Enterprise AI is moving beyond experimentation. As organizations put AI into production and expand it across the business, the focus is shifting toward scale and measurable business impact. In this environment, leaders need to know whether the AI model solves the right problem, fits into existing processes and can operate economically at enterprise scale.

Our latest research, The AI Impact Imperatives, 2026, illustrates the challenge. The report found that 43% of major AI projects initiated over the next 24 months are expected to fail. At the same time, the median expected payback period for major AI initiatives is roughly 18 months.

“The biggest difference between successful AI initiatives and those that fail is having a clear business objective from the beginning,” says Arunachalam Jayaraman, Vice President and Solution Principal at HCLTech.

“Many AI prototypes and proofs of concept perform well in a controlled environment but fail to deliver measurable business value because they are technology-driven rather than business-driven.”

That challenge extends beyond choosing the right use case. AI models depend on the quality of the data available to them and on their ability to work within the systems and processes an organization already uses. A model that performs well technically can still struggle in production if the underlying data is unreliable or the solution cannot connect effectively with enterprise workflows.

The research reflects these broader execution challenges. Among organizations surveyed, 40% cite cross-functional coordination as a barrier to AI adoption, 39% struggle to align AI initiatives with business strategy and 34% report difficulty establishing appropriate success metrics.

Choosing the right model for the task

The economics of enterprise AI are also changing how organizations think about model choice.

Large language models (LLMs) have an important role where broad knowledge and complex reasoning are required. But many enterprise workloads are narrower, repeatable and specific to an industry, function or organization. This is creating a stronger case for small language models (SLMs).

Jayaraman identifies three advantages: token economics, data privacy and differentiation.

“Large language models consume a significant number of tokens, making high-volume enterprise AI applications expensive and less cost-effective,” he says. “Smaller, domain-specific models can deliver similar performance for targeted use cases at a much lower cost.”

Keeping models and workloads within an organization’s own environment can also provide greater control over sensitive customer and business data. This can be particularly important where privacy, security or regulatory requirements constrain how information can be processed.

There is also a strategic dimension. An organization that fine-tunes models using proprietary data, processes and domain knowledge can create AI capabilities that reflect its own intellectual property rather than relying exclusively on capabilities available to the wider market.

“In my view, SLMs provide the right balance of cost, performance, security and business differentiation for most enterprise use cases,” says Jayaraman.

The case becomes more significant as AI consumption grows. Our research finds that 22% of organizations already cite a lack of transparent cost monitoring and unpredictable consumption rates as a technical challenge limiting AI adoption. Integration between multiple AI systems is cited by 30%, while 29% point to data quality, quantity and accessibility.

An enterprise will need more than one model

The choice between SLMs and LLMs does not have to be binary.

Jayaraman sees SLMs operating as one component within a broader AI environment that also includes larger models, AI agents and orchestration technologies. Specialized requests can be handled by smaller models, while more complex tasks can be directed toward models with greater reasoning capability.

The ability to make that decision dynamically becomes important as the number of models and use cases increases.

“The key enabler is an Intelligent Model Router, which dynamically selects the most appropriate model based on the complexity of the request, business context, data sensitivity and cost considerations,” says Jayaraman.

This allows enterprises to optimize model selection without requiring employees or customers to decide which underlying technology should handle each request. Combined with agents and orchestration layers, the architecture can route work according to what the task requires while maintaining a consistent user experience.

It can also help avoid designing an entirely new AI stack for every use case. For Jayaraman, scalability means creating an architecture capable of supporting multiple applications without significant rework or new operational bottlenecks.

That architecture still must fit into the wider technology estate. AI cannot operate independently of enterprise applications, data platforms, security controls and existing infrastructure.

This remains a significant issue. HCLTech’s research finds that only 21% of respondents consider their data estate modernized and operational, while organizations continue to report difficulties with data orchestration, intelligent data management and platform consolidation.

Build governance into the architecture

The same principle applies to governance. Security, privacy, compliance and Responsible AI cannot be treated as separate activities that begin once a model reaches production.

Jayaraman argues that these requirements need to be incorporated into the architecture itself, alongside integration with the organization’s existing technology ecosystem.

The research supports the importance of that foundation. Of the organizations surveyed, 90% say data modernization is important to operating compliant and Responsible AI, while 87% say it is important to improving model accuracy and reliability.

As enterprises introduce more models and increasingly autonomous agents, governance will also need to become continuous. Organizations need visibility into how models are performing, what they cost, how they are accessing data and whether their behavior remains within established security and compliance boundaries.

For leaders making AI architecture decisions today, Jayaraman’s argument is ultimately about discipline rather than scale for its own sake.

“Successful AI leaders will focus on creating sustainable enterprise AI capabilities that deliver long-term business value, not just impressive demonstrations of technology,” he says.

That means choosing models according to the problem, designing for reuse and scale, integrating AI into existing enterprise processes and putting governance around the entire environment. As AI becomes part of business-critical operations, those architectural decisions will increasingly determine whether experimentation turns into sustainable value.