Engageably

05 — Journal

← Back to Journal

Translation failures cause more AI write-offs than model failures

May 21, 2026 · 7 min read · Method

Executive TL;DR

  • AI investments fail at the translation layer between model output and business decision. Not at the model.
  • For the CFO, write-down risk is now a process and governance problem, not a capability bet.
  • For the CAIO, the function accountable for translation is usually undefined. That gap is where impairment starts.
  • Companies with weaker models but stronger translation protocols beat the reverse, consistently.

The Shift

For most of the early adoption cycle, companies diagnosed AI failure as a model problem. Accuracy. Hallucination. Training data quality. That diagnosis is out of date.

The data points the other way. MIT’s NANDA initiative found that 95% of enterprise generative AI pilots deliver no measurable P&L impact. It attributes the failure not to model quality but to a learning gap in how organizations integrate AI into workflows (Fortune / MIT Project NANDA, 2025). S&P Global reported that 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the prior year (S&P Global Market Intelligence, 2025). McKinsey surveyed nearly 2,000 organizations. It found workflow redesign is the single attribute most correlated with EBIT impact from generative AI. Yet only 21% of organizations have redesigned any workflows (McKinsey & Company, 2025).

The failure sits one layer above the model. Call it the translation layer: the decisions, interpretations, and handoffs that turn a model output into an action the business can take and defend..

Why It Matters Now

AI budgets have matured from exploratory line items to capitalized assets. That accounting shift changes everything about how failure surfaces. A failed experiment used to disappear quietly inside an R&D expense line. A failed capitalized asset becomes an impairment, recognized on the P&L, with finance owning the explanation. Finance has to defend the carrying value, and that defense requires documentation most companies have never produced.

Companies are auditing model performance while the actual write-off risk lives one layer up, in the gap between what the model said and what the organization did with it.

What Most Companies Are Still Doing

The dominant pattern is simple. Staff heavily for model development and infrastructure. Leave translation to the most available technical person, or to no one with real authority. The outputs come out technically correct and organizationally inert. A recommendation lands in a dashboard. Nobody owns what happens next.

The symptoms compound. Decisions contradict model recommendations with no recorded rationale. Audit trails document model behavior in detail but say nothing about decision logic. When someone questions an outcome six months later, the company can show what the model said. It cannot show why a human overrode it, accepted it, or ignored it.

This is not just wasted spend. It is lost leverage. Without translation documentation, there is no basis to renegotiate a vendor contract, no defensible record to course-correct against, and no audit trail to support the carrying value of the asset. RAND found that more than 80% of AI projects fail, roughly twice the failure rate of non-AI IT projects. It located the cause in organizational and process factors, not model capability (RAND Corporation, 2024). The model worked. The organization did not.

What the Best Operators Are Doing Instead

The operators pulling ahead treat translation as a discrete, staffed, and auditable function. Not an assumed byproduct of good tooling. They have noticed what the benchmarks miss: the gap between a correct output and a correct decision is where the money is made or lost. MIT’s NANDA research found that vendor partnerships succeed about 67% of the time. Internal builds succeed at roughly one-third that rate (Fortune / MIT Project NANDA, 2025). The differentiator is integration discipline, not model quality.

1. Translation ownership. A named role or team is accountable for converting model output into decision-ready framing. Not a committee. A person with a title, a budget line, and a defined scope. When the output is wrong, or right but unused, this role answers for it.

2. Decision logic documentation. A lightweight protocol records not just what the model recommended but why the organization acted on it, modified it, or set it aside. The cost of the protocol is small. The cost of not having it shows up the first time the company has to defend an asset on the P&L.

3. Friction by design. Deliberate checkpoints sit between model output and operational commitment. The point is not to slow the system down. The point is to keep a recommendation from becoming a decision until someone with authority has translated it into one.

4. Write-off prevention as a CFO metric. Translation fidelity is a financial control, not an engineering concern. It carries budget authority and reporting visibility. The CFO can see, before an impairment review, whether the documentation exists to defend the asset.

Implications for the Next 12 Months

Capitalization and Impairment Pressure

More AI spend is moving from opex to capex treatment. The absence of translation documentation will speed up impairment recognition and shrink the CFO’s room to defend asset carrying value at year-end review. Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls (Gartner, 2025). Each cancellation that hits a capitalized line becomes a write-down finance has to explain.

The CAIO Accountability Gap

Without a clear translation mandate, the Chief AI Officer role is structurally unable to prevent deployment failures. The failure is not technical. A technical leader cannot fix it alone. The CAIO needs explicit authority over the translation function, or the role becomes accountable for outcomes it does not control.

Competitive Separation

Companies that institutionalize translation now will build a durable execution advantage over peers still optimizing model selection. McKinsey’s 2025 survey found that only 39% of organizations report any enterprise-level EBIT impact from AI. Just 6% qualify as high performers, attributing 5% or more of EBIT to AI use (McKinsey & Company, 2025). The separation between the 6% and everyone else will not show up in benchmark scores. It will show up in outcome consistency, quarter after quarter.

Executive Next Step

Run one diagnostic. Map the last two AI initiatives that underperformed and identify the layer where the failure occurred. If the model performed but the decision did not follow, the problem is translation, and the fix is organizational, not technical. The CFO and CAIO own this review together. The output should sit on both desks within 30 days.

Sources

← Back to Journal