Impair, Terminate, or Restructure? A Framework for Distressed AI Investments
Once the AI record has been reconstructed, six workout dispositions are available. The choice is dictated by what the evidence can actually support — not by the sponsor's preferred outcome.
The purpose of reconstructing an AI investment record is not to render a verdict on the initiative. It is to establish what workout dispositions are actually available. Six are recognized in the framework AI Capital Recovery uses, and the evidence — not the original sponsor's preference — dictates which apply.
The six dispositions
Retain
The system is producing documented value, the record supports that finding, and the contract, provenance, and ownership positions are defensible. No restructuring action is required. Retain decisions are the least common in distressed portfolios and require the strongest evidence, because they authorize continued spend.
Remediate
The system's operational performance is acceptable, but the evidentiary record around it is not. Remediation restores the record — contracts, provenance, ownership, cost allocation, outcome measurement — so that a defensible retain decision becomes possible. Remediate is often the correct first action, even where impairment is the eventual outcome.
Replace
The function the system performs is required, but the specific system cannot be justified — on cost, on vendor risk, on provenance, or on capability. Replacement is scoped against the reconstructed requirements rather than the original business case, which is usually no longer accurate.
Restructure
The commercial terms, the vendor relationship, the ownership boundaries, or the internal accountability structure are wrong, but the underlying system is worth preserving. Restructuring is the workout action most sensitive to the quality of the reconstructed contract and provenance record.
Impair
The carried value of the investment is no longer supportable, and the record supports a write-down. Impairment decisions require a reconstruction disciplined enough to withstand audit and, where relevant, disclosure review.
Terminate
The system does not warrant continued operation. Termination is a straightforward action only when the reconstruction has surfaced the dependencies, contract exit terms, data retention obligations, and downstream integrations. In practice these are frequently the surprises that turn a simple termination into an extended workout.
How the choice is made
The evidence recoverability classification produced during reconstruction determines which dispositions are actually available. Where the record is fully recoverable, all six are open. Where it is partial, some collapse. Where the record is unrecoverable, the range narrows further, and terminate or impair may be the only defensible options.
The framework is deliberately narrow. It does not include 'wait,' 'expand,' or 'invest further.' Those are pre-workout decisions. In a post-bubble cleanup, the only decisions on the table are the ones that convert already-committed capital into a defensible position on the balance sheet.
The disposition the evidence supports is not always the disposition the original sponsor prefers. Reconstruction is what makes that distinction visible.
Continue reading
- July 28, 2026 · 9 minThe Slopification Crash: How Rushed Deployment and Usage-Based Pricing Broke the Numerator and the Denominator
MIT says 95 percent of generative AI initiatives returned nothing. Deloitte says only 10 percent of organizations saw significant returns on agentic AI. The crash was not a model failure. It was a governance failure, compounded by consumption pricing, over a portfolio no one kept records on.
ROI & Capital LossGovernance & ProvenanceCase Studies - July 17, 2026 · 7 minWhen the Model Changes and No One Documents It: The Provenance Problem
AI systems in production are not the systems that were approved. Vendors deprecate models, swap defaults, and revise prompts on their own schedule. Without a provenance record, the asset cannot be valued or unwound.
Governance & Provenance - July 14, 2026 · 6 minShadow AI Spend: The Cloud, License, and Labor Costs No One Reconciled
Approved AI initiatives are only the visible fraction of enterprise AI expenditure. The unallocated portion — cloud, seat licenses, embedded features, and human rework — often exceeds it.
ROI & Capital LossGovernance & Provenance