SERVICES
Cloud
Enabling Enterprise Agility & Resilience with a Robust Cloud Strategy
Highlights
- Enterprises can no longer rely on a single artificial intelligence (AI) model to meet every business requirement. Different workloads demand different capabilities, cost profiles, governance controls, and performance characteristics.
- Workload-based approach to model selection helps organisations align large language model (LLM) capabilities with business outcomes, risk requirements, and operational constraints.
- Multi-agent architectures powered by specialised models can improve quality, scalability, governance, and cost optimisation while accelerating enterprise AI adoption.
On this page
On this pageinpage
Shift in enterprise AI
Enterprise AI is shifting from model adoption to intelligent model selection.
The first wave of enterprise artificial intelligence (AI) focused on gaining access to powerful foundation models. Today, organisations face a different challenge: determining which model to use for which workload, under what conditions, and within what governance boundaries. Architectural decision-making can be overwhelming as AI leaders continually expand their domain-specific offerings, bringing different strengths in reasoning, cost efficiency, multimodal capabilities, grounding, latency, and tool orchestration. In this scenario, enterprises that can evaluate, select, and orchestrate the right models effectively will be best positioned to scale AI safely and sustainably.
Strategic model selection
Choosing the wrong model can increase cost, risk, complexity, and adoption challenges.
Not all AI workloads are created equal. A customer support classification process has very different requirements from a compliance review, a software development assistant, or a knowledge discovery platform. Models optimised for advanced reasoning may introduce unnecessary cost and latency for high-volume operational tasks, while faster models may not provide the reasoning required for complex decision-making. Some use cases depend on strong structured outputs and tool-calling capabilities, while others require multimodal understanding or long-context processing.
Model selection is no longer a technology choice; it is a strategic business decision as AI becomes embedded in business-critical processes. When selecting models, data residency and governance requirements, security controls, privacy boundaries, and regulatory expectations are important considerations. The decision is made more complex by the rapid pace of AI evolution. New releases continue to bring greater efficiency and more advanced reasoning, coding, multimodal, and tool integration capabilities. Organisations that standardise too early on a single-model strategy risk limiting future flexibility and innovation. However, waiting for a definitive "best model" is not a viable strategy. Enterprises need a framework that supports informed model selection, workload alignment, and governed adoption.
Choosing the right model
Model selection must centre on workload requirements, business objectives, and governance needs.
Many organisations continue searching for a single model capable of supporting every workload. In practice, no single model simultaneously optimises quality, speed, cost efficiency, governance, reasoning depth, and contextual understanding. Different model families are designed with different strengths, making workload fit a more effective decision criterion than model popularity.
Effective model selection begins with the understanding that workload requirements vary for every task. While some use cases demand advanced reasoning, others require low latency, multimodal capabilities, structured outputs, or tool orchestration. Governance obligations, security controls, data residency and regulatory requirements also play a key role when determining suitability. Instead of pursuing universal standardisation, organisations need to adopt a portfolio mindset that treats models as specialised assets for specific business outcomes.
The capability requirements of a workload tend to remain relatively stable even as the underlying AI models continue to evolve. For this reason, enterprises should prioritise capability-based selection and business outcomes over individual model or vendor preferences. The objective is not to identify the best model overall, but the most appropriate model for a specific business context (see Figure 1).
Figure 1: Workload-based model selection
Architecting for agents
As AI systems move from answering questions to performing work, orchestration becomes essential.
The rise of agentic AI is transforming how enterprises design intelligent systems. Business processes increasingly involve planning, research, retrieval, reasoning, execution, validation, and reporting. Relying on a single model to perform all functions efficiently introduces challenges related to cost, governance, and performance.
Agent-based architectures, in which responsibility is distributed among coordinated, specialised agents, offer a solution. A supervisor agent can coordinate workflows and routing decisions; a retriever agent can gather enterprise knowledge and contextual information; and a tool agent can interact with applications, application programming interfaces (APIs), and business systems. Combined with writer agents that generate user-facing outputs and verifier agents that ensure governance, compliance, and quality controls, the result can be seamlessly and autonomously executed workflows.
The approach combines specialisation with control, laying a more reliable foundation for enterprise-scale AI adoption.
Governance at scale
Enterprise trust requires visibility, control, and verification.
As AI becomes more advanced, governance can no longer be an afterthought; it must become a core architectural capability. In building this capability—and the trust that comes with it—organisations will need clear routing policies, model performance benchmarks, approval workflows, auditability, and continuous monitoring to keep execution within expected boundaries. What matters is enterprises' ability to translate these principles into operational practice, because that will determine their ability to adopt and scale AI responsibly.
Practical guardrails can be a good starting point:
- Clear routing policy and fallback mechanism that explain how requests are routed and what happens when a model fails.
- Validation of data and tool interactions to ensure they meet defined formats, requirements, and operating boundaries.
- Checks before outputs trigger action to assess outputs for accuracy, compliance, and potential risk before allowing systems to act.
- Human review of high-risk decisions and the definition of when exactly human review is required.
- Ongoing monitoring of models and agents to assess their performance over time.
These guardrails can help organisations govern and improve their AI operations and reduce the risk of systems straying from bounded action.
A new pattern
A common AI operating model is beginning to emerge across sectors.
Organisations across industries are moving beyond isolated AI initiatives to orchestrated networks of specialised agents spanning multiple business functions. The potential value lies in bringing together domain intelligence to improve decisions, streamline operations, and deliver greater business outcomes.
Retail companies can deploy planning, procurement, customer service, and governance agents to improve forecasting accuracy, operational efficiency, and customer experience. Banks can integrate fraud, risk, compliance, and customer service agents for better decision-making and regulatory compliance. For manufacturers, the opportunity may be in optimising performance and minimising disruptions leveraging engineering, maintenance, quality, and supply chain agents.
These examples show that while the business goals may differ, the underlying model is consistent: specialised models, coordinated agents, intelligent orchestration, and governance come together to create scalable, trustworthy, enterprise-wide AI.
Enterprises that master this model will move beyond fragmented experimentation to make AI a dependable engine of business performance.