1. Start with the business problem, not the AI tool.

That really comes down to the fact that many AI initiatives start with the wrong question: "What can we do with AI?" That question certainly triggers innovation, but it often leaves teams mesmerized by technically interesting (but not strategically significant) use cases.

A more relevant question is, “What business problem is important enough to be solved with AI?"

This shift forces leaders to define value before choosing technology. AI must be tied to tangible outcomes like revenue growth, cost reductions, customer satisfaction, operational efficiency, risk avoidance or rapid decision making. If you cannot correlate an AI project with a real business KPI, that goal should not be tried.

2. Data is the core of AI trust.

AI systems are only as good as the data behind them. Even the most sophisticated AI model will be hampered by fragmented, inconsistent or poorly governed data.

Unfortunately, most organizations are still stuck with disparate data scattered across legacy systems, business units, applications and manual processes. Definitions vary by department. Quality issues are discovered late. Ownership is unclear. When companies scale AI, these gaps transition from slight problems to major roadblocks.

The quality of data becomes even more paramount for generative AI and agentic AI. A connected chatbot with outdated knowledge will come up with unreliable answers. For example, an AI agent wired to poorly connected systems may make the wrong move. Predictive models trained on biased or incomplete data may perpetuate bad decisions.

Enterprise AI is reliant on trustworthy data platforms, automated processes, quality checks, lineage, metadata, cataloging and ownership. AI strategy—like data modernization—cannot be siloed.

3. Governance before AI creates risk.

Sure, speed is a big deal in AI, but we need to manage that speed, and doing it well creates risk. Use AI to make hiring, lending, legal review, customer service and other operational decisions: This is how you use AI systems for fraud detection and claims processing. The results of these systems can be significantly negative when they are biased, inaccurate or poorly monitored.

Governance is not enough in the post-deployment phase. It should be part of the AI lifecycle from the start.

The organization must be equipped with responsible AI principles, risk classifications, human oversight/audit trail requirements, explainability standards, bias tests, model monitoring and escalation processes. Leaders must also define accountability. Somebody must own model behavior, performance, risk and remediation.

Innovation is not the enemy of governance. Governance is what allows for scaling AI safely.

4. Define how AI work gets done across the enterprise.

An untold number of companies hire AI talent and procure AI tools, but never define how to organize work around AI. It adds to the confusion for business, data science, engineering, legal, security and operations teams.

Who owns the AI roadmap? Who prioritizes use cases? Who manages production models? Who approves of risk controls? Who ensures business adoption? Who should be held liable when an AI system fails?

When there are no clear answers, AI programs become pieces. They only get solutions that data science teams build but business teams do not adopt. Business teams take a whack at using the tools with insufficient technical oversight. Engineering teams build platforms without governance. Legal and compliance teams, too, come in far too late.

While some MNCs have a need for platform technology, governance and standards that can scale across the organization (a central AI function), others would prefer to use more of an embedded team approach as they understand what domain-specific business needs are. The actual structure varies, but the principle that AI requires ownership remains obvious.

5. Evolve out of pilot programs into production capabilities.

Most organizations have the capability to put together an AI proof of concept. Significantly fewer could extend production into AI.

Manual effort, few users and thin data are enough for a pilot but not for enterprise AI. Scaling needs production-ready infrastructure, MLOps, monitoring, retraining, security and reusable components, as well as workflow integration and change management.

This is where a great deal of AI strategies are failing. The demo was functional, but the business never implemented it. The model is used to test to some degree, but it is not built into the daily workflow. The pilot thrills, but no plans are in place for maintenance, governance or long-term ownership. A model working in a lab does not create AI value. When AI starts altering the nature of work throughout the enterprise, it is created.

6. Measure AI success by business capabilities transformed.

Companies that merely adopt the latest tools will not win the next phase in AI competition. These tools will be made readily available. The true edge will be with organizations that work the discipline around AI.

Instead of measuring the number of AI pilots launched, leaders should stop and think. Rather than success defined by incremental delivery, they should measure the number of business capabilities transformed.

AI is not a technology-related initiative. It is a leadership test. Ultimately, the companies that win will be those who connect AI to strategy, build strong foundations in data, govern responsibly, define ownership and design for scale from the start. That is the difference between AI adoption and impact.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?