/ all insights
Operations·Mar 17, 2026·11 min read
The first 90 days of an AI program, week by week.
What to ship, what to measure, what to refuse, and how to keep momentum past month two.
The first 90 days should create proof, not theatre. The mistake is treating AI as a company-wide transformation before one team has a working, measured workflow.
The 90-day shape
| Window | Focus | Output |
|---|---|---|
| Days 1-15 | Select one workflow with pain, data access, and willing users. | Workflow map, success criteria, failure examples. |
| Days 16-30 | Prototype the narrow loop. | Working draft with retrieval, prompts, schemas, or tool calls. |
| Days 31-45 | Test with real cases. | Error log, eval set, user corrections, first reliability fixes. |
| Days 46-60 | Add operational rails. | Review path, logs, fallbacks, permissions, handover notes. |
| Days 61-90 | Measure and decide what expands. | Usage, quality metrics, cost, next workflow recommendation. |
What to refuse early
- Do not start with an enterprise-wide agent platform.
- Do not automate a workflow nobody can explain manually.
- Do not skip human review for actions with customer, financial, or compliance impact.
- Do not measure success by demo excitement. Measure saved time, reduced misses, or faster decisions.