In the AI Expert Circle community, I’d like to hear from you:
Which of these five questions is most frequently overlooked in real-world AI projects?
My observation across AI Prototypes, Pilots, MVP/MLP or POC is that many teams can answer the technical questions, but struggle with ownership, testing/evaluation, and operational readiness when a prototype needs to move beyond experimentation. This becomes increasingly important as organizations seek to operationalize Agentic AI and create repeatable enterprise capabilities.
I’d be interested in hearing from you:
Which question do you think organizations underestimate the most, and why?
Florian, good article. The governance vs operating model distinction is the key part because that’s where I see teams get stuck. Everyone has a governance deck. Almost nobody can tell you how work actually moves when the AI gets something wrong.
So my answer is : failure management, and I’d argue ownership falls with it. Same root cause.
To explain in detail, it’s not that teams don’t care. It’s that in a prototype the failure never shows up. The agent answers something plausible, the demo lands, next slide. 2% error in a POC is nothing. 2% on a process running 10,000 times a week against a real book of business is an audit finding with your name on it.
Lately with banking and insurance customers I’ve been asking one thing before we even talk about accuracy: show me the failure path. What happens when the agent is wrong or unsure. Who gets it, what’s the SLA, is there a case behind it with an audit trail or does the error just die in a chat log. I’ve had rooms go quiet at that question more than once. And that silence usually answers the ownership question too, because you can’t own a decision if nobody has defined what a bad one looks like.
And finally your line about every prototype teaching the org how it will operate AI at scale is the one I’d underline.
So true @Fernando_Diaz Whenever I asked people the question “What must not go wrong” (in order to force thinking and a discussion about what makes the difference and what to prevent from happening), and people replied “but I don’t want things to go wrong” or there was this silence, this reveals so much. Being able to draw a line between wanted and unwanted outcomes, being able to explain the differentiators and what happens in both cases is essential for requirements and good solutions.