Why AI Initiatives Fail Even When Organisations Know Better

The problem is not that the playbook is unknown. It is that many organisations are still governing AI through a model built for a different category of spending.

Ninety-five per cent of organisations in MIT's GenAI Divide study report zero measurable return from their generative AI efforts.¹ Gartner warns that, through 2026, organisations will abandon 60 per cent of AI projects not supported by AI-ready data.² Cisco's AI Readiness Index finds only 13 per cent of organisations fully prepared across all six readiness pillars.³

None of this is hidden. The guidance for avoiding failure is widely available: prepare the data, redesign workflows before deployment, build adoption into the operating model, assign executive accountability for value capture rather than implementation alone.

Programmes fail anyway.

When an entire market knows the playbook and still fails to apply it, the problem is unlikely to be informational. Something in the way organisations approve, fund, and oversee AI investment is preventing the known solutions from being applied. The 5 Whys method is useful here because it separates symptoms from structure. It asks not only what is failing, but what system keeps reproducing the same failure.

One implication follows immediately. The standard board question — "Is the AI programme on track?" — may be too narrow. It tests execution inside a governance model that may itself be the primary constraint.

Why 1: Why do AI programmes fail despite the playbook being known?

Because organisations approve AI investment without confirming the conditions required for success.

The readiness gap is well documented. Data is not prepared. Infrastructure is not ready. Operating teams are not structured to absorb the change. Yet budgets are approved and pilots launched anyway. The conversion gap follows. Deloitte's 2026 State of AI in the Enterprise survey finds that only a quarter of enterprises have moved even 40 per cent of their AI pilots into production.⁴

That distinction is central. Readiness is often treated as the first operational problem. In practice, it is often the first sign that the wrong governance logic has already been applied.

The question, then, is not simply why readiness is weak. It is why weak readiness keeps being allowed through the gate.

Why 2: Why are organisations approving AI investment without confirming preconditions?

Because AI investment is still processed through approval mechanisms designed for discrete technology purchases.

A conventional capital approval process asks familiar questions: what will this cost, what return should it produce, when does it go live, is the technology feasible? Those are reasonable tests for an ERP module, a cloud migration, or a systems upgrade. They are not sufficient for an investment whose value depends on workflow redesign, behavioural adoption, data discipline, and cross-functional operating change.

That is the governing mismatch. AI transformation does not behave like a bounded technology purchase. Its value does not arrive at deployment. It arrives only when the organisation changes how work is done around the technology. Continuous, not discrete.

The same approval mechanism is applied to both categories of spending because no alternative mechanism exists.

Once that mismatch exists, the downstream pattern is predictable. Funding is approved one use case at a time. Teams build local tools and separate data pipelines. Technical complexity accumulates before scale is even attempted. The organisation mistakes authorisation for readiness. Deploying AI tools into unchanged workflows does not create value — it creates the appearance of progress while the underlying decision quality remains unaddressed.

Why 3: Why does this approval model remain the default?

Because competitive pressure compresses the decision timeline.

Boards are not wrong to feel pressure. AI is now treated not only as an innovation topic but as a strategic exposure. Delay appears risky and organisations do what institutions often do under pressure: they push new demands through the fastest existing channel.

The urgency is real. The governance conclusion that follows is wrong.

When serious institutions face real time pressure, they do not merely speed up inherited mechanisms. They redesign them. During COVID-19, regulators used rolling reviews so evidence could be assessed as it became available, rather than waiting for one final file at the end. The process changed so that speed could increase without abandoning rigour.

Many AI programmes are doing the reverse. The timeline is compressed, but the governing logic is left intact.

That choice matters more than it appears. Speeding up an unfit gate does not produce better decisions faster. It reproduces the same classification error at higher speed.

Why 4: Why does this not self-correct after deployment begins?

Because executive sponsorship is being mistaken for executive ownership.

The distinction matters. Sponsorship helps a programme begin. Ownership keeps it connected to the conditions that determine whether value ever appears: workflow redesign, adoption decisions, data remediation, capability build, and economic accountability after the technology is live. When AI governance is delegated to technical teams rather than owned at senior level, the organisation loses the ability to connect what AI produces to the decisions that actually matter.

Where that ownership is absent, reporting predictably drifts toward what is easiest to show. Milestones completed. Integrations delivered. Model performance improved. Those indicators may all be useful. None answers the deeper question: is the organisation converting technical deployment into operating value?

The evidence runs in the same direction across multiple studies. Prosci's latest change-management research finds that projects with extremely effective sponsors are 79 per cent likely to meet objectives, compared with 27 per cent where sponsorship is extremely ineffective.⁵ BCG's 2024 research on large-scale technology programmes points to the same governance pattern: more than two-thirds of major tech programmes are not expected to deliver on time, within budget, or within planned scope, with weak governance, poor business–technology collaboration, and insufficient senior-leadership involvement among the recurring causes.⁶

A programme can look technically healthy and fall short of meeting the desired outcomes.

Why 5: Why does no one redesign the governance model?

Because most organisations do not have a standing mechanism that connects the investment decision to the operating conditions required for success.

AI cuts across technology, finance, operations, data, risk, and workforce design at the same time. Most governance structures do not. Procurement approves spend. Technology implements. Finance tracks budget. Operations are expected to adapt. HR enters later, if at all. Each function may perform its own role reasonably well. What is often missing is a forum explicitly responsible for testing whether those roles remain connected over time.

Without that integrative mechanism, failure is almost always explained locally. The wrong use case. Weak data. Poor training. A bad vendor. An immature team. Any of those may be true in a given case. None explains why the same pattern keeps recurring across a portfolio. The same governance architecture is being applied repeatedly to a category of investment it was not designed to govern.

MIT CISR reaches the same governance conclusion from a different angle: traditional technology governance assumes stable technologies, predictable consequences, and manageable demand, while generative AI changes faster than conventional governance mechanisms can adapt. Its proposed answer is not looser governance, but minimum viable governance: structurally agile, trustworthy by design, integrated end-to-end, and opportunity-sensitive.⁷

Meanwhile, the workforce is not waiting. MIT CISR distinguishes between GenAI tools that enhance individual productivity and GenAI solutions designed to create organisational value by changing processes, systems, and offerings.⁸ Adoption of the first category — often through unvetted public tools, what MIT CISR terms shadow GenAI — typically outpaces the formal governance that the second category requires. When AI use outruns governance, the organisation is not innovating faster — it is accumulating risk it cannot yet see. The transformation is happening outside common controls, outside integrated design, and often outside any coherent value-capture model.

The Root Cause

Each layer points back to the same origin.

AI initiatives fail because organisations approve them before they are ready to absorb them. Readiness is not tested rigorously because the approval mechanism was built for a different category of spending. That mechanism persists because urgency speeds up the old gate instead of forcing a new one. The problem does not self-correct because sponsorship substitutes for ownership. And the pattern remains hard to see because no standing governance forum is designed to connect spend, workflow, adoption, data, and value across the life of the programme.

The root cause is architectural.

Boards are routing a continuous organisational transformation through a discrete capital approval gate. Until that governance model is redesigned with the deliberate urgency that regulators brought to vaccine approvals during COVID-19, the same failure pattern will continue to reproduce, even when the technology works and even when the advice is already known.

Five Questions for the Next Board Review

A board reviewing AI should test the governance model, not only the programme status.

Has readiness been tested independently before approval: not only technical readiness, but data, workflow, and workforce readiness as separate preconditions?

Does the approval process distinguish between a bounded technology purchase and a continuous transformation programme?

Where urgency is real, what has actually been redesigned in the governance channel itself, rather than merely accelerated?

Who owns the operating outcome after deployment: not sponsorship in name, but adoption, redesign, and value capture in practice?

Is the board reviewing AI as a portfolio pattern, so that recurring structural causes of failure can be seen early?

The right question at the next board review is not whether the AI programme is on track. It is whether the governance architecture was ever designed to let it succeed.

Notes

¹ MIT NANDA (Networked Agents and Decentralized AI) initiative, The GenAI Divide: State of AI in Business 2025, July 2025. The report presents directional findings on enterprise GenAI adoption and value realisation, including the finding that most initiatives showed no measurable P&L impact.

² Gartner, "Lack of AI-Ready Data Puts AI Projects at Risk," press release, February 2025. Underlying data from a Gartner survey of 1,203 data management leaders, July 2024.

³ Cisco, AI Readiness Index 2024, November 2024. Double-blind survey of 7,985 senior business leaders with responsibility for AI integration and deployment at organisations with 500 or more employees, across 30 markets. The six pillars assessed are: Strategy, Infrastructure, Data, Governance, Talent, and Culture.

⁴ Deloitte AI Institute, State of AI in the Enterprise 2026, January 2026. Survey of 3,235 business and IT leaders across 24 countries, conducted August–September 2025.

⁵ Prosci, Best Practices in Change Management, latest edition. The figures cited are drawn from Prosci's longitudinal sponsor-effectiveness research, comparing project outcomes by sponsor effectiveness rating.

⁶ Boston Consulting Group, Most Large-Scale Tech Programs Fail: How to Succeed, 2024. Drawn from BCG's 2024 Build for the Future global study, which surveyed more than 1,000 C-suite executives across 20 sectors. BCG defines "large-scale tech programs" as those involving more than 3 per cent of annual technology budget. Among the recurring failure modes, BCG identifies the absence of an active programme management office (cited by approximately 60 per cent of laggard organisations), insufficient senior-leadership involvement, and weak business–technology collaboration.

⁷ Nick van der Meulen, Jennifer Jewer, and Nadège Levallet, Minimum Viable Governance for Generative AI, MIT CISR Research Briefing No. XXVI-3, 19 March 2026.

⁸ Nick van der Meulen and Barbara H. Wixom, Managing the Two Faces of Generative AI, MIT CISR Research Briefing Vol. XXIV, No. 9, 19 September 2024. Findings drawn from three virtual roundtables and 23 semi-structured interviews with 93 data and technology executives from 50 global organisations with annual revenues above US$1 billion.

John Valavanis FCCA is the Founder of Noema Advisory, a boutique finance advisory practice providing senior finance leadership in complex, regulated, and cross-border environments, with a focus on the UK–Greece corridor.