ai_strategy_execution_berkins.html

AI Strategy Execution:

Why Most AI Initiatives Fail After Planning

Published by Berkins Consulting  | Strategy & AI Transformation Practice | 2025

80%

AI Project Failure Rate

RAND Corporation, 2024

42%

Companies Abandoned AI Initiatives

S&P Global, 2025

5%

Firms Realising Significant Value

BCG Global AI Survey, 2025

95%

GenAI Pilots Failing

MIT NANDA Initiative, 2025

The Execution Problem No One Wants to Talk About

Every boardroom has a version of the same story. The strategy deck was brilliant. The AI roadmap was approved with enthusiasm. The consulting firm delivered a polished framework. Pilot programs were launched. Leadership posted about innovation. And then — very quietly — the whole thing stalled.

That story has become so common that it barely registers as failure anymore. It gets reframed as 'learnings.' Teams are quietly reassigned. Budgets are repurposed. And twelve months later, a new vendor arrives with a new pitch, and the cycle begins again.

This article is about that gap — the dangerous stretch of ground between a well-designed AI strategy and actual, measurable business results. It is written for the executives, operations leaders, and transformation heads who are tired of paying for strategy documents that gather digital dust.

The numbers, frankly, are sobering.

KEY INSIGHT:  Organizations reporting 'significant' financial returns from AI are twice as likely to have redesigned end-to-end workflows before selecting modelling techniques — not after. The problem is not vision. It is sequencing, ownership, and execution discipline.

According to S&P Global's 2025 survey of over 1,000 enterprises across North America and Europe, 42% of companies abandoned most of their AI initiatives that year — up sharply from just 17% in 2024. RAND Corporation confirms the broader picture: over 80% of AI projects fail, a rate that is twice that of traditional IT projects. BCG's 2025 global AI survey found that only around 5% of companies are realising significant value from AI, while close to 60% report little to no benefit despite active investment.

The gap is not a technology gap. It is an execution gap. And it is widening.

The Anatomy of an AI Initiative That Looks Fine — Until It Isn't

Understanding why AI initiatives collapse requires understanding what they look like in their early stages. Most organisations do not fail at planning. They fail in the transition — the moment when the strategy document becomes a delivery mandate and no one quite knows who owns what.

The Pilot Trap

The most common entry point to failure is what practitioners have begun calling 'pilot paralysis.' Organisations launch proof-of-concept projects in controlled sandboxes, often with handpicked data and simplified workflows. The prototype performs well. Leadership is impressed. Confidence is high.

What happens next reveals the execution gap in full clarity. The technology worked in isolation — but integration with existing enterprise systems, compliance workflows, real user behaviour, and live data pipelines was never part of the pilot design. When the business asks 'when can we go live?', the team realises it has been optimising an F1-score rather than building a production system.

Gartner reports that it takes an average of eight months to move from AI prototype to production — for the projects that actually make it that far. Gartner also predicts that at least 50% of generative AI projects will be abandoned at the pilot stage due to poor data quality and misaligned expectations.

BERKINS PERSPECTIVE:  Pilot success is not the same as execution readiness. A proof of concept that cannot answer 'how does this integrate with our compliance team's workflow?' is not a proof of concept — it is a prototype in costume.

The Data Readiness Illusion

Ask any enterprise AI team whether their data is ready. Most will say yes. The honest answer, in most cases, is no.

Informatica's CDO Insights 2025 survey identifies data quality and readiness (43%) and lack of technical maturity (43%) as the top two obstacles to AI success. According to Wipro's State of Data4AI 2025 report, only 14% of business leaders believe their data maturity can support AI at scale. Yet 79% of the same respondents believe AI is essential to their company's future.

That disconnect — between data confidence and data reality — is one of the most consistent predictors of AI initiative failure. When organisations treat data readiness as a checkbox rather than a foundation, they build AI systems on unstable ground. The outputs are unreliable. Trust erodes. Adoption collapses.

Organisations that succeed with AI invert the typical spending ratio: they earmark 50–70% of the AI project timeline and budget for data readiness — extraction, normalisation, governance metadata, quality dashboards, and retention controls. That is not glamorous. But it is what works.

Assumption

Reality

Common Assumption

What Berkins Actually Finds on Engagement

'Our data is clean and centralised'

Siloed systems, inconsistent field definitions, multiple versions of truth across departments

'Our teams know which data to use'

Unclear data ownership, undocumented sources, no formal governance in place

'We can fix data issues during AI build'

Data debt accumulates. AI models trained on poor data propagate errors at scale

'Compliance will be straightforward'

Privacy, auditability, and legacy governance models create significant friction during deployment

The Leadership Alignment Gap

One of the most reliable leading indicators of AI initiative failure is the absence of genuine executive ownership. Not executive sponsorship — which is frequently symbolic — but actual accountability: a named executive whose performance metrics include AI delivery outcomes.

Harvard Business Review analysis reveals a striking internal disconnect: executives report feeling 82% aligned with their company's strategy, while actual measured alignment sits at just 23% — nearly three times lower than perceived. In AI programmes, this gap is particularly damaging. When the CTO believes the programme is on track and the CFO believes it has been deprioritised, and neither is wrong, it is because no one has defined what 'on track' means in shared terms.

A Gallup poll from late 2024 found that only 15% of US employees report that their workplace has communicated a clear AI strategy. Separately, nearly half of CEOs in one global survey acknowledged that most of their employees were resistant or even openly hostile to AI-driven changes — and cited the top obstacle as not a technology limitation, but a lack of effective change management.

LEADERSHIP NOTE:  Resistance is not irrational. It is what happens when organisations ask people to change the way they work without explaining why, without training them how, and without protecting them from the consequences. That is a leadership failure, not a technology failure.

The Tribal Organisation Problem

Modern AI initiatives require cross-functional collaboration at a depth most organisations have never attempted. Product teams chase feature velocity. Infrastructure teams harden security. Data teams clean pipelines. Compliance officers draft policies. Legal reviews vendor agreements. Each team operates rationally within its own domain — but no one is coordinating across all of them.

The result, as WorkOS research describes it, is 'disconnected tribes': teams working in parallel without shared success metrics, coordinated timelines, or a common definition of what deployment actually means. When these teams eventually surface for a go-live review, they discover that each department has been optimising for a different outcome.

This is not a personality problem. It is a structural problem — and it requires structural solutions.

Five Execution Failures That Kill AI Programmes

Based on Berkins Consulting's work across enterprise AI transformation engagements, and supported by the broader research literature, five categories of failure account for the vast majority of AI initiative breakdowns. They are rarely standalone — they compound each other. But they are also preventable.

01

Strategy Without Ownership

An AI roadmap is not a delivery plan. When every initiative 'belongs to everyone,' accountability lives nowhere. Programmes need a named executive whose compensation is tied to specific AI outcomes — not just a steering committee that meets quarterly.

02

Pilot Success Mistaken for Deployment Readiness

A proof of concept that works in a sandbox has demonstrated technical feasibility, not organisational readiness. Deployment requires integration architecture, change management, training pipelines, and governance — none of which pilots typically address.

03

Treating AI as a Technology Project

AI transformation is a business transformation that uses technology. When it is owned by IT rather than co-owned by business units, it optimises for technical performance rather than business outcomes. The gap between these two things can cost millions.

04

Skipping Workforce Readiness

Technology without workforce readiness is a costly experiment. Employees who do not understand, trust, or know how to use AI tools will route around them — reverting to old workflows, nullifying any efficiency gain. Adoption is not automatic; it is earned through preparation.

05

No Feedback Loop Between Strategy and Reality

AI is not static software. It requires continuous monitoring, retraining, and refinement. Organisations that treat deployment as a finish line will find that their models drift, their outputs degrade, and their ROI disappears — often without anyone noticing until significant damage is done.

A Berkins Case Study: From Stalled Pilot to Scaled Deployment

The following scenario reflects a pattern Berkins Consulting encounters regularly. The details have been composited to protect client confidentiality, but the dynamics — and the solutions — are real.

ENGAGEMENT PROFILE:  Sector: Financial Services | Organisation Size: 3,200 employees | AI Initiative: Customer risk scoring and loan decisioning | Status at Engagement: 14 months post-strategy, zero production deployments

The Situation

A mid-sized financial services firm had invested significantly in an AI-powered loan decisioning system. The strategy had been signed off at board level. The technology vendor had been selected after a rigorous RFP process. The data science team had built and validated a model with strong performance metrics in the test environment.

Fourteen months later, the system had not processed a single live loan application. The business had continued operating on its legacy decisioning rules. The AI model was sitting idle on a cloud server, maintained by two data scientists who were running increasingly infrequent retraining cycles on data that was growing stale.

When Berkins was brought in, the initial brief was to 'accelerate deployment.' Within the first two weeks, it became clear that acceleration was not the problem. The problem was that no one had ever defined what deployment actually required — from a compliance standpoint, an operations standpoint, or a workforce standpoint.

What Berkins Found

The diagnosis revealed five interconnected gaps that had been invisible during the planning phase:

•       The AI model had been built using historical loan data that predated a significant regulatory change. The model's decisioning logic, while technically valid, was non-compliant with current requirements. No compliance stakeholder had been part of the model design process.

•       The operations team responsible for loan processing had never been consulted about how the AI system would integrate with their existing workflow. Their process required a human-readable explanation for every credit decision — something the model's initial architecture could not provide.

•       The IT infrastructure team had not been included in the vendor selection process. The AI platform had been chosen without evaluating its compatibility with the organisation's core banking system, which ran on a legacy architecture.

•       There was no designated executive owner for the programme. The CTO owned the technology. The Chief Risk Officer owned the credit policy. The Chief Operations Officer owned the process. No one owned the programme.

•       The data pipeline feeding the model had never been stress-tested with live data volumes. In the test environment, inference ran in under two seconds. Under realistic production load, preliminary testing revealed latency spikes that would breach the firm's customer service commitments.

The Berkins Approach

Rather than attempting to accelerate a broken deployment, Berkins restructured the programme around a phased execution model — one built on business readiness, not technical milestones.

Phase 1

Governance First (Weeks 1–4)

A cross-functional AI Deployment Council was established, with a designated Programme Executive drawn from the business — not IT. Clear ownership was defined for every workstream: compliance, operations integration, infrastructure, data governance, and workforce readiness. A shared definition of 'deployment' was agreed across all stakeholders before any technical work resumed.

Phase 2

Data and Compliance Remediation (Weeks 5–12)

The existing model was reviewed by the compliance team, supported by Berkins' regulatory framework expertise. A model redevelopment sprint was conducted with compliance embedded in the process from the outset. Data lineage was documented. A governance dashboard was built to give the Chief Risk Officer real-time visibility into model inputs, outputs, and confidence scores.

Phase 3

Workflow Redesign (Weeks 10–16)

Operations leaders co-designed the new loan decisioning workflow — including the AI system's role, the human-in-the-loop checkpoints, and the escalation protocols for edge cases. The system was redesigned to produce explainable outputs that met the operations team's requirement for human-readable decision rationale.

Phase 4

Workforce Enablement (Weeks 14–20)

A structured training and change management programme was rolled out across the loan processing team. This included both technical training on the new interface and narrative communication about why the system was being introduced, how it would change their roles, and how their expertise would remain central to the final decision. Resistance dropped measurably after the first training cohort.

Phase 5

Controlled Deployment and Feedback Loops (Weeks 18–26)

A shadow deployment ran the AI system in parallel with legacy decisioning for six weeks, allowing the team to validate outputs, identify edge cases, and build operational confidence. Live deployment followed, with a monitoring framework that tracked model performance, data drift, compliance adherence, and user adoption in real time.

The Outcomes

Twenty-six weeks after Berkins' engagement began, the system was live. Not in a pilot. Not in a shadow mode. In full production, processing live loan applications with documented compliance sign-off and operational integration.

26 wks

From Engagement to Live Deployment

vs. 14 months stalled

91%

Loan Processor Adoption Rate

End of first month live

38%

Reduction in Decision Turnaround

vs. legacy process baseline

Zero

Compliance Findings Post-Deployment

First regulatory review

The financial outcome exceeded original projections — not because the AI model was more sophisticated than planned, but because it was actually being used.

BERKINS REFLECTION:  The model that had sat idle for fourteen months was not the problem. The absence of execution discipline was. Berkins did not build a better AI. We built the conditions in which AI could actually work.

Closing the Strategy-to-Execution Gap: What Actually Works

The research is consistent across McKinsey, BCG, Bain, Gartner, Deloitte, and the World Economic Forum. AI does not fail at the technical level. It fails at the organisational level. The gap is not about algorithms — it is about alignment, ownership, readiness, and the discipline to treat deployment as the beginning of a programme, not the end of one.

What follows is not a prescriptive playbook — every organisation's context differs. But it is a set of principles that separate organisations that are realising AI value from those that are not.

1. Make Business Outcomes the North Star, Not Technical Milestones

The most dangerous metric in an AI programme is model accuracy. Not because accuracy does not matter, but because it can be achieved in a sandbox while the business problem remains entirely unsolved.

Before any AI initiative begins, Berkins advises establishing a set of business outcome metrics that will define success — not technical performance indicators, but real-world measures: decision turnaround time, cost per transaction, customer satisfaction score, error rate on live data. Every technical decision in the programme should be traceable to one of these metrics. If it cannot be traced, it should be questioned.

McKinsey's 2025 findings confirm that organisations reporting significant financial returns are twice as likely to have redesigned end-to-end workflows before selecting modelling techniques. The implication is clear: start with the business problem, not the technology.

2. Establish a Cross-Functional Execution Council with Real Authority

The steering committee model — where stakeholders meet monthly to receive updates and provide symbolic endorsement — is insufficient for AI transformation. What is required is a cross-functional council with actual decision-making authority, clear ownership of workstreams, and accountability for delivery.

This council should include a designated Programme Executive from the business (not IT), representatives from compliance, legal, operations, data, IT infrastructure, and HR. It should meet weekly during critical phases. It should have a shared definition of deployment readiness — agreed before the first line of model code is written.

3. Invest in Data Before You Invest in Models

Winning AI programmes invert the typical budget allocation. Instead of spending 70% on model development and 30% on data, they spend 50–70% of the project budget on data readiness — and treat that investment as non-negotiable.

This means establishing a single source of truth, implementing quality controls, tracing data origin and usage, documenting lineage, and building governance frameworks that satisfy both business and regulatory requirements. It means treating data as a product rather than a byproduct.

Only 14% of business leaders believe their data maturity can support AI at scale, according to Wipro. If your organisation is in the other 86%, that is the place to start — not with the model.

4. Design for Adoption from Day One

Workforce readiness is not a deployment activity. It is a programme design activity. The people who will use, oversee, and be affected by an AI system need to be part of its design from the earliest stages — not informed of its existence at go-live.

This means co-designing workflows with operations teams, addressing job security concerns openly and honestly, building training programmes that explain both how the system works and why it has been introduced, and measuring adoption as a core delivery metric alongside technical performance.

Employee resistance is not an obstacle to be overcome after deployment. It is a signal that the programme failed to engage its stakeholders during design. In Berkins' experience, the organisations with the highest AI adoption rates are those that began workforce engagement before the model existed.

5. Build for Continuous Learning, Not Point-in-Time Deployment

Treating AI deployment as a finish line is one of the most expensive mistakes an organisation can make. AI systems require continuous monitoring, retraining, and refinement. Data drifts. Business conditions change. Regulatory environments evolve. A model that was accurate at deployment can become a liability within months if it is not actively maintained.

Effective AI programmes establish monitoring frameworks from day one — tracking model performance, data drift, compliance adherence, and user adoption in real time. They build feedback loops between the AI system and the business stakeholders who depend on it. They treat AI as a living capability, not a product to be shipped.

Bain's research on AI strategy and execution emphasises that transformation should be treated as a business endeavour — not a technology project — with a 'self-funding journey that balances short-term gains with long-term capabilities.' That framing requires a very different operating model than a traditional technology deployment.

The Berkins AI Execution Framework

Drawing on our engagement experience and the broader research literature, Berkins has developed an execution framework for AI initiatives designed to close the strategy-to-execution gap systematically. It is built around five dimensions of organisational readiness — because readiness, not technology, is what determines outcomes.

Dimension

Key Readiness Questions

Strategic Alignment

Are AI initiatives directly traceable to board-approved business priorities? Is there a named executive accountable for business outcomes — not technical delivery?

Data & Infrastructure Readiness

Is the data foundation documented, governed, and validated for the specific AI use case? Has infrastructure been stress-tested under production conditions?

Governance & Compliance

Are compliance, legal, and risk stakeholders embedded in model design — not consulted at deployment? Is there an explainability framework for AI decisions?

Workforce Readiness

Have operations teams co-designed the AI-integrated workflow? Is there a training and adoption programme with measurable targets?

Monitoring & Iteration

Are real-time performance dashboards in place? Is there a defined retraining cadence? Who owns ongoing model governance post-deployment?

Organisations that score strongly across all five dimensions before committing to production deployment consistently outperform those that prioritise speed. The organisations that fail are almost always the ones that scored well on technical dimensions but neglected governance, workforce, and monitoring.

What Leadership Must Do Differently

The research is unambiguous: AI transformation fails when organisations treat it as a technology implementation. Successful adoption requires rethinking not only who does the work but how the work gets done — which means leadership must lead differently.

This is not abstract. It means specific behavioural changes at the executive level.

•       Stop measuring AI progress by pilot count or model accuracy. Start measuring by production deployments, adoption rates, and business outcome impact.

•       Stop separating AI strategy from business strategy. Every AI initiative should be traceable to a specific business outcome, owned by a business leader — not an IT function.

•       Stop treating workforce resistance as an implementation problem. It is a design problem. Address it before the model is built, not after it is deployed.

•       Stop accepting 'the data isn't ready' as a constraint. Make data readiness a programme prerequisite — and resource it accordingly.

•       Stop defining 'done' as deployment. Define 'done' as sustained adoption with measurable business impact. Everything before that is pre-flight checks.

As Sylvain Duranton, Global Leader of BCG X, has stated: 'Companies cannot simply roll out GenAI tools and expect transformation. The real returns come when businesses invest in upskilling their people, redesign how work gets done, and align leadership around AI strategy.'

That observation reflects what Berkins sees in practice, consistently. The firms that are realising genuine returns from AI did not have better technology than their peers. They had better execution discipline — and leadership that understood the difference.

Conclusion: The Competitive Advantage Is Not the Algorithm

The AI landscape in 2025 is not short of ambition. The strategies are impressive. The technology is genuinely capable. The investment is substantial — total corporate AI investment reached $252 billion in 2024, and enterprise generative AI spending grew sixfold in the same year.

What is short is execution rigour. And that gap — between strategy and delivery — has become the defining competitive differentiator for organisations navigating the AI era.

The firms that will define their industries over the next decade will not be those with the most sophisticated models. They will be those with the clearest understanding of how their people, processes, and governance structures need to change to make AI work at scale. Those that treat AI deployment not as the end of an initiative, but as the beginning of a sustained capability.

The window for competitive advantage is open. But it does not stay open indefinitely. Organisations that invest now in execution discipline — not just strategy documentation — will be the ones that look back on this period as the moment they pulled ahead.

Those that keep producing strategy decks without closing the execution gap will have very good documentation of why they fell behind.

ABOUT BERKINS CONSULTING:  Berkins Consulting works with executive teams across financial services, healthcare, manufacturing, and professional services to design and deliver AI transformation programmes that close the strategy-to-execution gap. We do not sell technology. We build the conditions in which AI can work — and stay working.

Research Sources & References

This article draws on the following research and industry sources:

•       S&P Global Market Intelligence, Enterprise AI Initiatives Survey (2025) — AI abandonment rates and pilot-to-production conversion

•       RAND Corporation, AI Project Failure Rate Analysis (2024) — 80% failure rate benchmark

•       BCG, Global AI Survey (2025) — Value realisation and scaling barriers

•       MIT NANDA Initiative, GenAI Divide: State of AI in Business (2025) — Pilot failure rates

•       McKinsey & Company, State of AI Report (2025) — Workflow redesign and ROI correlation

•       Gartner, AI Deployment Research (2025) — Pilot-to-production timelines

•       Informatica, CDO Insights Survey (2025) — Data quality as AI obstacle

•       Wipro, State of Data4AI Report (2025) — Data maturity and AI ambition gap

•       Harvard Business Review / McKinsey — Strategic alignment gap research

•       World Economic Forum, Closing the Intelligence Gap (2025) — Workforce readiness and data governance

•       Bain & Company, The Gap Between AI Strategy and Reality is Execution (2025)

•       Gallup, AI Strategy Communication Survey (2024) — Employee awareness of AI strategy

•       DXC Technology / Insight Jam, The AI Strategy Execution Gap (2025)

•       Atlassian, State of Teams (2025) — Strategic alignment and team performance

© 2025 Berkins Consulting. All Rights Reserved. For reprint permissions or to discuss an engagement,