Since 2023, AI has been in every enterprise meeting. Vendors promise transformation, consultancies draw roadmaps, boards allocate budget. Twelve months later, the number of working systems falls well short of expectations.
This post is not a failure autopsy. It shares three differences we have observed in the field — between systems running in production and those stuck at the pilot stage.
1. Is there a concrete business problem, or just an "AI project"?
The vast majority of failed projects start with an "AI strategy." Working projects start with a far more mundane question: "How much time are we losing in this process?" or "Why does this error keep recurring?"
Hotel example: rather than "making AI do guest questions," identify that 40% of guests filling in self-check-in forms ask the same three questions, and automate only those three. The second approach is measurable; the first stays in the deck.
Build around the business problem, not the AI project. AI is the tool, not the goal.
2. Layer on top, or rebuild from scratch?
When it comes to factory automation, ERP integration or hotel PMS connectivity, we frequently see proposals to "remove the existing infrastructure and rebuild from scratch." That approach typically means an 18–24 month project, a large budget and high cancellation risk.
Working systems are built layer by layer. Connect to the existing PLC over standard protocols, put a REST adapter in front of the ERP, read PMS data one-way. Addition, not replacement. With this approach a 30-day PoC is genuinely possible — because you are not touching production infrastructure.
3. Who will use it, and how?
Even the most sophisticated model is worthless if users reject it. This needs to be solved as a business problem, not a technical one.
If a factory technician checking an error code on the morning shift does not have a phone to hand, they will not open a web interface. If they query over WhatsApp, put the interface there. A hotel receptionist cannot learn 10 different systems on onboarding day — the system you build must integrate into existing screens, not route them to a new one.
Skipping the usage context is the primary reason for the "technically deployed but nobody uses it" outcome.
Working in production means the system people actually use, not just the one that deployed successfully.
Summary
Why enterprise AI projects fail is rarely technical. No concrete problem, disconnected from existing infrastructure, user context ignored. When all three are right, the system is still running 12 months later.
When we deploy 1click.chat as an enterprise assistant, the process always starts with those three questions. That is also why the pilot model works: a concrete business problem, limited scope, real users, 30 days.