Few technologies have been discussed as intensely and deployed as narrowly as AI has been in the past few years. The gap is not caused by the technology underperforming. It is caused by starting from capability rather than from problem — which produces pilots that demonstrate something impressive and change nothing about how the business runs.
The alternative is unglamorous and considerably more productive: identify decisions that are currently made badly, late, or not at all; determine whether the data to make them better actually exists; then choose the simplest technique that answers the question reliably.
Where it earns its place
A handful of applications recur because they share a structure — a repetitive judgment, made often, on data the business already holds.
Segmentation. Most businesses segment customers by something convenient: size, industry, region. Behavioral segmentation — grouping customers by how they actually buy and engage — routinely cuts differently, and the differences are actionable because they are about behavior rather than description.
Churn indicators. Not a prediction delivered as a score with no explanation, but an early, explainable signal that a specific relationship is weakening, delivered while intervention is still possible and accompanied by enough context for the person receiving it to act.
Customer service automation. The value is rarely in replacing the team. It is in resolving the substantial share of contacts that are genuinely routine, and in giving the person handling the rest immediate access to the full history rather than making the customer repeat it.
Document and knowledge work. Extracting structured information from unstructured documents, and making an organization's accumulated knowledge searchable by question rather than by filename. Both are unglamorous, both compound, and both remove work that no one should be doing manually.
Decision support. Assembling the analysis that a person would assemble if they had the time, so the decision happens within the window in which it matters rather than the week after.
Why the pilots stall
The pattern is recognizable. A capable model produces a credible output; the output arrives somewhere nobody works; no workflow changes; the pilot is judged successful and quietly ends.
This is a design failure rather than a model failure. An AI capability creates value only where it changes what someone does — which means it must be delivered inside the system where the work already happens, at the moment the decision is made, with enough explanation to be trusted and enough accountability that a person remains responsible for the outcome.
The second common stall is data. A model is a function of what it can see, and in most organizations what it can see is partial, inconsistently defined and spread across systems that disagree. Applied to that foundation, AI produces confident answers of uncertain value. This is why connected data is a precondition rather than a parallel workstream.
Choosing the simplest technique that works
A meaningful share of problems presented as AI problems are better solved by a clear definition and a rule. If 'at risk' can be defined as a specific observable condition, a rule is more explainable, cheaper to run, easier to adjust and considerably easier for the business to trust. Reserve the heavier techniques for problems where the pattern genuinely cannot be specified in advance.
This is not conservatism. It is the same judgment applied to any technology choice: the system must earn its ongoing cost, and explainability has real operational value when a person has to act on the output and answer for the result.
Keeping a person accountable
For customer-facing applications in particular, the durable arrangement is that the system informs a decision someone remains answerable for. That constrains where automation is appropriate — routine and reversible, yes; consequential and irreversible, with a person in the loop — and it tends to produce systems the organization actually adopts, because the people using them understand what the system is doing and retain the authority to override it.
The honest starting point
AI is one capability within a broader technology strategy, and it is most valuable where the groundwork already exists: connected customer data, defined lifecycle stages, instrumented workflows. Where that foundation is present, applied intelligence produces real advantage quickly. Where it is absent, the foundation is the higher-return project — and building it first is what makes the AI work when it arrives.