85% of AI POCs never make it to production — a Go/No-Go checklist changes the game
The "valley of death" between the POC and production is the main graveyard of AI projects. Several studies suggest that a large majority of AI projects remain at the experimental stage. The causes are always the same: no formalized decision criteria, commitment bias (you keep going because you've already invested), and underestimation of the industrialization effort. A 12-criteria Go/No-Go checklist lets you make an objective decision in 2 hours, saving months of investment on projects doomed to fail.
The problem
The transition from POC to production is the most critical phase of an AI project — and the most poorly managed. SMBs and mid-market companies systematically underestimate the gap between a functional prototype and an industrialized solution.
The reasons POCs stay POCs:
- The demonstrator bias: the POC works perfectly on a carefully selected dataset, but collapses on real data (noise, edge cases, volumes). The performance gap between POC and production is typically 20 to 40%.
- The absence of objective criteria: no one defined upfront what constitutes success. The POC is judged "interesting" without anyone being able to say whether it deserves 50,000 euros of additional investment for industrialization.
- The under-costing of industrialization: going into production costs 3 to 5 times the cost of the POC. You have to add IT system integration, monitoring, error handling, security, documentation, training. This budget is often not planned.
- Commitment bias: after investing 2-3 months and 15,000-30,000 euros in a POC, it's psychologically difficult to say "stop." You keep going out of inertia, even when the results are insufficient.
The result is a portfolio of AI POCs "awaiting decision" that consume attention and resources without ever creating tangible value. Each zombie POC costs on average 15,000 to 25,000 euros per year in minimal maintenance and opportunity cost.
The AI solution
Our Go/No-Go checklist is based on 12 objective criteria split into 3 categories. A project must satisfy at least 10 of the 12 criteria to justify a Go, including the 5 "non-negotiable" criteria.
Performance criteria (4 criteria)
1. Model performance exceeding the baseline by at least 30% on the test set. 2. Acceptable error rate defined and met (e.g., < 5% false positives for a classifier). 3. Response time compatible with real usage (e.g., < 3 sec for a real-time assistant). 4. Robustness tested on edge data (edge cases, noisy data, peak volumes).
Business criteria (4 criteria)
5. Positive projected ROI at 6 months with documented assumptions. 6. Confirmed business sponsor with validated production budget. 7. End users involved and positive feedback (NPS > 3/5). 8. Stable use case (no major scope change since the POC launch).
Technical criteria (4 criteria)
9. IT system integration plan validated by IT. 10. Reproducible and automatable data pipeline. 11. Defined monitoring and alerting plan (drift, performance, costs). 12. Regulatory compliance verified (GDPR, AI Act if applicable). Discover our LLMOps offering for industrialization.
Implementation
Here are the three steps to organize an effective Go/No-Go review and make an informed decision in 2 hours.
Prepare the decision file (week -1)
The project manager compiles a structured file: POC results vs. success criteria defined upfront, scoring of the 12 Go/No-Go criteria (green/orange/red for each), estimate of the industrialization budget (development + integration + monitoring + training), provisional production rollout schedule, and risk analysis with mitigations. This file is 5 to 8 pages maximum and serves as the basis for discussion.
Conduct the Go/No-Go review (2 hours)
Bring together the business sponsor, the AI lead, the CIO and a representative of the end users. Review structure: presentation of the POC results (30 min), review of the 12 criteria with discussion on each orange or red point (45 min), evaluation of the production budget and schedule (15 min), formal decision and next steps (30 min). The decision is documented and communicated within 48 hours.
Execute the decision (week +1)
In case of a Go: immediately launch the industrialization phase with a detailed schedule and a ring-fenced budget. Appoint an industrialization lead (distinct from the POC lead if possible). In case of a No-Go: document the learnings in a closure report, evaluate the options (pivot, data improvement, archiving), and reallocate the budget to the next project in the roadmap. In case of a conditional Go: define the conditions to be met and the reassessment date (4 weeks maximum).
Results
Frequently asked questions
How long should an AI POC last?
An AI POC should last between 4 and 8 weeks, no more. Beyond that, you're in the eternal POC syndrome. The first 2 weeks are used to prepare the data and configure the environment, the next 2-4 weeks for development and testing, and the last 1-2 for evaluation and Go/No-Go. If the data requires more than 2 weeks of preparation, that's a warning sign.
What are the minimum criteria for a Go to production?
The 5 non-negotiable criteria are: model performance exceeding the baseline by at least 30%, sufficient and representative training and test data, positive projected ROI at 6 months, confirmed business sponsor and validated production budget, and a technical integration plan validated by IT. If even one of these 5 criteria is not met, it's a No-Go.
What should you do after a No-Go on an AI POC?
A No-Go is not a failure, it's an informed decision. Three options: pivot the use case (same technology, different problem), improve the data (if quality was the limiting factor, invest 4-6 weeks in data engineering then relaunch), or archive the project with a learning report that documents the lessons for future projects.
What is the difference between a POC and an MVP in AI?
The POC (Proof of Concept) validates technical feasibility: can AI solve this problem with this data? The MVP (Minimum Viable Product) validates business value: do users adopt the solution and obtain the expected gains? The POC is done in 4-8 weeks with a throwaway prototype, the MVP in 2-4 months with a solution that can be integrated into the IT system.
For technical profiles
Comparison of industrialization platforms for moving from POC to production:
| Criterion | n8n / Make + LLM API | LangServe / custom FastAPI | MLOps platform (MLflow + Kubernetes) |
|---|---|---|---|
| Deployment complexity | Low (no-code/low-code) | Medium (Python code) | High (full infra) |
| POC → Production time | 1-2 weeks | 2-4 weeks | 4-8 weeks |
| Built-in monitoring | Basic (logs) | Custom (to develop) | Complete (drift, perf, costs) |
| Scalability | Limited (1,000 req/h) | High (auto-scaling) | Maximal |
| Monthly production cost | 50 – 300 euros | 200 – 1,000 euros | 500 – 3,000 euros |
| Skills required | Advanced user | Python developer | ML Engineer + DevOps |
| Suitable for | Quick wins, < 500 req/day | MVP, 500-5,000 req/day | Critical production, 5,000+ req/day |
Our recommendation: for the first AI projects in an SMB, use n8n or Make coupled with an LLM API (OpenAI, Mistral). Going into production takes 1-2 weeks and the cost stays under control. Migrate to a LangServe/FastAPI architecture when volumes exceed 500 requests/day or when customization requirements increase. The full MLOps stack is only relevant from 5+ models in production with demanding SLAs.