Foundations for Success

A use case is not a value case

Organisations should be wary of AI programmes that begin with a catalogue of use cases. A long list creates movement, but it rarely creates accountability. The stronger starting point is one consequential decision whose economics can be seen. It might involve approving a transaction, intervening in a customer relationship or escalating an exception before it causes material harm.

A use case describes what the technology could do. A value case defines what must improve, by how much, for whom and over what period. If it does not identify the behaviour that converts an output into a result, the financial return is still hypothetical.

This is a widespread weakness. The Governance Institute of Australia found that 93 per cent of respondents could not effectively measure return from AI. The Reserve Bank of Australia reports shallow adoption and mixed returns across many firms. Access is spreading faster than accountability for value.

Defining the value case is the first discipline. Establishing how it will be proven is the second.

Measurement belongs at the beginning

An initiative without a baseline cannot establish a return. Before a model is selected, leaders should record the economics of the decision as it exists today.

Every initiative needs 4 measures. First, a business outcome such as revenue, cost or risk. Second, an operational signal such as decision time or rework. Third, a measure of unintended consequences such as reversals or overrides. Fourth, the cost of a successful outcome at production volume.

That final measure is easily misunderstood. The useful number is not the price of a model call. It includes data preparation, integration, human review, monitoring and remediation. A capability that looks inexpensive in a controlled trial can become uneconomic when volumes rise or every result requires verification.

Measurement is therefore not a reporting exercise performed after delivery. It is part of the design. It tells the team which data to capture, which controls matter and when to stop, adjust or increase investment.

A baseline tells the organisation what must improve. Understanding the work reveals what must change for that improvement to occur.

Build around the work as it really happens

AI never enters a blank process. It enters a role, an approval path, a set of permissions and real operating pressure. Experienced employees hold context that is rarely present in a system of record. They understand an informal dependency, an emerging risk or an exception that makes the obvious answer the wrong one.

That knowledge is not resistance to be managed away. It is design material.

The people closest to the decision should help define when the AI is useful, what they need to trust it and which circumstances demand human judgement. Their repeated overrides are valuable data. They can expose a weak recommendation, a poorly designed workflow or incentives that reward a different behaviour.

Adoption on its own is not success. A heavily used tool can still create rework, transfer risk or make a fast decision worse. The stronger test is whether the capability improves the quality and consistency of work without hiding its true cost.

Autonomy should be earned in stages

The move from AI that recommends, to AI that decides, to AI that acts is not a feature upgrade. Each stage is a different commercial and governance proposition.

When AI recommends, a person retains the decision and the organisation should measure relevance, acceptance and override reasons. When AI decides within defined rules, exceptions need a route to a person and measurement should include reversals and their cost. When AI acts across systems, identity, permissions, spending limits, audit records, rollback and an immediate stop mechanism become prerequisites.

No organisation should grant broader autonomy because the model appears more capable. Authority should expand only when evidence from the previous stage shows that the business can predict performance, detect failure and recover safely.

Organisations rarely get into difficulty because they started with too little autonomy. They get into difficulty because authority expanded before the evidence and controls were ready. Agentic AI makes this discipline more urgent. The distance between an answer and a business consequence is rapidly disappearing.

The architecture must carry accountability

Earned autonomy depends on evidence that survives production. The organisation must be able to establish which data informed a decision, which permissions governed an action, what the system did and how the outcome was measured.

Codex treats the enterprise stack as that chain of accountability. A governed AWS Landing Zone establishes the boundaries. SageMaker Lakehouse with DataGovOps, Enterprise Ontology and AWS Context establish trusted data and shared meaning. Amazon Bedrock, SageMaker AI, Amazon Bedrock AgentCore and Amazon Quick apply intelligence within those boundaries.

Kiro’s specification-driven artefacts preserve the connection between executive intent and technical implementation. Requirements define the intended outcome, acceptance criteria define the evidence and review gates protect the path to release.

The point is not to own more technology. It is to make each recommendation traceable, each action bounded and each outcome measurable.

The questions leaders should insist on

At board level, the argument reduces to evidence, economics and control. Management should show the counterfactual. What would have happened without the AI? Can the result be separated from market conditions or process changes? What prevents a locally sensible action from creating a larger problem elsewhere? What happens to the economics at twice the volume? Who can stop the system, and how quickly?

These questions are not barriers to innovation. They are what make sustained investment possible.

Enterprise AI will not become valuable simply because it becomes more persuasive. It will become valuable when the organisation can prove, repeatedly, that a better decision occurred at an acceptable cost and level of risk. The organisations that lead will be those whose foundations allow them to grant AI more authority without surrendering control.

For every proposed capability, the most useful executive question is now this: what must be true before an organisation allows AI to do that here?

Read more of Adrian’s ‘Return on AI’ series here.