· Company Agility · 5 min read
AI Isn’t Failing — Our Operating Systems Are
95% of AI projects fail — not because the models are wrong, but because organizations use project-mode management for high-uncertainty work. How agile operating principles change the AI success rate.
Executives keep hearing that “95 percent of AI projects fail.”
The truth is uglier: it’s not the models that are failing, it’s the management operating systems we’re using to leverage them.
💡 Insight: The AI/OS Mismatch
AI is high-potential/high-uncertainty work. Managing it with 1990s waterfall thinking is why 95% of projects fail. AI needs a Product Mindset, not a Project Plan.
Most organizations treat AI like a project — scope, budget, work breakdown structure, milestones. But looking for ways to leverage/operationalize AI is prime example of high potential with high uncertainty. Will this be a practical use case? Is it feasible? Viable?
Why AI Adoption Often Stalls
| Traditional Approach | The AI-Ready Reality |
|---|---|
| ”AI is an IT project.” | AI adoption challenges live in Sales, Marketing, and Legal. It’s an organizational shift. |
| ”Delivering scope = Value.” | You can deliver every feature on your AI roadmap and still move zero business metrics. |
| ”We need better models.” | You need better context. AI effectiveness is capped by how well you define the problem. |
| ”Agile is for developers.” | AI requires “Agile” thinking at the leadership level: sensing, responding, and pivoting. |
Managing such work like a project often results in nasty surprises - projects don’t deliver on time, and even when they do, they fail to make an impact.
How to Find “AI Gold”
To find “AI Gold,” you have to stop managing deliverables and start managing outcomes. This requires three fundamental shifts in your Operating System:
- Context is the New Code: In the age of AI, Context is the training data. Your OS must provide clear boundaries, defined users, and specific success criteria. The smarter your context, the faster your AI learns.
- Bounded Learning Loops: Instead of a 12-month “AI Roadmap,” you need short “Learning Cycles.” The goal is to gather evidence. If the evidence says “no,” you pivot immediately—minimizing the cost of failure.
- Outcome-Driven Governance: Traditional PMOs focus on schedule. An AI-ready OS asks “Are we moving the needle?” using Leading Indicators and OKRs.
The Hidden Bottleneck: Your Operating System
Think of how you manage strategic projects in your organization:
Crawl - It’s so chaotic, it’s impossible to deliver even a pretty stable project.
Sit - We deliver stable projects through professional project management and accountability (e.g. through a PMO, adopting EOS). High-potential, high-risk projects are still a crapshoot.
Walk - We can deliver high-technical/feasibility-risk projects through effective collaboration and fast feedback loops. Some of these projects provide the promised outcomes, but some of the most strategic ones don’t.
Run - This is the holy grail. We are able to maximize returns and minimize investment even for high-potential high-risk investments by leveraging derisking mechanisms, whether the risk is implementation, need, or total cost of ownership.
When leaders pull me in for advice, I see distinct differences between product-focused initiatives and the rest.
Most organizations have learned they need to walk and run when it comes to their product/tech initiatives. But when it comes to other initiatives, even this taxonomy is foreign to most company-level PMOs and leaders.
What happens when you plug AI into this environment?
Companies that are working to leverage AI in their products are doing reasonably well. There’s plenty to learn and do when it comes to designing, developing, and maintaining AI features, as well as leveraging AI to accelerate product development, but product/tech organizations are at least aware of what they should be aiming for.
This is true when integrating AI capabilities into your internally focused products and systems as well.
The challenge arises when organizations try to leverage AI beyond the boundaries of their traditional IT. It’s these use cases that have high potential but often fail miserably because they’re not managed properly.
Apply Product Thinking to AI
Organizations that do find gold with AI borrow habits from great product teams:
Start with a problem, not a technology.
Ask, “Where would we hire AI to move the needle?” not “Which tool should we buy?”Work in bounded, outcome-driven loops.
Treat AI efforts like products with clear hypotheses, customers, and measures of success.Continuously develop context.
The more clarity you give teams about constraints, users, and success criteria, the smarter their experimentation becomes. Context is the real training data.
At a BioTech firm I worked with, for example, scientific teams used Scrum-like loops to test how AI models could accelerate gene-therapy design. The secret wasn’t better algorithms; it was better empiricism — visible hypotheses, short learning cycles, and shared evidence across disciplines.
Upgrading the Company, Not Just the Code
You don’t need to replace your entire operating system, but you do need to upgrade it for learning speed. Ask yourself:
Are strategic initiatives expressed as outcomes or projects?
Do our scorecards show leading indicators we can steer with, or lagging ones we report after the fact?
How often do we inspect and adapt our bets based on evidence, not hope?
The Real Playbook
The organizations turning AI into business impact share one discipline: agility. Not the capital-A flavor with ceremonies and roles, but the practice of aligning, inspecting, adapting, and flowing work toward outcomes.
They learn faster, and pivot cheaper. They’ve moved from projects to products way beyond the classic boundary of the product/tech organization.
AI isn’t a silver bullet. It’s a forcing function. It exposes how slowly our organizations learn.
Before you invest in more models, invest in an operating system that can think and adapt like one.
Dive deeper in this recent Scrum Community Podcast conversation: Why AI adoption requires an agile operating system.
Frequently Asked Questions
Why do AI pilots stall even when the models are promising?
Because execution systems are often still project-oriented, siloed, and weak on learning loops. The constraint is usually operating model design, not model capability.
What is the first operating-system change to make?
Reframe AI initiatives around outcomes and hypotheses, then run short inspect-and-adapt cycles with cross-functional ownership.
Can a traditional PMO still play a role in AI transformation?
Yes, if it evolves from schedule policing to enabling evidence-based governance, risk management, and decision quality across initiatives.
About Yuval Yeret
Yuval is a rare practitioner who has shaped the agility path of dozens of organizations and influenced the frameworks used across the industry. He helps product and technology leaders move from agile theater to evidence-informed, outcome-oriented delivery that creates better value sooner, safer, and happier.
