Why I Am Writing This
I am writing this from the perspective of someone who has been on the data and AI journey for a long time – from the Hadoop era, through large-scale data platform modernisation, machine learning programmes, cloud data platforms and now enterprise AI. Over that time, I have run teams, delivered platforms and seen both sides of the market: major investments in machine learning that did not deliver the expected value and focused use cases that created clear business impact.
One of the recurring patterns has been fragmentation. Many organisations have approached AI and data transformation through a technology-led lens, often driven by platform replacement, data modernisation, or vendor capability rather than a clear business outcome or strategy. Some investments worked, but many became difficult to scale, govern, or sustain.
AI is creating a similar moment. Technology feels generous because it appears to do almost everything. But when you lift the hood, not every AI solution is equally secure, scalable, observable, cost-effective or production-ready. As a practising architect, I see customers trying to navigate this complexity while also avoiding lock-in, short-term thinking and architectures that cannot adapt as models, costs and regulatory expectations change.
This model has come from practical work across sovereign cloud and enterprise AI environments, including Microsoft Foundry, Snowflake, Databricks, and open-weight model approaches. The point is not that one level is better than another. The point is that organisations need an architecture that helps them choose the right level for the right business outcome, while managing cost, sovereignty, security, operational maturity and long-term sustainability.
To bring clarity to this complexity, I propose a four-tier model for thinking about AI enterprise architecture and the strategic direction organisations should take over the next few years.
The Four Levels of AI Architecture
Level 1: Frontier AI Model APIs
This is where many organisations start: direct access to frontier model APIs such as models from OpenAI, Anthropic’s Claude family and DeepSeek. The value is clear, these services provide the newest capabilities, quicker release cycles and richest feature sets with very little infrastructure effort.
Key benefits include:
- Rapid experimentation and validation
- Low barrier to entry
- Typically token or consumption-based pricing
- Continuous access to the latest capabilities
- Easier productivity gains
The trade-off is commercial and architectural control. Organisations are dependent on vendor-led data terms, pricing changes, product roadmaps and release decisions. Level 1 is powerful for discovery, prototyping and productivity, but it should not be the default pattern for every enterprise workload.
As organisations mature, they often discover that productivity gains alone do not create sustainable enterprise AI capability. Especially as the workforce changes.
Level 2: Cloud-Managed AI Model Services
The second level is where models are consumed through an enterprise public-cloud and AI control plane, such as Microsoft Foundry, Snowflake, Databricks, SAP or Amazon Bedrock. This pattern improves enterprise readiness because identity, governance, security policies and commercial management can sit around the model consumption layer. One potential advantage is tighter integration with data already governed within the platform. Instead of users manually supplying context, AI can access information already residing within enterprise systems.
Benefits include:
- Existing security controls
- Integrated experience with cloud, data and application tooling
- Reduced implementation effort
- Better governance
- Quicker use case build cycle
Level 2 is often where organisations begin to productionise AI because it connects more naturally into existing cloud, data and application estates. Pricing remains influenced by model and platform economics and varies by deployment model. Features or releases can be constrained by the cloud platform’s packaging, roadmap and regional availability.
Level 3: Hosted Commercial AI Models
Level 3 is about using commercial models in a hosted pattern, deployed in cloud, private cloud or on-premises environments. Examples include models and stacks from providers such as Mistral AI, Cohere and Red Hat (multi-AI LLM hosting framework). This level gives organisations greater control over deployment, security posture, sovereignty and commercial arrangements than purely API-led consumption.
The value is stronger isolation, clearer enterprise support, more predictable operating patterns and better alignment to regulated or sensitive workloads. The trade-off is slower access to release cycles, narrower capability and procurement or licensing arrangements that may require enterprise software commitments.
Where Level 3 is useful:
- Regulated use cases that require enterprise support, controlled release cycles and clear accountability.
- Workloads where data residency, private deployment or network control are important.
- Organisations that want more sovereignty and commercial control without taking on full responsibility for operating open-weight models themselves.
Organisations still need to invest in AIOps as part of level 3 to manage the workflow, model and infrastructure.
Level 4: Open-Weight AI Models
Level 4 is the most controlled pattern: self-managed model weights deployed on organisation-controlled infrastructure. This can include open-weight models such as Kimi and Qwen, operated through the organisation’s preferred cloud, GPU platform, private infrastructure or sovereign environment.
The benefits are:
- Maximum sovereignty, lower to no model license cost, full runtime control and the ability to optimise models for specific workloads.
The trade-offs are:
- Infrastructure, security operations, model evaluation and AIOps responsibility sit with the organisation.
Best fit for strategic AI capabilities where long-term ownership, cost control, portability and regulatory confidence matter most.
Together, Levels 3 and 4 represent the shift from consuming AI capability to operating AI capability. They require more mature engineering practices, including DataOps, MLOps, LLMOps and AIOps, but they provide a clearer pathway to secure, governed and production-grade enterprise AI.
Final Thoughts
The future of enterprise AI architecture is not about committing to one model, platform or vendor. It is about having enough flexibility to choose the right AI pattern for each business outcome.
Most organisations will likely use all four levels over time. Level 1 supports productivity and experimentation. Level 2 helps move AI into production workflows. Levels 3 and 4 are better suited to use cases that require ownership, sovereignty and long-term cost control. AIOps will become a necessary foundation for managing security, intellectual property, operational risk and enterprise value.
The organisations that succeed will not simply be the ones that adopt the most AI. They will be the ones that build architectures that are outcome-led, data-first, modular, secure, observable and production-ready. AI will keep changing, but these foundations will determine whether organisations can scale value sustainably.
