John Radosta, Principal AI Engineer & Partner
February 20, 2026
15 min read
AI Transformation
After more than a hundred enterprise AI transformations across industries — financial services, healthcare, manufacturing, retail, and government — I've watched every flavor of failure and a much smaller set of genuine successes. The question I hear most often:
Roughly 80% of AI projects fall short of expectations, according to RAND Corporation research — a failure rate double that of conventional IT projects. BCG's 2025 survey of 1,250 companies found that just 5% generate AI value at scale, while 60% see no material impact at all. I've seen both sides of that distribution firsthand. What follows are the six lessons that separate the winners from everyone else.
In This Article
- The Data Problem Nobody Wants to Budget For
- Pilot Purgatory Is the Default — Escaping It Requires Workflow Redesign
- People and Process Account for 70% of the Equation
- Mid-Market Companies Move Faster But Face Different Constraints
- Generative AI Rewrote the 2025 Playbook — and Created New Pitfalls
- What the Top 5% Actually Do Differently
The Data Problem Nobody Wants to Budget For
Of every lesson from 100+ transformations, this one is the most universal and the least heeded. Data readiness is the single most consistent barrier we encounter, across every engagement we run. 43% of chief data officers cite data quality as their top obstacle. Gartner found that 63% of organizations either lack or are unsure whether they have adequate data management practices for AI.
The pattern is painfully predictable. Teams build an impressive proof-of-concept on clean, curated data in a sandbox. Then they try to scale against the messy reality of production systems — siloed databases, inconsistent formats, missing fields — and the model falls apart. Gartner projects that through 2026, organizations will abandon 60% of AI projects that lack AI-ready data foundations.
How Organizations Allocate AI Project Budgets
Failing vs. high-performing programs (% of total budget)
Data First, Then AI: Turning Scrap Logs into Margin
I worked directly on this engagement. A copper fabrication manufacturer wanted AI to cut scrap, but their MES exports were inconsistent across jobs, workcenters, and time periods. Before a single model was trained, we spent weeks standardizing part numbers, operations, workcenter names, time buckets, and cost fields across historical scrap and production files. Only then could we reliably tie scrap events to dollar impact. The result: a scrap-reduction copilot that highlights the highest-value hotspots, quantifies 5–10% improvements as a baseline, and lays a path toward 30%+ reductions. The real unlock wasn't the model — it was turning messy logs into an AI-ready asset.
Pilot Purgatory Is the Default — Escaping It Requires Workflow Redesign
AI Project Outcomes: Where Initiatives Break Down
% of original initiatives surviving each stage
Nearly two-thirds of organizations remain stuck in pilot mode, according to McKinsey. IDC puts it more starkly: for every 33 AI pilots launched, only 4 ever reach production — an 88% failure rate at the scaling stage.
The core issue isn't that pilots don't work. They usually do. The problem is organizations try to bolt AI onto existing processes rather than redesigning the processes themselves. A pilot that works in isolation exposes all the friction points of legacy workflows the moment you try to scale it.
McKinsey's 2025 survey tested 25 management attributes against EBIT impact and found one that towered above the rest: workflow redesign is the single highest-contributing factor to realizing business value from AI. Yet only 21% of organizations have fundamentally redesigned any workflows around AI.
Workflow Redesign: From Manual Data Ops to AI-Ready
I led this engagement firsthand. An automotive marketing provider ran dealer campaigns through brittle, manual workflows: daily FTP pulls, hand-cleaned CSVs, ad hoc SQL suppression, and one-off address checks, even as volumes hit 5,000–6,000 records a day. Instead of bolting AI onto this process, we rebuilt the workflow around an automated cloud pipeline that ingests feeds, validates addresses, applies suppression, and outputs campaign-ready lists with no human touch. Only on top of that redesigned flow could an AI layer learn from results, optimize targeting, and eventually decide who to market to and how — illustrating that real AI value came from reengineering the workflow, not just piloting a model.
People and Process Account for 70% of the Equation
BCG's widely validated 10/20/70 framework captures an uncomfortable truth: successful AI transformation is 10% algorithms, 20% technology and data, and 70% people and processes. Most organizations invert this ratio, pouring 60–70% of their budgets into technology while scrambling to fund organizational change. We see this constantly — it's the fastest path to a failed transformation.
46% of CxOs cite talent skill gaps as the primary reason AI projects stall. Only 35% of the workforce has received any AI training in the past year, despite 75% of companies actively adopting AI. AI specialist salaries average $400,000 — and 85% of tech executives report postponing projects due to talent shortages.
The companies seeing results invest in broad upskilling. EY committed $1.4 billion to AI transformation and put 83% of its workforce through foundational AI learning, logging 2 million hours and 115,000+ badges. Walmart is upskilling 50,000+ employees for AI roles.
BCG's 10/20/70 Framework
Where AI value actually comes from
Mid-Market Companies Move Faster But Face Different Constraints
Time from Idea to Production
Average months to deployment by company type
A significant portion of our work is with mid-market companies ($100M–$1B revenue), and they hold structural advantages over large enterprises. RSM's 2025 survey found mid-market GenAI adoption surged to 91%, with 88% reporting more positive impact than expected. Top performers achieve 90-day pilot-to-implementation timelines.
Mid-market advantages are real: smaller organizations treat AI as P&L assets, not science experiments. They target specific pain points without needing enterprise-wide consensus. Their cultures of ownership mean AI initiatives have clear business sponsors from day one.
The constraints are equally real. Mid-market companies lack dedicated AI Centers of Excellence. 70% report needing outside assistance to implement GenAI effectively — which is exactly where we spend most of our time.
Generative AI Rewrote the 2025 Playbook — and Created New Pitfalls
Enterprise GenAI spending hit $37 billion in 2025, a 3.2× increase year-over-year. Code generation emerged as AI's first true killer use case, with 50% of developers now using AI coding tools daily and teams reporting 15%+ velocity gains.
But the ROI picture is complicated. Microsoft-sponsored IDC research shows an average $3.70 return per $1 invested — yet McKinsey's 2025 survey found over 80% of respondents say their organizations are not seeing tangible enterprise-level EBIT impact. Only 1% of executives describe their GenAI rollouts as "mature."
The newest frontier is agentic AI — autonomous systems executing multi-step tasks without human intervention. McKinsey found 62% of organizations already experimenting, with 23% actively scaling. Gartner projects 40% of enterprise applications will integrate agents by end of 2026, up from less than 5% today.
Enterprise GenAI Spending Growth
Annual enterprise spending ($B) — 2023 to 2025
Where GenAI Falls Short
- Productivity gains don't aggregate to EBIT impact
- No governance for autonomous agents
- Only 1% of rollouts described as "mature"
Where GenAI Delivers
- Code generation: 15%+ velocity gains
- Agentic AI cutting cycle times from weeks to hours
- $3.70 return per $1 for structured deployments
What the Top 5% Actually Do Differently
Top 5% vs. Average: Key Performance Metrics
Indexed performance (1.0 = industry average)
They start with business outcomes, not technology. JPMorgan's 450+ production AI use cases generate an estimated $2 billion in annual benefits, tracked at the individual initiative level — not platform-wide vanity metrics.
They build platforms, not point solutions. JPMorgan's model-agnostic LLM Suite reached 200,000 daily users in eight months. Moderna's same AI pipeline that designs routine mRNA sequences accelerated COVID vaccine development without custom engineering.
They redesign work, not just add tools. Walmart's GenAI improved 850M+ product data points — a task requiring 100× the headcount manually — and eliminated 30 million unnecessary delivery miles, saving $75M in a single year.
They treat change management as the main event. Morgan Stanley's GenAI assistant achieved 98% adoption by wealth management teams because they didn't roll it out until it met advisers' quality standards.
The Defining Insight
The defining insight from 100+ AI transformations is counterintuitive: the technology is the easy part. The 5% of companies generating real value at scale aren't using fundamentally different algorithms or more advanced models. They're doing harder, less glamorous work — cleaning data for years before building models, redesigning workflows rather than adding AI to broken processes, investing 70% of transformation budgets in people and change management, and holding AI initiatives to the same ROI discipline as any other business investment.
The GenAI era has made the starting line more accessible than ever. Prototyping takes days, not months. Off-the-shelf tools deliver real productivity gains. But the gap between individual productivity and enterprise-level financial impact persists — and closing it still requires the fundamentals: executive sponsorship, data readiness, workflow redesign, and relentless organizational change.