Part of the AI, Future and War series. Analysis and hypothetical examples are identified in the text.
A demonstration is not a procurement case
An AI supplier may demonstrate fluent summaries and rapid analysis on carefully selected examples. The purchasing organisation still needs to know how the service behaves on its own documents, unfamiliar events and incomplete information. A compelling presentation does not establish that the system is suitable for decisions made under pressure.
Joint AI deployment guidance from the Australian Signals Directorate and international partners emphasises the security work involved in operating externally developed systems. Procurement should connect to that operating responsibility, rather than end when a contract and a successful demonstration are obtained. ASD and international partners: Deploying AI systems securely
Define the decision before the feature list
Start by describing what staff will do differently if the tool is introduced. Will it help find documents, draft a briefing or recommend an action? Which errors would be inconvenient, and which could create serious consequences? The acceptable design depends on these answers.
A hypothetical crisis-management team might permit AI to assemble relevant source material while requiring a qualified reviewer to produce the final assessment. That boundary is clearer than a vague promise that the tool will improve strategic intelligence. It also provides a basis for testing and training.
Demand evidence that resembles real use
A proposed acceptance exercise should include incomplete records, contradictory reports and questions the system should decline to answer. Retain the test inputs, outputs and reviewer decisions. A small set of attractive examples cannot show how often the system makes material errors.
Ask whether the evaluation applies to the version that will actually be delivered. A service can change after procurement through updates to its model, retrieval system or interface. The contract and operating process should explain notification, reassessment and options when a material change reduces suitability. These are negotiation considerations rather than claims about mandatory terms in any jurisdiction.
Examine the supply chain and exit route
The main vendor may depend on separate hosting, data or model providers. Request enough information to understand important dependencies and incident responsibilities. Where commercial confidentiality limits disclosure, record what cannot be verified and decide whether the uncertainty is acceptable for the intended use.
An exit plan should be tested before it is needed. Can the organisation export its records in a usable format? Can staff continue essential work without the product? Does ending the contract affect access to evidence required for an investigation? These questions concern continuity as well as commercial bargaining power.
Put responsibility into daily practice
Name an operational owner, a technical owner and a person authorised to suspend consequential uses. Staff need a simple way to report unreliable outputs. Procurement should also fund the time required for evaluation and training; an unfunded review obligation is unlikely to work under pressure.
For AI used in crisis decisions, the buying decision is ultimately about an operating arrangement. The model, supporting data, supplier behaviour and human review process must work together. Leaders should prefer evidence of that arrangement over broad claims of accuracy, autonomy or strategic advantage.