superseded by better opportunities discovered along the way.
Timeline inaccuracy. AI project timelines are notoriously unreliable. Gartner found that 30% of generative AI projects will be abandoned after proof of concept—not delayed, but cancelled. The roadmap assumption that projects proceed through predictable phases to predictable completion is systematically violated.
Value uncertainty. Roadmaps typically include business case projections: "This AI capability will generate $X million in value." Yet over 80% of organizations report no tangible EBIT impact from AI. Value projections in AI roadmaps are aspirational, not predictive.
1.2 The Planning Fallacy Amplified
The planning fallacy—the tendency to underestimate time, costs, and risks while overestimating benefits—affects all project planning. But AI amplifies this fallacy through multiple mechanisms:
Technology uncertainty. AI capabilities evolve rapidly. Roadmaps made in 2023 couldn't anticipate GPT-4's capabilities; roadmaps made in 2024 couldn't anticipate agentic AI's emergence. Planning assumes stable technology; AI delivers constant disruption.
Discovery requirements. Unlike traditional software, AI requires discovery: Is this use case feasible? Will the model perform adequately? Will users adopt? Roadmaps assume these questions are answered before planning; in reality, they're answered through building.
Integration complexity. AI value depends on workflow integration, which depends on organizational factors that can't be known in advance. A technically successful AI capability may create no value because integration fails. Roadmaps can't predict organizational receptivity.
Compound uncertainty. AI initiatives face technical uncertainty (will it work?), adoption uncertainty (will users use it?), integration uncertainty (will it fit workflows?), and value uncertainty (will it create impact?). Each uncertainty multiplies; the compound result is profound unpredictability.
---
2. Why Roadmaps Fail for AI
Understanding why roadmaps fail requires examining the assumptions embedded in traditional roadmapping and why those assumptions don't hold for AI.
2.1 The Roadmap Assumptions
Traditional roadmaps assume:
Known destinations. We know what we want to build. The roadmap specifies the path to predefined endpoints.
Predictable paths. We can estimate effort, dependencies, and timelines. Planning enables resource allocation and coordination.
Stable scope. What we plan to build today is what we'll want to have built tomorrow. Requirements persist through execution.
Linear value. Value is proportional to delivery. Build more, get more.
2.2 Why These Assumptions Fail
For AI, each assumption is violated:
Destinations are discovered, not known. AI success depends on finding use cases where AI creates value—and this requires experimentation, not just execution. Organizations don't know in advance which AI applications will work, be adopted, and create impact. Roadmaps that specify destinations prevent the exploration that reveals where value actually lies.
Paths are unpredictable. AI development involves surprises: models that don't perform as expected, data that doesn't support intended use cases, users who reject technically successful solutions. These aren't failures of execution but inherent properties of working with probabilistic, adaptive systems.
Scope should change. As organizations learn from AI experiments, they should pivot toward higher-value opportunities and away from lower-value ones. Roadmap fidelity—delivering what was planned—is a bug, not a feature. The goal is value, not plan adherence.
Value is nonlinear. AI value depends on adoption and integration, not just delivery. An AI capability that isn't adopted creates zero value, regardless of technical sophistication. Value emerges from user behavior changes, not from technology completion.
2.3 The Organizational Dynamics
Roadmap failure isn't just about uncertainty—it's about how organizations respond to roadmaps:
Commitment escalation. Once roadmaps are approved and communicated, organizations feel committed to delivering them. Pivoting away from roadmapped initiatives feels like failure, even when pivoting is the right strategic choice.
Resource lock-in. Roadmaps drive resource allocation. Teams are staffed, budgets are committed, and dependencies are established. Changing direction becomes expensive, so organizations persist with losing bets.
Accountability structures. Leaders are accountable for roadmap delivery. Admitting that the roadmap was wrong threatens careers. The incentive is to redefine success rather than acknowledge roadmap failure.
Planning theater. Organizations perform roadmapping because it's expected, not because it's useful. The roadmap satisfies governance requirements and executive expectations regardless of its accuracy.
---
3. What Good AI Planning Looks Like
If traditional roadmaps fail, what should replace them? Effective AI planning embraces uncertainty rather than ignoring it.
3.1 Outcome Objectives, Not Feature Roadmaps
Instead of specifying what AI capabilities will be built, define what business outcomes AI should improve.
Outcome objective example:
"Reduce customer support cost per interaction by 30% within 18 months through AI-assisted resolution."
This objective:
- Specifies the outcome (cost reduction) rather than the solution (chatbot, copilot, etc.)
- Sets a measurable target (30%) that can be tracked
- Provides time boundaries (18 months) while allowing path flexibility
- Permits multiple solution approaches
The team pursuing this objective might build a chatbot, a knowledge base, an agent assist tool, or something not yet invented. The roadmap doesn't constrain the solution; the outcome constrains the objective.
3.2 Investment Theses, Not Project Plans
Instead of planning projects, articulate investment theses about where AI might create value.
Investment thesis example:
"We believe AI-assisted code review can improve developer productivity by 20-40% based on early experiments. We will invest $X over the next 6 months to validate this thesis and determine whether to scale."
This thesis:
- Acknowledges uncertainty (20-40% range, not false precision)
- References evidence (early experiments)
- Bounds investment (6 months, $X)
- Defines decision criteria (scale or not)
Investment theses create option value. The organization learns whether AI creates value in a domain before committing to large-scale deployment. Failed theses are learning, not failure.
3.3 Decision Frameworks, Not Timelines
Instead of specifying when things will happen, define how decisions will be made.
Decision framework example:
"We will scale AI in domain X when we observe: (a) >90% user satisfaction in pilot, (b) >20% productivity improvement versus baseline, (c) <5% error rate with business impact. We will expand to domain Y if domain X achieves these criteria within 6 months."
This framework:
- Specifies criteria for progression
- Links future investments to demonstrated value
- Preserves optionality based on learning
- Doesn't commit to dates independent of results
3.4 Portfolio Management, Not Project Scheduling
Instead of treating AI initiatives as discrete projects to be scheduled, treat them as a portfolio to be managed.
Portfolio approach:
- **Explore investments:** Small bets on unproven AI opportunities. High failure rate expected. Goal is learning.
- **Scale investments:** Larger bets on validated AI opportunities. Lower failure rate. Goal is value capture.
- **Core investments:** Ongoing investment in proven AI capabilities. Goal is optimization.
The portfolio is rebalanced as experiments succeed or fail. Resources flow toward validated opportunities and away from failed hypotheses. No multi-year commitment locks resources into predetermined paths.
---
4. Implementing Adaptive AI Planning
Moving from traditional roadmaps to adaptive planning requires changes in process, governance, and culture.
4.1 Process Changes
Quarterly planning horizons. Replace annual roadmaps with quarterly planning cycles. Commit to 90-day goals while maintaining directional intent for longer horizons.
Discovery sprints. Before committing to build, invest in discovery: user research, technical feasibility, value estimation. Discovery sprints answer questions; development sprints build solutions.
Evidence-based progression. Define criteria for moving from exploration to scaling. Require demonstrated value, not elapsed time, as the progression trigger.
Kill criteria. Explicitly define conditions under which initiatives will be terminated. Make stopping as legitimate as starting.
4.2 Governance Changes
Outcome accountability. Evaluate AI leaders on outcomes achieved, not roadmap adherence. Pivoting toward higher-value opportunities should be rewarded, not penalized.
Portfolio reviews. Replace project status meetings with portfolio reviews. Ask not "Is this project on track?" but "Is this investment creating value? Should we increase, maintain, or decrease investment?"
Learning orientation. Treat failed experiments as learning, not failure. Post-mortems should ask "What did we learn?" not "Who's responsible?"
4.3 Communication Changes
Directional communication. Communicate direction and intent, not specific commitments. "We're investing in AI for customer service" is honest; "We'll deploy AI chatbots in Q3" may not be.
Uncertainty acknowledgment. Be explicit about what's known and unknown. "Based on current experiments, we believe..." is more honest than "Our roadmap shows..."
Progress measures. Report on learning and validated value, not on roadmap completion percentage. "We've validated that AI reduces handling time by 25%" is meaningful; "We're 60% through the roadmap" is not.
---
5. Case Study: From Roadmap to Adaptive Planning
A large healthcare organization had developed a detailed 3-year AI roadmap: 15 use cases across clinical, administrative, and operational domains, with specific delivery dates and projected ROI for each.
Eighteen months in, reality diverged dramatically from plan:
- 4 roadmapped use cases had been cancelled (technical infeasibility or low value)
- 3 use cases not on the roadmap had been developed (opportunities discovered through experimentation)
- 5 use cases were behind schedule (integration complexity exceeded estimates)
- 3 use cases were on track but showing no measurable value
The organization had achieved 70% "roadmap completion" while creating minimal business impact.
5.1 The Pivot
The organization abandoned traditional roadmapping for adaptive planning:
Outcome objectives defined:
- Reduce clinical documentation time by 25%
- Improve diagnostic accuracy for top 5 conditions by 10%
- Decrease administrative cost per patient by 15%
Investment theses articulated:
- "We believe ambient AI documentation can reduce clinician documentation time. We'll invest in a 50-clinician pilot to validate."
- "We believe AI-assisted imaging can improve radiology accuracy. We'll deploy in one facility to measure."
Portfolio established:
- Explore: 6 small experiments, $200K each, 90-day validation
- Scale: 2 validated initiatives, $2M each, 12-month deployment
- Core: 1 proven capability, $500K/year ongoing investment
Decision criteria specified:
- Move from explore to scale when pilot shows >15% improvement with >80% clinician satisfaction
- Kill initiatives that don't meet explore criteria within 90 days
5.2 The Results
After one year of adaptive planning:
- Clinical documentation time reduced 30% (exceeding 25% target)
- Three explore bets validated and promoted to scale
- Four explore bets killed without extended investment
- Total investment lower than roadmapped plan
- Actual value delivered higher than projected
The organization spent less and achieved more by embracing uncertainty rather than pretending to predict.
---
6. Implications for Leaders
6.1 For Executives
Demand honesty about uncertainty. AI roadmaps that claim certainty are lying. Insist on plans that acknowledge what's unknown and how learning will occur.
Reward adaptation. Leaders who pivot away from failing initiatives are making good decisions. Leaders who persist with the roadmap despite evidence are not.
Fund learning explicitly. Budget for exploration, experimentation, and potential failures. Organizations that only fund known wins won't discover new value.
6.2 For AI Leaders
Resist roadmap pressure. When stakeholders demand detailed multi-year plans, explain why such plans are unreliable and propose adaptive alternatives.
Build learning infrastructure. Invest in the measurement, experimentation, and feedback mechanisms that enable adaptive planning.
Communicate appropriately. Share direction, intent, and progress toward outcomes—not fictional timelines for fictional deliverables.
6.3 For Strategists
Reframe planning purpose. Planning should enable good decisions, not predict the future. Good AI planning creates decision frameworks, not Gantt charts.
Design for optionality. Preserve choices rather than foreclosing them. The value of planning is the ability to respond to learning.
Measure plan quality. Plans should be evaluated on the quality of decisions they enable, not their detail or confidence.
---
7. Conclusion: Embracing Productive Uncertainty
AI roadmaps fail because they impose certainty on inherently uncertain endeavors. They predict what can't be predicted, commit to what should remain flexible, and create accountability structures that punish learning.
The alternative isn't no planning—it's different planning. Outcome objectives define what success looks like without constraining how to achieve it. Investment theses create option value through bounded experiments. Decision frameworks guide resource allocation based on evidence. Portfolio management enables continuous rebalancing toward demonstrated value.
This approach requires organizational courage. It means admitting uncertainty rather than pretending confidence. It means evaluating leaders on outcomes rather than roadmap adherence. It means treating pivots as learning rather than failure.
Organizations that make this shift gain competitive advantage. They deploy resources toward value rather than toward predetermined plans. They learn faster because their planning enables rather than prevents experimentation. They achieve more because they're optimizing for outcomes, not deliverables.
The fiction of the AI roadmap is comfortable. The reality of adaptive planning is effective. The choice determines whether AI investment creates value or merely creates plans.
---
Extended References
- Kahneman, D. (2011). *Thinking, Fast and Slow*. Farrar, Straus and Giroux. Foundational work on the planning fallacy and systematic prediction errors.
- Cagan, M. (2018). *Inspired*. Wiley. Framework for outcome-driven product planning versus feature roadmaps.
- Ries, E. (2011). *The Lean Startup*. Crown Business. Principles of hypothesis-driven development and validated learning.
- McGrath, R. G. (2013). *The End of Competitive Advantage*. Harvard Business Review Press. Portfolio approach to strategy under uncertainty.
- MIT NANDA Initiative. (2025). *The GenAI Divide*. Evidence on AI initiative volatility and success factors.
---
Glossary
Investment Thesis: A structured hypothesis about where AI might create value, with bounded investment and clear validation criteria.
Outcome Objective: A measurable business result AI should achieve, specified without constraining the solution approach.
Decision Framework: Criteria for making AI investment decisions based on evidence rather than predetermined schedules.
Planning Fallacy: The systematic tendency to underestimate time, costs, and risks while overestimating benefits in planning.
Portfolio Management: Treating AI initiatives as a portfolio of investments with varying risk profiles, continuously rebalanced based on results.
---
*This article is the fourth in the Product × AI series. Previous: "AI Value Realization vs AI Adoption." Next: "AI as a Capability, Not a Feature"*