We run a paid assessment before any machine learning engagement. A meaningful share of those assessments end with a recommendation not to build the model. Clients occasionally find this odd for a firm that would rather sell the build. The reasoning is straightforward: a model that ships on bad foundations does not fail loudly, it fails quietly, and quiet failure is far more expensive than an early no.
The data does not exist in the form you think
Nearly every organisation believes it has more usable data than it does. Transactional records exist, but the label you need to predict often does not. A client wanting to predict customer churn had eight years of transaction history and no definition of churn: no cancellation event, no clear inactivity threshold, no field marking a lost account. Before any model could be trained, the business had to decide what churn meant, and that turned out to be a genuinely contested question internally.
The data is censored by your own operations
This is the failure mode we see most often and the one that does the most damage. Sales data does not record demand, it records demand that was successfully met. If an item was out of stock, sales show zero, and a demand model trained on that data learns to expect zero. It will then recommend not stocking the items that sell out fastest, which is precisely backwards.
Correcting for censored demand is possible if you have stock-out timestamps. If you do not, the honest answer is that the data cannot support the model until you start recording them.
There is no baseline to beat
Any model must be compared to what is happening now. Often nobody has measured what is happening now. If your planners currently forecast with a mean absolute percentage error of twelve percent, a model at fifteen percent is a costly step backwards, and without the baseline you would report the fifteen percent as a success.
Measuring the baseline sometimes ends the project on its own, because it reveals that the current process is already close to the achievable ceiling.
Nobody will act on the output
The most avoidable failure is the accurate model that changes no decision. If the prediction arrives after the decision point, or reaches someone with no authority to act, or contradicts an incentive nobody is willing to change, then accuracy is irrelevant. We ask early who receives the output, when, and what they are empowered to do differently. A vague answer is a serious warning sign.
What a good assessment produces
An honest data quality report, a measured baseline, a written success criterion, a small prototype, and a clear recommendation. When the answer is yes, everyone proceeds with realistic expectations. When it is no, the client has spent a small fraction of a build budget to avoid spending the rest.