AI business cases need auditable assumptions

AI can now prepare an infrastructure business case in minutes. On August 24, Google Cloud announced AI-powered Quick Assessments in Migration Center: infrastructure inventories or aggregated inputs can produce TCO estimates, target architectures, bills of materials and exportable reports. This shortens discovery. It does not replace a decision on the assumptions behind it.

A model is not a decision

A TCO model looks precise even when its inputs are not. What utilisation was assumed? Which licence and staffing costs are included? Which region, availability requirement or migration sequence was used? AI can explain these parameters and calculate scenarios. It must not turn an open assumption into a number that merely appears certain.

Every AI-assisted business case therefore needs an auditable appendix:

• Data snapshot: Which inventories, invoices and time periods were used?
• Assumptions: Which cost, usage and growth values were set?
• Scenarios: What changes with a different region, utilisation level or term?
• Approval: Which business, finance and IT role owns the final number?

Google describes adjustable financial parameters for Quick Assessments and an agent that explains the underlying logic. That explainability should become an acceptance criterion: not only the recommendation, but also its provenance must be reviewable.

AI tools belong in management control

The same logic applies to developer agents. For Antigravity, Google consolidates usage metrics for tokens, API calls and developer activity, as well as central audit logs. Microsoft Defender now also reports posture risk for AI agents, incorporating configuration, permissions, runtime activity and alerts.

This creates two management loops that belong together: economic evidence for an initiative and operational visibility into what an agent actually does afterwards. An approved business case remains incomplete when usage, permissions and deviations in production are not visible.

What DACH organisations should do now

Start with a bounded modernisation initiative. Let AI prepare the first assessment, but version input data and assumptions separately from the result document. Before presenting to a steering committee, define at least one sensitivity scenario and one accountable owner for every material number.

The relevant metric is not how quickly an assistant creates a slide. It is whether, months later, the organisation can still explain why that number was approved — and whether the agent funded by it operates within its approved boundaries.