Everyone Knows Something Is Wrong With Enterprise AI. Here Is Why Nobody Says It
Orignally published January 27 2026 on LinkedIn
Note
Enterprise AI failure is not invisible. It is legible across boards, audit trails, capital allocation and legal exposure. The reason it is rarely stated plainly is not lack of evidence. It is the economics of silence.
Boards are being asked to approve AI investments they cannot defend. Audit committees are signing off systems they cannot independently verify. Legal teams are absorbing liability they cannot trace.
This is not a future risk. It is current practice.
The Depreciation Trap
Hyperscalers are building infrastructure on 18-24 month obsolescence cycles. Enterprise governance resolves trust on 36-60 month cycles. The capital is real. The depreciation is real. The confidence is not.
Nvidia’s release cadence proves the point: A100 to H100 to Blackwell in 36 months. Economic obsolescence arrives before utilisation peaks. Data centres look permanent. Accelerators age like fruit.
Recent financial commentary is now questioning whether the useful life assumptions behind AI infrastructure are credible at all. Assets written down over four or five years may be losing economic value in two. That is not a technology problem. It is an accounting problem.
Hyperscalers are responding by stretching depreciation schedules rather than confronting the economics. Alphabet’s shift from four to six years reduced reported depreciation by $3.9bn in a single year. Microsoft disclosed a $3.7bn uplift to operating income from the same adjustment. Amazon’s extension from five to six years added an estimated $3.1bn to 2024 operating income. Meta’s adjustments account for a further $1.5-$2bn. These are disclosed accounting effects, not speculative estimates. Across just these four firms, over $12bn of economic cost was deferred through policy change rather than technological improvement. That is before including second-tier hyperscalers or the Blackwell-era CAPEX wave.
This is beginning to resemble arbitrage dressed up as innovation: hyperscalers sell shovels at scale while the gold remains unrealised.
By the time boards reach epistemic confidence, the silicon is commercially obsolete. This is not railways before industrialisation. It is empty airports built before airlines exist.
Vendor narratives continue to frame this as the largest infrastructure build-out in history. That rhetoric sustains confidence in capital markets. It does not resolve the underlying utilisation gap inside enterprises.
The Economics of Silence
Most people inside the system already see the gap. Consultants see stalled pilots. CIOs see fragile deployments. Analysts see weak realised ROI. VCs see second-order monetisation struggle.
Yet honesty carries career risk. Ambiguity carries none.
Publish ‘AI infrastructure is racing ahead of economic reality’ and you risk being labelled a sceptic who missed the cycle. Publish ‘AI is transformative but early’ and you remain directionally safe. Only one of those positions ends careers.
So rational actors choose survivable vagueness over precise critique.
This produces what might be called comfort compliance: benchmark scores, policy artefacts, model cards and process attestations that satisfy immediate reporting needs but do not guarantee reconstructability. They do not prove independent verification occurred before approval.
Comfort compliance often escalates into comfort outsourcing: advisory engagement treated as verification, comfort opinions treated as evidence, readiness assessments treated as assurance, and consultant sign-off treated as risk transfer. Under V = 0, that is structurally false. Confidence is procured. Verification is not supplied.
McKinsey reports sub-15% material impact from AI initiatives. Gartner reports over half of enterprise programmes being restructured or abandoned. These are not fringe signals. They are structural ones.
MIT research found that 95 per cent of enterprise generative AI projects delivered zero measurable financial return within six months. Not marginal gains. Zero bottom-line impact despite significant investment in infrastructure, training and deployment.
Even the most data-driven investors, those whose reputations rest on disciplined evidence, are signalling caution rather than acceleration. Capital is not behaving the way bubble narratives suggest.
This is what Epistemic Debt looks like in practice: an invisible liability accumulating while firms implicitly bet that future value will arrive before present assets become indefensible.
The gap is visible to millions. It just cannot be stated plainly without professional consequence. Because the failure does not become visible when a decision goes wrong. It becomes visible when justification is demanded, by force.
A regulator asks for a sample-based reconstruction: ten credit rejections, source lineage, override rationale, proof of human verification. The bank can produce packs and outputs, but not a decision record that links inputs to claims to verification to sign-off. Suddenly the institution is not arguing about model performance. It is arguing about capital and personal accountability.
A court asks a public body to produce reasons. Not policy statements, not dashboards, not ‘the system said so’. A lawful basis, a record of evidence considered, the human review applied, the chain of authorisation. What shows up is templated boilerplate and missing artefact trails. The outcome is not a technical debate. It is a legitimacy crisis.
An external auditor asks an insurer to show the evidential chain behind reserve movements. Data linkage, assumptions, independent challenge, period-on-period reconstruction. What they get is narrative, not auditability. When the trail collapses, confidence becomes irrelevant. So does the vendor’s slide deck.
And sometimes the compulsion is accounting, not courts. Capex was approved on enterprise deployability. Then customers ask for defensibility evidence that cannot be produced. Pilots stall. Utilisation stays in the teens. Under IAS 36 logic, impairment becomes unavoidable. The write-down is not a surprise. It is the bill for an assumption no one independently verified.
In each case, when the compulsion arrives, the upstream chain tends to disclaim in sequence: vendors disclaim outputs, advisors disclaim verification, assurance providers disclaim scope. The board retains the fiduciary, regulatory and litigation exposure. Mandated assurance is not mandated verification.
The Deeper Failure
Public AI discourse remains focused on hypothetical future misalignment. Russell. Bengio. Long-term risk.
Meanwhile, present misalignment is already operational.
Boards cannot defend AI-assisted decisions. Audit trails cannot reconstruct reasoning. Accountability dissolves under scrutiny.
The bottleneck is not compute capacity. It is epistemic confidence.
Until organisations can trust outputs, defend decisions and assign accountability, they will not deploy AI deeply enough to justify this infrastructure. So utilisation stays thin. Pilots proliferate. Real transformation stalls.
Organisations are not losing control of machines. They are losing control of their own decision legitimacy.
The risk is no longer speculative. It is already embedded in governance processes, financial approvals and legal exposure. That makes this a board-level failure, not a research debate.
The constraint is not technology. It is the absence of a governance architecture capable of restoring decision authority.
A practical test for any steering committee: can we reconstruct and defend this decision without relying on the originating system, the vendor, or the consultancy that sold it? If the answer is ‘not cleanly’, the organisation is not managing innovation risk. It is accumulating exposure and calling it progress.
What Happens Next
The arithmetic makes the mismatch impossible to ignore. Current infrastructure spending implies AI revenues must rise from roughly $20bn today to around $2tn annually by 2030 to justify the build-out. That is not incremental growth. That is a hundred-fold expansion in five years, while most enterprise deployments still report no measurable financial return.
I have spent the last year documenting this failure mode formally across architecture, governance, economics and market behaviour.
Not because I oppose AI adoption. But because I have watched too many organisations accumulate risk they cannot see, cannot measure and cannot explain.
If any of this feels familiar, you are not imagining it.
Footer
This article sits within a wider linked body of work on verification absence, AI governance and decision-level defensibility. The citable versions are available on Zenodo and SSRN. Readers who want the broader framework can explore the linked papers, companion papers and technical notes there.