There's a specific pattern behind most stalled AI initiatives: a vendor demo looks impressive, budget gets approved, and six months later the project quietly stops being mentioned in leadership meetings. The model usually isn't the problem. The company skipped questions that had nothing to do with AI and everything to do with whether the organization was ready to use it.
First: is there actually a knowledge base worth training against? A conversational agent is only as good as what it can retrieve, and 'we have documentation somewhere in Confluence' is not the same as structured, current, accessible content. If your human agents already struggle to find the right answer, an AI agent pointed at the same mess won't do better. It'll just be confidently wrong faster.
Second: what's the actual approval boundary? Every serious AI deployment inside a workflow needs a clear line between what the agent can do autonomously and what needs a human to sign off. Teams that skip defining this either over-restrict the agent until it's useless, or under-restrict it until it does something nobody wanted, out loud, to a customer.
Third: who owns the model's mistakes? Not legally. Operationally. When the agent gets something wrong, is there a person whose job includes noticing that, fixing the underlying data, and retraining? If the honest answer is 'nobody, we'll figure it out,' the project doesn't have an owner, it has a launch date.
Fourth: what does 'success' actually mean, numerically, before you start? Deflection rate, resolution time, customer satisfaction. Pick one or two, agree on the target before launch, and measure from day one. Projects that define success after the fact tend to define it generously.
Fifth: is this replacing a manual handoff, or just adding a new fragile process on top of the old one? The AI projects that actually stick are the ones built around removing a specific, named bottleneck. If you can't name the exact handoff being removed, you're not ready to buy yet. You're ready to go find that handoff first.