The numbers are in, and they tell a story that should worry every CTO: despite spending $37 billion on generative AI in 2025 (up 3.2x from 2024), nearly two-thirds of organizations haven't begun scaling AI across their enterprise.
Let me say that again: 65% of companies investing heavily in AI still can't get it out of the pilot phase.
This isn't a technology problem. It's a process problem. And it perfectly illustrates why the traditional “AI strategy” playbook is broken.
The Scaling Gap Is Real
The research from Andreessen Horowitz and Menlo Ventures paints a clear picture:
- Companies are spending more than ever on AI ($37B in 2025 vs $11.5B in 2024)
- 37% of enterprises now use 5 or more different AI models
- The competitive landscape has shifted dramatically (Anthropic: 40% market share, up from 12% in 2023; OpenAI: 27%, down from 50%)
- Yet 65% still haven't scaled AI beyond experiments
What's happening? Organizations are treating AI adoption like traditional enterprise software procurement: lengthy planning cycles, massive upfront investments, and “big bang” deployments that never quite materialize.
Sound familiar? It should. It's the exact problem I outlined in my last post on moving from AI hype to value.
Why Traditional Planning Fails AI Initiatives
The data shows what I've observed repeatedly in the field: the problem-first companies are succeeding while the AI-first companies are stuck.
McKinsey's research confirms this: high-performing AI organizations are three times more likely to have senior leaders demonstrate genuine ownership and commitment. But here's the critical insight—that leadership isn't about grand AI strategies. It's about disciplined experimentation and rapid learning.
The companies still stuck in pilot purgatory share common patterns:
- They lead with technology, not problems. “We need an AI strategy” sounds proactive, but it's actually backward. You need a business strategy that might leverage AI where it creates value.
- They plan for perfection. Six to twelve months of planning before any real experimentation. By the time they're ready to pilot, the landscape has shifted, requirements have changed, and the team has lost momentum.
- They bet big before proving value. Massive platform investments, enterprise-wide rollouts, vendor lock-in before understanding what actually works for their specific problems.
- They lack the data foundation. You cannot skip data quality and governance. Period. AI built on messy data delivers messy results, no matter how sophisticated the model.
The Portfolio Approach Is Winning
Here's what the successful minority is doing differently:
- Multi-model deployment. 37% now use five or more models. This isn't indecision—it's pragmatism. Different problems need different tools. The winners maintain a portfolio of experiments, each matched to specific business problems.
- Buy vs. build shift. The ecosystem has matured enough that buying purpose-built AI applications often beats custom builds. This lets teams prove value faster and allocate engineering resources to truly differentiating capabilities.
- Collaborative leadership. Success requires a triumvirate of CIO/CTO, CFO, and CSO working together. AI initiatives fail when they're siloed in IT or treated as pure technology plays.
- Problem-first orientation. Every successful initiative I've studied started with a clear pain point, not a technology vision.
This is exactly what the 9-week experiment cycle enables.
From Experiments to Scale: The Missing Playbook
The gap between pilot and production isn't a mystery. It's a result of skipping the systematic learning phase.
Here's what should happen:
Weeks 1–9: First experiment
- Pick one specific problem
- Test the hypothesis with minimal investment
- Gather real data on feasibility and value
- Make a go / pivot / kill decision based on evidence
Weeks 10–18: Parallel experiments
- Launch 2–3 experiments simultaneously
- Share learnings across the portfolio
- Build organizational capability through doing
- Identify patterns in what works
Weeks 19+: Scale what proves out
- Productionize the 1–2 experiments that delivered
- Kill the rest without guilt
- Apply learnings to new experiments
- Build a culture of rapid iteration
Notice what's missing? Year-long roadmaps. Massive upfront investments. “AI Centers of Excellence” that never ship.
The Real Competitive Advantage
The a16z research reveals something crucial: the winners aren't spending the most. They're learning the fastest.
Anthropic's rise from 12% to 40% market share didn't happen because enterprises made one-time vendor decisions. It happened because organizations running portfolios of experiments found specific use cases where Claude outperformed alternatives—then shared those learnings across their portfolio.
This is the experiment-driven approach in action.
The companies still stuck aren't failing because they picked the wrong model or vendor. They're failing because their organizational process can't support rapid learning.
What This Means For You
If you're in the 65% still struggling to scale:
This week
Audit your current “AI initiatives.” How many are experiments with clear go/no-go criteria versus open-ended pilots? If you can't clearly articulate the success criteria and decision timeline, you're already stuck.
This month
Pick your most promising pilot and apply the 9-week framework to it. Set a decision date. Define what “success” looks like with specific metrics. Empower a small team to move fast.
This quarter
Build a portfolio. Launch 3–5 small experiments in parallel. Celebrate the kills as much as the wins—every failed experiment that costs $50K saves you from a failed platform that costs $5M.
The Hype Correction Is Your Opportunity
MIT Technology Review called 2025 “the start of a much-needed hype correction” for AI. The GPT-5 launch disappointed. The reality is setting in: AI isn't magic, and it won't solve problems you can't clearly define.
This is actually great news.
The hype correction means:
- Executives are ready for honest conversations about value
- Budgets will flow to proven results, not promises
- The experiment-driven approach becomes easier to sell
- Failed pilots stop being covered up and start being learning opportunities
The companies winning with AI aren't the ones with the biggest “AI strategies.” They're the ones with the most disciplined learning loops.
65% of enterprises are stuck in pilot purgatory. Don't be one of them.
Moving Forward
The path from pilot to production isn't mysterious. It requires:
- Data-as-a-product foundation
- Problem-first orientation
- Experiment-driven culture
- Collaborative leadership
- Willingness to kill what doesn't work
The 9-week experiment cycle gives you the framework. The research proves it works. The only question is: are you ready to stop planning and start learning?
Join the Conversation
Where is your organization in the scaling journey? Are you stuck in pilot mode or building a portfolio of experiments? I'd love to hear what's working (or not working) for you—reach out on LinkedIn.