Carry phase

IAS 36 impairment triggers for AI models

Carry, live question

IAS 36 requires an entity to assess at each reporting date whether any indicator suggests an asset may be impaired, and to test if one exists IAS 36 §12. For a token-built model the indicators that bite are obsolescence-driven: deprecation of the underlying model, a superior superseding model, or a collapse in inference price that undercuts the asset's economics.

The indicators IAS 36 lists

IAS 36 sets a minimum list of external and internal indicators to consider at each reporting date IAS 36 §12. External indicators include significant adverse changes in the technological or market environment and declines in value. Internal indicators include evidence of obsolescence or physical damage and evidence that economic performance is, or will be, worse than expected.

Which apply to a token-built asset

External: a superseding foundation model, or a sharp fall in inference prices that changes the build-versus-buy calculus.
External: an announced deprecation date for the underlying model or endpoint the asset depends on.
Internal: evidence of obsolescence, such as adoption falling as users move to a newer capability.
Internal: worse-than-expected economic performance from the asset relative to the case that justified capitalising it.

When a test is required

The indicator review is done at each reporting date; a formal recoverable- amount test is required only when an indicator exists IAS 36 §12. For AI assets the review is not a formality, because indicators arise between reporting dates whenever a provider announces a deprecation or a materially better model ships. A mid-period announcement is itself the trigger to test.

Evidencing the assessment

Whether or not an indicator is found, the assessment should be documented: the indicators considered, the evidence reviewed, and the conclusion. Where an indicator is found and a test performed, the recoverable-amount workings and the resulting loss are retained. This file is what an auditor tests, and its absence is a common weakness in fast-moving AI portfolios.