Value

The measurement that does not happen

Boards see AI benefit cases. They see them in detail, with projections, and they approve against them.

The gap appears when the same measure is not taken afterward. A proposal may state that handling time will fall by a specified amount against a stated baseline. If handling time is never measured and the baseline cannot be retrieved, the board cannot establish whether the benefit was realized. The team's dependence on the system can then become the reason to keep it running.

This is why Value sits third rather than last. It is not a financial control appended to the end of a governance framework. It is the domain that determines whether the organization can distinguish an AI system that works from one that is merely in use, and that distinction becomes unavailable within months of deployment unless it was arranged for beforehand.

The practical remedy is small. VAL-1 requires the measure and the baseline to be recorded at approval. VAL-2 requires the measurement date to be set at the same time. Both are easier to establish at approval than to reconstruct afterward.

Efficiency is not effectiveness

VAL-6 separates two claims that are routinely presented as one.

Efficiency concerns the cost of producing an output: the same work performed faster, more cheaply, or in greater volume. Effectiveness concerns the quality of the outcome: better decisions, fewer errors, improved results for the people affected.

AI systems frequently deliver the first while the case was argued on the second, and the conflation is rarely deliberate. A system that drafts responses faster has improved efficiency. Whether the responses are better, or merely more numerous, requires separate evidence that is harder to obtain and is therefore usually not obtained.

The distinction matters to a board because the two carry different risks. An efficiency gain that turns out to be smaller than claimed is a financial disappointment. An effectiveness claim that turns out to be wrong means the organization has been making worse decisions at scale, which belongs to Exposure and may belong to Obligations.

The cost side moves

VAL-3 exists because AI cost behaves differently from the cost of the systems boards are used to approving.

Build and license costs may be known at approval, but usage-based AI services add costs that change with adoption. Where model inference is charged per unit of work, a successful deployment can increase operating costs beyond what the original business case anticipated.

Two further components are commonly omitted entirely. Human review effort, where the system's outputs are checked, is a recurring operational cost created by the deployment. Assurance is a recurring cost under ASR-1. Both are frequently absorbed into existing team capacity and therefore never appear against the system, which makes the system look cheaper than it is and makes the benefit case look better than it is.

Decommissioning

VAL-4 is the control that distinguishes measurement from reporting.

Organizations rarely switch off AI systems that do not deliver. The reasons are predictable: the investment is sunk, the team has adapted around the system, the original sponsor is committed to it, and the alternative process has been reduced. Each is a reason to continue that is unrelated to whether the system works.

Stating the decommissioning trigger in advance, as VAL-4 requires, moves the decision to a point where it can be taken on the evidence. Deciding afterward means deciding under advocacy, and the outcome of that decision is reliably to continue.