Most AI investments do not fail because the model was wrong. They fail because the business case was loose. The pilot ran, the demo impressed, and the value never compounded. Impact is not produced by the model. It is produced by the case the model is wrapped in.

The Six Tests Of An IMPACTFUL Case

Across the AI work we have funded, sponsored, or rescued, the cases that survived contact with reality shared six properties. Each property is a hurdle. Together they form a useful filter for which cases to back and which to politely defer.

Identifiable. The decision the AI is helping with can be named, described, and located in the operating model. If you cannot say what decision is changing, there is no case to evaluate.

Measurable. The outcome that matters can be observed before and after, in the same units, on a timeframe the business actually runs on.

Provable. A reasonable counterfactual is available. You can argue, with evidence, that the outcome was caused by the AI rather than by everything else that changed at the same time.

Auditable. The trail from input to recommendation to action is recoverable on demand. A regulator, a board, or a customer can be answered without forensics.

Compoundable. The case extends naturally to adjacent decisions. Each deployment makes the next one cheaper.

Tolerable. The cost of the AI being wrong, occasionally, is bounded and survivable. Where it is not, the AI is advisory, not autonomous.

The strongest AI cases are unglamorous. They sit inside high-volume operational decisions where the cost of a slow answer compounds quietly, every hour of every day.

Where The High-Quality Cases Live

Three patterns produce the highest-yield AI cases we see. They are not the cases that make for the best demos. They are the ones that make for the best second year.

Triage. A high volume of inbound items has to be sorted into routes. Claims, tickets, exceptions, alerts, applications. The AI does not decide the outcome. It decides the queue.

Reconciliation. Two or more sources disagree about the same operational fact. The AI proposes the most likely truth, the human confirms, and the disagreement collapses faster than a person could sort it manually.

Recommendation under policy. A decision is bounded by policy and informed by data. The AI surfaces the most defensible action under the policy and presents the evidence. The human owns the commit.

Cases to be careful with

Generative content for external publication, fully autonomous decisions in regulated environments, and “AI-led strategy” exercises consistently fail the IMPACTFUL tests. They can still be valuable, but the case has to be honest about what is being measured and why.

Why Sovereign Matters For The Case

For Australian organisations operating under IRAP, Essential Eight, or sector-specific obligations, sovereignty is not a nice-to-have on the AI case. It is part of the case. A model that cannot be deployed inside the customer’s environment, with the customer’s data, under the customer’s controls, fails the auditable and tolerable tests by default.

DOLIUM is built for this constraint. The AI runs where the work runs. The provenance trail stays inside the boundary the organisation already governs. The case becomes defensible in the rooms where it has to be defended.

Funding The Right Twelve Months

The single most useful change we recommend is to stop funding pilots and start funding programmes. A programme commits to a portfolio of three to five high-quality cases over twelve months, with shared infrastructure, shared governance, and shared learning. The pilots that do not pass the IMPACTFUL filter never get budget. The ones that do compound.

To stress-test your AI portfolio against the IMPACTFUL filter, book a briefing.