The first investment question is not how many AI use cases the enterprise can launch. It is which decision deserves better intelligence—and what would make that intelligence dependable enough to use.
Define the decision before choosing the capability
“Improve productivity” is a direction, not yet an investment case. Identify the work: investigating an invoice exception, assessing an asset intervention, preparing a service response or comparing supply alternatives. Then identify the person accountable for the decision and what a better outcome means.
A baseline should include the current effort, elapsed time, error or rework, and business consequence. Faster preparation is useful only if it changes the work that follows. Time released, cash saved and revenue enabled are different forms of value; the business case should keep them distinct.
Make the relevant context dependable
An assistant may produce a fluent answer while drawing on an outdated policy, the wrong asset record or information the user should not see. An agent can execute a well-formed action against an incomplete understanding of the customer commitment.
Readiness therefore begins with the sources, their owners, permissions and meaning. Reconcile the records required for the chosen decision. Define which system governs the fact and how changes reach the workflow. The objective is sufficient dependable context for the bounded use case; it need not become an indefinite programme to perfect every enterprise dataset.
Set the boundary of authority
Assistance and authority are separate design decisions. A capability may retrieve evidence, compare options, prepare a recommendation or execute an approved action. Each step changes the consequences of failure and the controls required.
Decide which actions require approval, when the system should stop, and who handles an exception. Give the responsible person enough context to challenge the recommendation. “Human in the loop” has little meaning if the reviewer lacks time, evidence or the ability to intervene.
Evaluate the operating system around the model
A useful evaluation represents real work, including missing information, conflicting records and unusual requests. Test whether outputs are supported by the permitted sources, whether proposed actions respect the boundary and whether exceptions reach the right owner.
The economic view must include integration, evaluation, monitoring, model usage, human review and support. A successful prototype can still be an unattractive operating proposition. Establish acceptance conditions before testing, then retain the option to reshape or stop.
Scale a capability the business can own
Expansion should follow evidence from use: service performance, adoption, exception patterns and realised benefit against the baseline. Name who maintains the knowledge, reviews evaluations, manages changes and owns the operating budget.
Our view is that enterprise AI earns its place through this discipline. The dependable core and useful intelligence develop together, at the pace the business can absorb and govern.