AI roadmaps have a unique challenge: the foundation (models, their capabilities) improves on its own schedule, outside your control. A 12-month roadmap written in detail assuming today's models is obsolete in six months. This post is the planning approach that accommodates model progress without surrendering to chaos — and the roadmap artifacts that survive model releases.
Three planning horizons
0-3 months: commit. Current model capability is known. Features deterministic enough to commit. Detailed specs appropriate here.
3-9 months: intent. Models likely to improve during this window. Plan by intent and capability requirement, not by specific implementation. Revise quarterly.
9-18 months: thematic. High-level bets about where the company is heading. Not specific features. Strategic direction that guides engineering investment.
Plan by capability, not by model
Instead of 'integrate GPT-5 for Feature X,' plan 'achieve accuracy bar Y on task Z using best-available model at time of ship.'
Swap models as they improve. Quarterly (or more often) model evaluation: has any new model improved quality enough to swap in? Is cost down enough to enable previously-infeasible features?
Decouple feature spec from model spec. Feature spec: what users can do. Model spec: how the AI is configured. These evolve on different timelines.
Handling model releases
New models often rewrite 3-6 month horizons. A capability that required a workaround is now native. A feature that was cost-prohibitive is now economic.
Process: quarterly model review. Evaluate frontier models against your eval sets. Identify shifts in capability that affect roadmap. Revise plan.
Avoid reactive thrash. Not every model release requires roadmap change. Discipline in evaluating rather than acting emotionally is the skill.
See vendor guide 2026 for the current landscape of major model providers.
Managing dependencies
Data dependencies. Features may depend on data collection, labeling, or user-generated content reaching certain volumes. Plan for dependencies, not just model arrivals.
Infrastructure dependencies. Moving from experimental to production AI requires observability, evaluation, cost monitoring infrastructure. Plan this.
Team dependencies. AI product work needs specific expertise. Hiring plans intersect with roadmap.
External dependencies. Regulatory changes, competitor launches, major partner events. Factor in, hedge where possible.
Capacity considerations
AI features require more ongoing engineering than SaaS features. Evaluation, monitoring, model updates, prompt iteration. Allocate capacity appropriately.
Typical: 30-50% of engineering capacity on ongoing maintenance and evolution of shipped features. Not new features.
Research-adjacent work. Some AI features require prototyping against uncertainty. Budget for research vs product work appropriately.
Roadmap artifacts
Committed features list. Specific, tested, scheduled. 0-3 month horizon.
Capability goals. 'Enable users to do X at quality Y by Q4.' 3-9 month horizon. Implementation details flexible.
Strategic themes. 'Become the trusted AI layer for [vertical].' 9-18 month horizon. Directional.
Eval set evolution plan. What eval sets are you building, what quality bars are you raising, what capability expansions are you pursuing? This is AI-specific artifact.
Communication patterns
Board and executives. Thematic communication. Capability milestones. Quality measures. Business outcomes. Avoid getting into model-specific details.
Customer communication. Capability language ('our AI now does X'), not model language ('we use Claude 4.7'). Customers care about outcomes.
Team communication. More specific. Engineering teams need to know model choice, cost targets, quality bars, integration points.
Common planning failures
Over-committing to frontier. Planning features that assume capability that doesn't quite exist yet. Slippage results.
Under-committing to capacity. Ongoing maintenance underestimated. New feature velocity slows as debt accumulates.
Ignoring cost trajectory. Features planned assuming current costs; cost drops make new things possible; plan doesn't adapt.
Over-detailed long-term plans. Quarter-by-quarter detailed specs that are wrong by the time they arrive. Lose credibility with team.
Closing
AI roadmapping is an exercise in managing uncertainty without becoming paralyzed. Short-term commits, mid-term intents, long-term themes, quarterly reviews. See prioritization post.