Why most AI projects never reach production
Industry estimates suggest that between 60% and 80% of AI projects in mid-size companies do not survive the pilot phase. Not because the technology fails. Because the project was poorly designed from the start: no clear success criteria, no adoption plan, no governance, and no internal team capable of operating what was built.
The problem is not the AI. It is how the engagement is structured.
A COO evaluating an AI provider today does not need to understand language models. They need to know which questions to ask before signing. The following eight are the ones that matter most.
1. What are the success criteria for the project, and who defines them?
If the provider responds with vague metrics — "improve efficiency", "reduce turnaround times" — that is a warning sign. A well-designed project has concrete success criteria from week one: hours saved per process, error rate before and after, cycle time for a specific workflow.
If they cannot define it before the work begins, they will not be able to measure it when it ends.
2. How many weeks until something is running in production?
Not in a demo. Not in a sandbox. In production, with real data, operated by the team.
A serious provider can commit to a first functional deliverable within six to ten weeks. If the answer is "it depends on many factors" with no concrete range, the project will likely run longer than planned — and the opportunity cost is borne by the client.
3. Who operates the system once the provider leaves?
This is the question most often omitted in commercial conversations, and the one that causes the most damage afterward.
If the answer is "we handle everything", the client is buying dependency, not capability. A provider genuinely oriented toward results transfers knowledge to the internal team throughout the implementation. By the time the project closes, the client's team must be able to operate, modify, and expand the system without external intervention.
4. How are AI model costs managed in production?
Language models carry variable costs that can scale unexpectedly without governance in place. A provider that does not address observability, spending limits, and usage policies from the outset is ignoring a real operational risk.
For a mid-size company processing moderate volumes, monthly model costs can range between 200 and 2,000 euros depending on system design. Without controls, that range widens in both directions.
5. What happens when the AI model produces an incorrect output?
Every AI system produces errors. The question is not whether it will happen, but how the system is designed to detect, contain, and correct it.
A provider that has no clear answer on validation mechanisms, human review at critical checkpoints, and output traceability is not designing for production. They are designing for a demo.
6. How does it integrate with the systems you already use?
ERP, CRM, reporting tools, internal databases. An AI agent that does not connect with existing systems creates additional work rather than eliminating it.
Integration is not a secondary technical detail. It is part of the project design from the start. If the provider treats it as a later step, the real scope of the work is being underestimated.
7. Do they have experience with companies of a similar size and sector?
Solutions designed for global corporations do not scale down directly. A company of 80 people has team, budget, and risk-tolerance constraints that are fundamentally different from those of a multinational.
Ask for concrete cases. Not generic testimonials: real cases with context, problem, solution, and measurable outcome. If they don't have them, the client absorbs the cost of the provider's learning curve.
8. What is the governance model once the system is in production?
An agent in production is not static software. Models are updated, business workflows change, volumes fluctuate. Without an active governance model — someone monitoring output quality, costs, incidents, and system evolution — the project degrades over time.
Ask the provider what is included after go-live. If the answer is "incident-based support", the real maintenance work will go uncovered.
A concrete example: what happens when these questions are not asked
A distribution company with operations in three countries hired a consultancy to automate its monthly financial reconciliation process. The project ran for nine months, cost approximately 120,000 euros, and produced a system the finance team did not know how to operate. Six months after project close, the process was being done manually again.
The problem was not the technology. It was that no one defined success criteria, no one planned the knowledge transfer, and no one established post-implementation governance.
An equivalent project, well designed — with clear criteria, an enabled team, and active governance — can be executed in eight weeks at a cost between 15,000 and 40,000 euros, and deliver savings of 20 to 40 hours per month in the finance team from the first month of operation.
The difference is not in the technology. It is in how the project is designed and contracted.
What to do if the provider cannot answer these questions
Three options:
Request written answers before moving forward. A serious provider has no difficulty documenting their commitments. If there is resistance, that is relevant information.
Ask for direct references from current clients. Not edited success stories: real conversations with the client team that operated the project.
Run an internal diagnostic first. Before evaluating providers, it is worth gaining clarity on which processes carry the greatest potential impact, what data is available, and what internal capacity exists to adopt the system. Without that foundation, any external proposal is difficult to evaluate with sound judgment.
Conclusion
Contracting an AI project well does not require deep technical knowledge. It requires the right questions and the discipline not to move forward when the answers are not concrete.
The eight points above cover the vectors where most projects fail: success criteria, time-to-value, adoption, governance, integration, and knowledge transfer. A provider that answers them clearly reduces risk significantly. One that avoids them increases it.
If you want to assess where your company stands before beginning a conversation with any provider, complete the free diagnostic form. You will receive an initial assessment within 48 hours, with no obligation.
