By Vanguard Enterprise Intelligence Unit with the work of Erik Brynjolfsson, Andrew McAfee, Andrew Ng, Fei-Fei Li, and Eric Horvitz.

The first wave of enterprise artificial intelligence was defined by access to models. The next wave will be defined by the ability to turn models into durable economic value. That shift sounds simple, but it is proving to be the central management problem of 2026. Companies have bought tools, launched pilots, tested copilots, experimented with agents, and encouraged employees to use generative AI. Yet many still struggle to show meaningful enterprise-level returns.

This is not because AI lacks power. The models are improving rapidly. Coding, reasoning, multimodal analysis, search, synthesis, and automation capabilities continue to advance. Open-source models are becoming more competitive. Cloud providers are building larger AI infrastructure footprints. Governments are investing in sovereign AI capabilities. Enterprises are experimenting with AI factories, agentic workflows, model operations, retrieval systems, and internal platforms. The AI ecosystem is expanding quickly.

The problem is that many corporate AI strategies remain model-centered rather than value-centered. They begin with the question, “What can this model do?” instead of “What economic system are we trying to improve?” That distinction explains why so many initiatives stall. A powerful model attached to a weak workflow creates limited value. A brilliant pilot without production infrastructure becomes a demonstration. A chatbot without trusted data becomes a novelty. An agent without orchestration becomes a risk.

The companies that will lead the next phase of AI will not be the companies with the most public enthusiasm. They will be the companies that understand AI as an enterprise resource: a governed, measured, scalable capability that requires infrastructure, context, operations, security, and organizational redesign. The strategic question is no longer whether the model is impressive. It is whether the organization is built to convert intelligence into performance.

The Pilot Ceiling

Most companies do not fail at AI because they cannot produce a promising pilot. They fail because pilots are structurally different from production systems. A pilot can be narrow, manually supervised, lightly integrated, and supported by a small expert team. A production system must be reliable, secure, observable, cost-controlled, integrated with enterprise data, aligned with workflows, and accountable to business outcomes.

This gap is underestimated. Executives often see an impressive prototype and assume the organization is close to value realization. In reality, the prototype may represent only the easiest part of the journey. The hard work begins when the system must function across real users, messy data, legacy platforms, compliance constraints, vendor dependencies, security requirements, and changing business conditions.

The pilot ceiling appears when experimentation outpaces institutional capability. Teams launch use cases faster than the organization can support them. Business units adopt tools before data architecture is ready. Leaders approve agents before identity, access, and monitoring systems are mature. Employees generate output faster than managers can establish quality standards. The organization creates motion, but not infrastructure.

This is why many AI programs feel busy but underwhelming. They produce demos, dashboards, internal excitement, and scattered productivity gains, but they do not change the economics of the business. They remain trapped between possibility and performance.

The Infrastructure Turn

The next phase of AI is less glamorous than the first, but more important. It is the infrastructure phase. Companies are discovering that AI is not simply software. At scale, it is an operating environment that depends on compute, data pipelines, model governance, security architecture, orchestration layers, evaluation systems, cost management, and human oversight.

This is the logic behind the rise of the “AI factory.” The phrase is useful because it reframes AI from a collection of tools into a production system. A factory does not exist to admire machinery. It exists to convert inputs into outputs repeatedly, reliably, and economically. An AI factory must do the same. It must convert data, models, prompts, context, workflows, and human judgment into measurable business outcomes.

For many firms, this requires a major shift in thinking. In the pilot phase, AI can be treated as an innovation project. In the scaling phase, AI must be treated as infrastructure. That means leaders must ask different questions. Where does the data come from? Who owns it? Which systems can the model access? How are outputs evaluated? What happens when the model is wrong? How are costs monitored? What level of latency is acceptable? Which tasks require human approval? Which vendors are strategic dependencies? Which capabilities should be built internally?

These are not technical details. They are strategic decisions. A company’s AI infrastructure determines what it can automate, how fast it can experiment, how safely it can scale, and how much control it retains over its future.

The Context Problem

One reason AI strategies fail is that leaders overestimate the value of the model and underestimate the value of context. General intelligence is impressive, but enterprise value depends on specific intelligence. A model must understand the company’s products, customers, pricing logic, policies, workflows, risk standards, historical decisions, and operating language. Without that context, AI produces generic output. It may sound polished, but it does not move the business.

This is why context engineering is becoming a core enterprise discipline. It is not enough to prompt a model well. Companies must design the information environment around the model. That includes retrieval systems, knowledge graphs, structured data access, approved documents, business rules, user permissions, workflow memory, and feedback loops. The goal is to make AI systems aware enough of the enterprise to be useful without giving them uncontrolled access to everything.

The context problem explains why many AI tools disappoint after early enthusiasm. A public model can draft a memo. It cannot automatically understand the company’s margin structure, customer history, regulatory exposure, or internal decision logic. An agent can execute a task. It cannot safely execute the right task unless the organization has defined the right information, permissions, thresholds, and escalation paths.

The winners will not simply use better models. They will build better context systems around the models they use.

Generative AI Is Not Enough

The public conversation often treats generative AI as if it replaced the rest of machine learning. That is a mistake. The most valuable enterprise systems will combine generative AI with predictive models, optimization methods, rules engines, simulation, search, and traditional analytics. Generative AI is powerful at language, synthesis, interface design, and reasoning-like interaction. Predictive machine learning remains essential for forecasting, classification, risk scoring, demand planning, pricing, fraud detection, personalization, and operational prediction.

The strategic opportunity is convergence. A sales system might use predictive models to identify accounts likely to convert, generative AI to prepare tailored outreach, retrieval systems to surface account context, and workflow automation to schedule follow-up. A supply-chain system might use predictive models to detect disruption risk, generative AI to summarize implications, optimization tools to propose alternatives, and agents to initiate approved actions. A finance system might use anomaly detection to identify irregularities, generative AI to draft explanations, and human review to approve final treatment.

This convergence is where sustained ROI becomes more plausible. Generative AI alone often improves individual productivity. Integrated AI systems can improve the economics of whole processes. The difference is the difference between a smarter worker and a smarter operating model.

Many companies remain stuck because they isolate AI capabilities. The data science team builds predictive models. The innovation team tests generative tools. The IT team manages platforms. The business unit owns workflows. The risk team reviews compliance. Value leaks between these functions. Top performers will connect them into a single operating discipline.

Open Source and Strategic Control

Open-source AI has become a strategic force because it changes the economics and control dynamics of enterprise adoption. Companies no longer face a simple choice between building frontier models internally and relying entirely on closed external systems. Open models, smaller specialized models, fine-tuning methods, retrieval architectures, and domain-specific deployments give enterprises more flexibility.

This flexibility matters. Some companies need lower cost. Some need data control. Some need customization. Some need latency improvements. Some need to deploy models in restricted environments. Some need to avoid excessive dependency on a single vendor. Open-source momentum gives leaders more architectural options, but it also increases complexity. More choice requires better technical judgment.

The point is not that open source is always superior. Closed frontier models may still offer stronger performance in many use cases. The strategic issue is portfolio design. Which workloads require frontier capability? Which can be handled by smaller models? Which should run in the cloud? Which should run in a private environment? Which models should be fine-tuned? Which should be accessed through retrieval? Which vendors create acceptable dependency? Which capabilities are too strategic to outsource entirely?

This is where AI strategy begins to resemble capital allocation. Leaders must decide where to spend for performance, where to optimize for cost, where to preserve control, and where to maintain optionality. The companies that make these decisions deliberately will have an advantage over those that simply adopt whatever tool is most visible in the market.

Sovereign AI and the New Geography of Intelligence

AI is also becoming geopolitical infrastructure. Sovereign AI reflects a growing recognition that control over models, data, compute, chips, cloud regions, and energy capacity has national and regional significance. Governments increasingly view AI capability as part of economic competitiveness, public-sector modernization, defense readiness, and technological independence.

For corporations, this matters because AI strategy will be shaped by geography as much as software. Data residency rules, sector-specific regulations, supply-chain exposure, chip availability, cloud concentration, energy constraints, and regional model ecosystems will all affect enterprise AI decisions. A multinational company cannot assume that one AI architecture will fit every market. It may need different deployment models for different jurisdictions, customers, and risk environments.

Sovereign AI also forces executives to reconsider dependency. If AI becomes central to operations, then compute access, vendor terms, model availability, and regulatory permissions become strategic vulnerabilities. A company that depends heavily on a single provider, region, or infrastructure pathway may find itself exposed in ways that were not obvious during the pilot phase.

This does not mean every company must build its own sovereign AI stack. Few can. But every serious company needs a sovereignty analysis. It should know which data must remain controlled, which workloads are mission-critical, which vendors are systemic dependencies, and which jurisdictions create operational constraints. AI value depends not only on what the system can do, but on whether the company can continue to use it under pressure.

MLOps Maturity and the Discipline of Production

The move from models to value requires operational maturity. MLOps, ModelOps, and emerging LLMOps practices are not administrative burdens. They are the production discipline of AI. They define how models are deployed, monitored, updated, evaluated, secured, and retired. Without these capabilities, companies cannot reliably scale AI beyond isolated use cases.

Traditional software can be tested against relatively stable requirements. AI systems are more dynamic. Their outputs can vary. Their performance can drift. Their behavior can change when data changes, prompts change, vendors update models, users adapt, or workflows evolve. Generative systems add further complexity because fluency can conceal error. A model can be wrong in a highly convincing way.

Production AI therefore requires continuous evaluation. Leaders need to know whether systems are accurate, useful, biased, secure, cost-effective, and aligned with business objectives. They need incident response procedures when AI fails. They need version control for prompts, models, datasets, and retrieval sources. They need human review thresholds. They need audit trails. They need cost visibility at the use-case level.

The firms that lack this discipline will continue to stall. They may produce exciting prototypes, but they will struggle to build trust. The firms that build operational maturity will be able to scale faster because they can see, measure, and govern what their systems are doing.

The Economics of Compute

AI strategy increasingly depends on compute economics. Leaders who once viewed infrastructure as a background technology issue must now understand its strategic implications. Training, fine-tuning, retrieval, inference, agent orchestration, and real-time automation all carry cost, latency, and capacity considerations. At scale, inference costs can become material. Poor architecture can turn a promising use case into an expensive habit.

This is one reason AI ROI is often elusive. A pilot may appear inexpensive because usage is limited. Production deployment changes the equation. More users, larger context windows, higher query volume, more agents, more retrieval calls, and more model interactions can drive costs upward quickly. A system that saves employee time may still fail economically if compute costs are not managed against value created.

This requires financial discipline. AI programs should have unit economics. Leaders should know the cost per interaction, cost per workflow, cost per resolved ticket, cost per qualified lead, cost per generated recommendation, and cost per automated decision. They should compare those costs against labor savings, revenue lift, risk reduction, customer retention, and speed improvements.

AI cannot be managed only as innovation spend. It must be managed as operating capital.

The Roadmap to Value

The companies that fix failing AI strategies will begin by shifting the unit of analysis from models to business systems. They will identify the processes where intelligence is economically valuable: sales conversion, customer retention, service resolution, fraud prevention, supply-chain resilience, product development, software delivery, compliance review, pricing, forecasting, and decision support.

They will then build a portfolio of use cases instead of a collection of experiments. A portfolio has priorities, owners, metrics, dependencies, and funding logic. It distinguishes between quick wins, strategic platforms, risky bets, and foundational investments. It prevents the organization from confusing scattered activity with strategy.

Next, they will create an AI operating architecture. This architecture will define model access, data access, retrieval systems, orchestration tools, governance controls, evaluation methods, security standards, human review processes, and cost management. It will make clear which capabilities are centralized, which are embedded in business units, and which are provided by external partners.

They will also develop context engineering as a formal capability. The organization’s proprietary advantage will not come only from using general models. It will come from connecting those models to trusted enterprise context. This means investing in knowledge systems, data quality, metadata, permissions, process documentation, and feedback loops.

Finally, they will measure value at the level of business outcomes. The question should not be how many employees use AI. It should be whether AI improves conversion, reduces cycle time, raises quality, lowers cost, decreases risk, increases retention, or accelerates learning. Adoption is not the outcome. Performance is.

The Executive Test

AI is moving from spectacle to infrastructure. That transition will disappoint companies that expected models alone to create competitive advantage. It will reward companies that build the management systems required to turn intelligence into value.

The executive test is whether leaders can make AI boring enough to scale. That does not mean stripping it of ambition. It means giving it the discipline that every serious enterprise capability requires: architecture, ownership, security, measurement, governance, economics, and continuous improvement.

The firms that fail will continue to chase models, tools, and announcements. They will buy new capabilities before absorbing the old ones. They will confuse experimentation with transformation. They will treat AI as a technology race when the real competition is an operating-model race.

The firms that succeed will do something harder. They will build AI factories rather than AI showcases. They will converge generative and predictive intelligence. They will treat compute as a strategic resource. They will govern context as carefully as they govern capital. They will design systems that are secure, scalable, measurable, and adaptable.

The next advantage in AI will not come from having access to intelligence. Access is spreading. The next advantage will come from converting intelligence into differentiated capability faster, safer, and more economically than competitors.

That is why AI strategy must now move from models to value. The model may be the engine, but the enterprise is the machine. If the machine is poorly designed, even the most powerful engine will waste its force. If the machine is built well, intelligence becomes more than a tool. It becomes a durable source of competitive power.