From pilot to scale: the 5 stages of AI rollout
The pilot worked great. Scaling it is a disaster. Why this happens and how to avoid it.
According to McKinsey State of AI 2025: 88% of companies use AI, but only 7% have it fully deployed across the organization. Nearly two thirds are still stuck in the experiment-and-pilot phase. The technology isn't the bottleneck. The hard part is getting from pilot to scale. Here's a roadmap that works.
Stage 1: Discovery (Week 1-2)
Before you write a single line of code, answer these questions:
- • What specific problem are we solving?
- • Who is the business owner?
- • How will we measure success?
- • Do we have data? At what quality?
Output: a 1-pager defining the project and its success criteria.
Stage 2: Proof of Concept (Week 3-6)
A PoC is a test of technical feasibility. It doesn't have to be pretty, it just has to answer the question: "will this even work?"
WHAT YOU DON'T DO IN A POC
- ✗ UI: Excel or a notebook is enough
- ✗ Scalable infrastructure: a laptop is enough
- ✗ Integrations: mocked data
- ✗ Security review: that comes later
Output: a working prototype + an assessment of whether it's worth pushing on.
Stage 3: Pilot (Month 2-3)
A pilot is a test with real users, but at limited scale. 5-20 people, one department, one region.
What you're testing:
- • Do users actually use it?
- • Does the solution deliver value?
- • What needs to change before scaling?
- • What does the feedback loop look like?
"A pilot that just 'works' is only half the job. The real result is a list of things to fix before you scale."
Output: a go/no-go decision + a list of production requirements.
This is what I do hands-on — advising on AI strategy and building agents that survive the demo.
Stage 4: Production Ready (Month 4-6)
This is where most projects die. Because the move from pilot to production requires:
PRODUCTION READINESS CHECKLIST
- ✓ Infrastructure: Scalable, with redundancy
- ✓ Monitoring: Alerting, logging, dashboards
- ✓ Security: Review, penetration testing
- ✓ Integrations: Real APIs, not mocks
- ✓ Documentation: For users and support
- ✓ Training: Onboarding for users
- ✓ Support: Who answers when something breaks?
- ✓ Rollback: How do we revert to the old process?
Output: a production-ready system + a rollout plan.
Stage 5: Scale (Month 6+)
Scaling isn't "turn it on for everyone." It's a controlled rollout:
- Wave 1: 10% of users, you watch the metrics
- Wave 2: 25% of users, you gather feedback
- Wave 3: 50% of users, you optimize
- Wave 4: 100%, full rollout
Between each wave: review, fixes, a decision on whether to continue.
Why do projects die between stages?
COMMON CAUSES OF DEATH
- PoC → Pilot: No business owner, "interesting but not now"
- Pilot → Production: Underestimated production cost, no budget
- Production → Scale: Legacy systems, organizational resistance, no change management
How to improve your odds of success
- 1. Have a C-level sponsor from day one
- 2. Budget the whole path, not just the PoC
- 3. Start with a simple problem with clear ROI
- 4. Involve end users from the pilot onward
- 5. Plan change management in parallel with the technology
How to land in the 7%
Pick one process with clear ROI and take it through all five stages, instead of firing off ten pilots at once. Close every stage with a go/no-go call, and only move on with budget for the whole path. That's the difference between the 7% of companies running AI in production and everyone else still testing.
