How to choose the right AI approach in Azure

The decision is not simply SaaS versus custom development. It is about choosing the approach that best fits your technical and business requirements.

1. Ready-to-use Copilots (SaaS Agents) 

Microsoft Copilots are ready-to-use SaaS solutions designed to improve productivity with limited setup. Microsoft 365 Copilot adds AI capabilities across familiar Microsoft applications. Other in-product and role-based Copilots support specific business functions and industries. 

This includes Microsoft 365 Copilot, role-based Copilots and Copilots built into products such as Dynamics 365, Power BI and GitHub. 

Choose a ready-to-use Copilot when: 

• The use case is already supported by an existing product 
• You need a fast implementation 
• Standard functionality is sufficient 
• Your organisation has limited AI development capacity 
• You don’t want to manage infrastructure 
• A licence-based cost model is suitable 

The main advantage is speed. Teams can start using AI without building models, APIs or hosting environments. 

Trade-off: Copilots deliver the fastest results but offer less flexibility and control than a custom AI solution.

Tradeoff: Low-code SaaS development opens AI development to more teams, but complex customisation may require a different approach. It also still has limits. As requirements become more technical, regulated or specialised, you may need to move towards a managed Azure platform.

3. Managed PaaS development  

Platform as a service (PaaS) is often the starting point for custom AI applications and agents. Choose managed PaaS development when low-code tools don’t offer enough control, but you still want Microsoft to manage the underlying platform. 

Microsoft Foundry provides a managed environment for building agents, RAG applications and model-based solutions. Developers can use APIs, SDKs and platform tools without managing the full infrastructure stack. Other managed services may include Azure Machine Learning and Microsoft Fabric for machine learning and data workflows. 

Choose PaaS when: 

• You need custom agents or applications 
• The solution uses private or specialised data 
• You need control over model selection 
• you want to use RAG 
• Evaluation and observability are important 
• You may need fine-tuning, enterprise networking, policies and access controls 
• You don’t want to manage servers or clusters directly 

A managed platform gives development teams more control over models, data orchestration and application logic. It brings agents, models and tools into one managed environment, with built-in monitoring, evaluations, access controls and policies. 

For example, a fraud detection agent may require more control than Copilot Studio can provide. The team may need to select specific models, evaluate accuracy, manage sensitive data and adjust the application based on risk.  

Trade-off: While managed PaaS offers greater flexibility and control than low-code SaaS, it requires stronger engineering and Azure expertise. 

4) AI on Azure infrastructure (IaaS) 

Infrastructure as a service (IaaS) provides the highest level of control over your models and runtimes. Depending on your requirements, you can use Azure Container Apps (ACA), Azure Kubernetes Service (AKS), virtual machines or a combination of these services.

It typically takes the longest to build and requires the most effort to maintain over time.

Opt for IaaS when: 

• You need to host your own model, or managed platforms cannot meet your requirements 
• You have strict regulatory or privacy requirements 
• Latency must be controlled closely or you need a custom runtime 
• The architecture requires specialised infrastructure 
• You need detailed control over networking, security or scaling 

For example, a product safety and compliance system may need detailed control over data handling, access, model hosting and auditability. IaaS gives you that fine-grained control. 

Trade-off: Azure infrastructure offers the most control, but it requires the most time, skill and ongoing operational effort. 

Azure AI decision framework 

Some solutions may also move between approaches over time. For example, a sales and inventory agent could begin as a low-code solution, then move to PaaS when the organisation needs more control over data analysis, models or integrations. 

Also, the architecture does not have to remain fixed forever.
 
 

Microsoft AI decision tree 

Use the decision tree in the following section to narrow between the options of selecting the right AI solution in Azure: 

What to consider before choosing your AI solution 

Use the following guidance to weigh four factors for AI solution: 

• Capabilities: Review the capabilities of Microsoft and Azure AI solutions to see if they meet the needs of your use case. 
• Data needed: Confirm the required data exists and is accessible for the scenario. 
• Skills required: Validate that each use case is achievable with current capabilities before you select a solution. 
• Cost: Assess whether the solution fits your budget 

1. Capabilities

Start with what the solution needs to do. Review the capabilities of Microsoft and Azure AI services and compare them with your use case.

Ask yourself:

• Does an existing Copilot already support this scenario?
• Can Copilot Studio handle the workflow?
• Do we need custom models or RAG?
• Do we need to fine-tune a model?
• Are standard integrations sufficient?
• Do we need custom tools or APIs?
• Do we need full control over the infrastructure?

Avoid choosing a platform based on popularity alone. Start with the business and technical requirements.

2. Data

AI depends on accessible, relevant and usable data. Your data strategy determines whether your priority use cases have governed, high-quality data to work with.

Confirm which data is needed, where it is stored, whether it can be accessed, whether it is sensitive and which retention and compliance requirements apply. Also determine whether the data requires retrieval, indexing or transformation.

• A simple productivity use case may work with data already available through Microsoft 365.
• A custom RAG application may require document ingestion, chunking, indexing, embeddings and access controls.
• A regulated system may require custom storage, networking and infrastructure.

The data requirements can quickly narrow your technology choices.

3. Skills

Validate that the solution is achievable with your current capabilities. Low-code tools may suit business users and developers working together. Managed PaaS development requires software engineering, Azure and AI application skills.

Infrastructure-based AI may also require platform engineers, data engineers, security specialists and machine learning expertise.

Ask yourself:

• Can our current team build this?
• Can we operate it after launch?
• Do we understand the security requirements?
• Can we monitor and troubleshoot the solution?
• Do we need training, additional expertise or external support?

Choosing a highly customised approach without the required skills increases costs and delays delivery.

4. Cost

Consider both the initial and ongoing costs, including licences, model consumption, tokens, compute, storage, data retrieval, integration, monitoring, security, testing and training.

• SaaS options often provide more predictable costs and require less operational effort.
• PaaS costs may depend on model, token, storage and compute consumption.
• Infrastructure provides greater technical control, but your organisation is responsible for operating the environment.

The lowest licence price does not always result in the lowest total cost. Include development and operational ownership in your comparison.

Closing thoughts 

Choosing the right AI approach in Azure starts with the use case.

• Ready-to-use Copilots offer speed
• Low-code platforms add flexibility
• Managed PaaS services provide more control
• Azure infrastructure supports the most demanding custom or regulated scenarios

The right choice depends on your business and technical requirements, including data, skills, security, cost and operational ownership. 

Remember, choosing the technology is only one part of AI adoption. You also need to prepare the organisation, validate the approach through a focused proof of concept (POC), and confirm that your team can support the solution responsibly.