A 12-month roadmap: from isolated pilot to knowledge platform
Wave sequencing, exit criteria per phase, what to centralize and what to federate — and the mistakes that cost a year.
*Eleventh article in our series on enterprise AI, legacy integration and information governance.*
Executive summary
The predictable path is neither one giant platform project nor fifty parallel pilots: it is four waves, each with an explicit exit criterion, where every wave leaves a reusable asset for the next. Twelve months are enough to move from scattered usage to a governed knowledge platform — provided the company resists the temptation to skip wave one.
Wave 1 — Months 1 to 3: see and channel
Goal: stop knowledge loss and invisible usage.
- Inventory of usage and existing cases, including informal ones.
- An official path available to everyone, with corporate login and company-side logging.
- One-page policy and risk matrix published.
- A baseline measured in one chosen process.
Exit criterion: most usage flowing through the official path and a trustworthy baseline.
Wave 2 — Months 4 to 6: first case with real context
Goal: prove value in a process that crosses an actual legacy system.
- Integration with one corporate source, honoring source permissions.
- A cleaned-up document archive in the chosen domain, with a named owner.
- An evaluation set with known correct answers.
- A working audit trail for that case.
Exit criterion: primary metric met, answers citing valid sources, and per-profile leakage tests with no incidents.
Wave 3 — Months 7 to 9: turn it into an asset
Goal: make the second case cost less than the first.
- Curation running: templates promoted, poor cases flagged.
- Prompts and instructions versioned, with before-and-after evaluation for each change.
- Second and third use cases reusing integration, archive and governance.
Exit criterion: cost and lead time of the third case materially lower than the first.
Wave 4 — Months 10 to 12: scale with control
Goal: make adoption predictable and reversible.
- Federation: departments propose and operate low-risk cases within the central standard.
- Review of vendors, cost per unit of work and exit strategy.
- A timed audit rehearsal and a risk matrix review.
Exit criterion: new low-risk cases reach production without central committee review and without incidents.
What to centralize and what to federate
| Centralize | Federate |
|---|---|
| Identity, permissions and logging | Use-case selection per department |
| Legacy system integrations | Curation of the domain archive |
| Policy, risk matrix and audit | Process-specific templates |
| Contracts and vendor relationships | Local outcome measurement |
Centralize what is expensive to duplicate and dangerous to diverge on; federate what requires domain knowledge.
Five mistakes that cost a year
- 1.Starting with the platform before having a use case with an owner and a metric.
- 2.Indexing the entire network drive instead of one well-maintained domain.
- 3.Deferring permissions to phase two — fixing it later costs reindexing and trust.
- 4.Keeping ten pilots open with no exit criteria.
- 5.Treating curation as an IT task rather than the domain expert's.
What to do on Monday
- Put the four waves on the calendar with owners and review dates.
- Choose the wave 2 process now — preferably one that crosses the ERP.
- Write the exit criteria before approving any budget.
- Schedule a cost-per-unit-of-work review starting in month 4.
Conclusion
Durable AI adoption is not the fastest but the best sequenced: each wave delivers value and leaves an asset that makes the next one cheaper. The series closes with the verification questions in The executive checklist.
Further reading
Engineering track:
- Internal AI platform: from pilot to product — the technical deep dive on this topic.