The roadmap isn’t a list of projects — it’s a sequence of bets ranked by value, risk, and learning. The wave-based approach ensures each stage validates what the next one needs to work.
Roadmap in waves — without a validation milestone, the next wave starts amid uncertainty.
🌊 Waves: crawl, walk, run
Each wave has a business objective, not just a technical deliverable. Crawl proves it’s possible and useful. Walk verifies that it scales operationally. Run is when the standard becomes standard practice.
Crawl — prove value
Minimum scope that answers the question: “Does this work here?” With real data and real users, but controlled volume. Expected result: yes or no, backed by data.
Typical duration: 4-8 weeks. Exit criterion: predefined metric achieved.
Walk — expand coverage
Expand what crawl validated to more use cases, more users, or more volume. Build reusable capabilities and integrate with more systems.
Typical duration: 8-16 weeks. Prerequisite: validated crawl.
Run — scale
The pattern is rolled out across the entire organization or operation. Monitoring and maintenance processes are already in place. The team knows how to operate it.
Typical duration: ongoing. Prerequisite: validated walk with a defined SLA.
⚠️ The most costly mistake
Go straight to Run without going through Crawl. Rolling back a solution deployed across the entire operation without validation costs 10× more than a pilot that failed.
🔗 Dependencies: data and foundation first
The biggest source of delays in AI projects isn’t the model—it’s unmapped dependencies: data that doesn’t exist or isn’t good enough, integrations that take months, permissions that require legal approval.
Common dependencies
- 🗄️Data: existence, quality, access, privacy
- 🔌Integrations: Legacy system APIs, CRM, ERP
- 👥Team: technical capability, training, time
- ⚖️Governance: compliance, LGPD, approvals
Recommended order
💡 Practical tip
Map dependencies in the first sprint, before writing code. Use a simple table: dependency, owner, deadline, blocker. Review weekly. Invisible dependencies are the biggest schedule risk.
🏆 Sequence by value, risk, and learning
When there are multiple candidate use cases, the sequencing criterion isn’t “the most exciting one” or “what the CTO asked for.” It’s a a combination of potential value, execution risk, and what the case teaches for the next ones.
📐 Sequencing criteria
| Criterion | What to assess | Ideal for the first wave |
|---|---|---|
| Value | How much impact if it works (R$ or time) | Visible and meaningful, but not critical |
| Risk | Chance of failure × cost of error | Controllable, errors are tolerable |
| Learning | What this case teaches us for next time | High — will unlock reusable capabilities |
🧱 Reusable capabilities
A consultant who thinks about the platform before the feature reduces the marginal cost of each subsequent wave. A reusable capability is any block built for a use case that can be reused in future projects without rebuilding.
Data ingestion, cleaning, and normalization. Built once, it serves all models.
Module that identifies people, products, and dates. Reusable in any case involving text.
Integration with business systems. Built for one use case, it serves the next ones.
Alert system for when model behavior changes. Applies to all models.
Tested prompt patterns for recurring tasks. Reusable with minimal variation.
Output quality testing framework. Reused across all AI modules.
🗓️ Visual roadmap with milestones and owners
The visual roadmap serves a practical purpose: align expectations. When everyone looks at the same diagram, disagreements surface at the planning meeting—not in the month the deadline arrives.
📋 Minimum roadmap structure
Crawl / Walk / Run clearly separated over time, with estimated durations and anchor dates.
Each block in a row or column, with a name, owner, and execution wave.
Go/no-go at the end of each wave, with visible success criteria.
Arrows connecting what depends on what. Criticisms are in red.
💡 Practical tip
The tool doesn’t matter: a table in Google Docs or a board in Miro work just as well. What matters is that it’s public, updated weekly, and everyone knows where to find it.
🔁 Replan based on what the pilot taught you
The pilot is the best consultant you’ll have. It reveals what planning couldn’t see: worse-than-expected data, more complex integration, or a use case that delivered 3× more value than expected. Using this data to replan is mandatory.
What the pilot often teaches
- →The data has more gaps than the mapping showed
- →Integrating with the legacy system will take 2× longer
- →An unmapped adjacent use case has greater value
- →The model works well, but the team resists adoption
Formal post-wave review
🎒 Module summary
Next module:
4.3 — Choose the pyramid level for each use case