A credible AI roadmap in the AI industry sequences foundations before scale rather than chasing capability first. The pattern that survives contact with production is a four-quarter arc: build the data and evaluation foundation, ship two or three governed pilots, harden the winners into production with cost and safety controls, then scale under governance. Skipping the foundation is why 70 to 80 percent of pilots stall. This playbook lays out a phased plan across four quarters, the exit criteria for each phase, and the sequencing that gets an AI-native organization from foundation to governed scale without pilot purgatory.
Sequence foundations before scale, or repeat the pilot graveyard
Most AI roadmaps fail the same way: they start with a flashy capability, demo it, and then discover there is no data foundation, no eval harness, and no governance to carry it to production. That is the mechanism behind the 70 to 80 percent pilot stall rate. A roadmap that survives contact with reality inverts the order. It treats the data and evaluation layer as quarter-one work, because without a vector store, a golden eval set, and lineage, every later phase is building on sand. Frontier model access is assumed, not a milestone; the milestones are the surrounding infrastructure and the discipline that turns experiments into shippable systems.
The right shape is a four-quarter arc with hard exit criteria between phases. Quarter one builds the foundation and picks the use cases. Quarter two ships two or three governed pilots against real baselines. Quarter three hardens the winners into production with cost engineering, safety evals, and monitoring, and kills the losers. Quarter four scales what works under full governance, with an inventory, approval checkpoints, and audit logs in place. Each phase has a gate: you do not advance until the exit criteria are met. This sequencing is what lets an AI-native organization reach governed scale in a year instead of accumulating a graveyard of demos that never shipped. The roadmap is not a list of features; it is a set of gates that force foundations, evidence, and governance in the right order.
The four-quarter AI roadmap
Each quarter has a theme, the core work, and an exit criterion that must be met before the next phase begins. The gates are the point: they prevent scaling on an unproven foundation.
| Quarter | Theme and core work | Exit criterion |
|---|---|---|
| Q1 | Foundation: vector infra, golden eval set, lineage, use-case selection | Retrieval measured, eval set live, top three use cases funded |
| Q2 | Governed pilots: ship two or three against real baselines | Each pilot has a numeric baseline and passing evals |
| Q3 | Harden: cost engineering, safety evals, monitoring; kill losers | Winners meet cost, latency, and safety bars in production |
| Q4 | Governed scale: inventory, approval checkpoints, audit logs | Every scaled system classified, monitored, and human-gated |
Execute the roadmap phase by phase
- Spend quarter one on the data and evaluation foundation and use-case selection, treating frontier model access as assumed rather than a milestone.
- Limit quarter two to two or three governed pilots, each with a numeric baseline defined before any code, so lift is provable.
- In quarter three, harden only the pilots that showed value, adding cost engineering, safety evals, and monitoring, and kill the rest without sentiment.
- Reserve quarter four for scaling under governance, with a model inventory, human approval checkpoints, and queryable audit logs in place first.
- Enforce the exit criterion at each gate, and do not advance a phase until it is met, so no capability scales on an unproven foundation.
How AI roadmaps derail
- Leading with capability and postponing the data and eval foundation, which guarantees the demos stall on the way to production.
- Running too many parallel pilots, so none get the depth to reach a real baseline and the portfolio produces no clear winners.
- Skipping the kill decision in quarter three, letting weak pilots consume engineering that the winners needed to harden.
- Scaling before governance is in place, then retrofitting inventory, checkpoints, and audit logs under regulatory pressure at far higher cost.
Track roadmap progress against gates
- Gate pass rate: share of phases that met their exit criteria on schedule, the honest read on execution discipline.
- Pilot-to-production conversion: proportion of quarter-two pilots that reached governed scale, targeted well above the industry norm.
- Foundation coverage: retrieval quality and eval-set completeness at the end of quarter one, the leading indicator for everything after.
- Governed-scale share: proportion of production AI systems with inventory, approval checkpoints, and audit logs by year end.
Frequently asked questions
Why start with data and evals instead of a flagship AI feature?
Because the flagship feature will stall without them. The data foundation and eval harness are what carry a demo to reliable production, and their absence is the main reason 70 to 80 percent of pilots fail. Starting with the foundation feels slower in quarter one but is far faster to governed scale, because every later phase builds on solid ground instead of collapsing on contact with real usage.
How many pilots should we run at once?
Two or three, no more. The temptation is to run many and see what sticks, but spreading engineering thin means no pilot gets the depth to reach a real baseline, and the portfolio produces no clear winner to harden. A focused set of two or three, each with a defined baseline and passing evals, converts to production far more reliably than a broad scatter of shallow experiments.
When do we put governance in place on the roadmap?
Governance foundations belong before scaling, in quarter three and locked by quarter four, not retrofitted afterward. Scaling first and adding inventory, approval checkpoints, and audit logs later is far more expensive and usually happens under regulatory pressure. Building the checkpoints and logs as the winners harden means governed scale is the default state rather than an emergency remediation project.
Related reading
Go deeper on this sector and topic.