Insights

Many AI projects get evaluated the wrong way

The first questions about an AI project are often about the model, the demo, or how fast a team can ship a prototype.

The real question is different: can this intelligence be trusted, embedded in real work, absorbed by the organization, scaled safely, and, if it faces customers, defended? That is how I estimate the odds that an AI project will actually succeed.

A. Can it be trusted in the task?

This is about reliability, recoverability and the burden of trust. The right bar depends on the use case. In some, 80% accuracy is enough to save time and still create value. In others, you may need 99.9% because the cost of being wrong is too high.

The question is not whether the system is perfect. It is whether it is reliable enough for the task, and whether it fails in a way that can be caught, corrected or escalated.

B. Does it fit the work?

This is about workflow, context and governance. Is the AI causing the workflow to be rethought in a better way, or is it just patched onto an existing process without removing friction? If it is not reducing handoffs, compressing cycle time, improving decisions or simplifying execution, it may not be helping much.

C. Can the organization absorb it and prove value?

This is about adoption and outcomes. A practical question here is: what is in it for the end user? If the user experience is worse, the extra effort is unclear, or the benefit goes mainly to leadership while the frontline carries more burden, adoption will stall.

Even a strong system can fail if users do not see clear personal value and the business does not define success in measurable terms.

D. Can it scale safely?

Security and resilience are not always the first gating issue, but they quickly become critical as projects expand. Can the system handle sensitive data, misuse and operational failure without becoming a liability?

E. If it is an external product, can it be defended?

For internal AI, the main question is usually value capture through workflow and adoption. For a customer-facing product there is a second question: if this works, what prevents it from becoming a commodity?

Many thin AI wrappers will disappear as the major AI labs keep adding capabilities that used to look defensible, from dashboarding to interface layers to basic orchestration. So the bar is higher: where is the real advantage?

Where projects go wrong

My bias is that most AI projects fail less because the model is weak and more because they were scoped wrong early, especially on trust, workflow fit and adoption.

Originally posted on LinkedIn on April 21, 2026, lightly edited. See the original post and discussion.

Start with a 30-minute conversation

Tell me where AI stands in your company today. I will tell you plainly where I think it can help, and whether I am the right person to help with it.

Email me to set up a 30-minute call