Qlik has published research identifying what it calls “the hard part of AI” in enterprise deployments. The framing is self-serving — Qlik sells data integration and analytics software — but the underlying observation is accurate and underreported.

The hard part of AI is not the model. It is everything the model depends on.

Enterprise AI projects fail at a predictable set of choke points. Data that exists in the organization but is not accessible to the model. Data that is accessible but inconsistent, unlabeled, or structured in ways that the model cannot parse reliably. Outputs that are technically correct but operationally useless because they do not map to the decisions the organization actually makes. Governance requirements that make deployment legally or politically untenable. Integration costs that exceed the value of the capability being deployed.