Generative Artificial Intelligence, or Gen AI, has exploded in use these past two years. And as it continues to evolve, organizations seeking to integrate AI capabilities into their workflows must make key decisions about how much control they want to have over their AI models.
The Generative AI Scoping Matrix serves as a framework for understanding the different levels of ownership and complexity involved in working with generative AI. In this post, I will explore five primary categories within the scoping matrix, and examine the level of ownership associated with each:
1. Public Generative AI Services
Public generative AI services include platforms like OpenAI’s ChatGPT, Google Gemini, and Anthropic’s Claude. These services allow users to interact with AI models without requiring any technical expertise or infrastructure management.
Level of Ownership
- Control: Minimal. Users rely entirely on third-party providers.
- Customization: Very limited. Users can interact with the model but cannot fine-tune it beyond prompt engineering.
- Infrastructure & Maintenance: None. The service provider manages everything.
- Cost: Typically usage-based, with free and paid tiers.
Ideal For
- Individuals and businesses that need AI-powered text or image generation quickly.
- Users who want to avoid infrastructure management.
- Organizations with limited AI expertise.
2. An Application with Generative AI Features
Some companies embed generative AI within their applications, but offer a more tailored experience. For example, Microsoft Copilot integrates AI-driven features directly into Office applications.
Level of Ownership
- Control: Moderate. The company provides some level of customization for its users but still relies on third-party APIs.
- Customization: Limited to predefined settings and API calls.
- Infrastructure & Maintenance: Managed by the app provider.
- Cost: Subscription-based or included in enterprise software pricing.
Ideal For
- Businesses that want to integrate AI into their existing tools without developing custom models.
- Organizations that need AI capabilities with minimal setup effort.
3. Using a Pre-Trained Model
This is the category I’ve been most involved in over the past year. This is where developers can integrate open-source or proprietary pre-trained models into their own applications using services like Amazon’s AWS Bedrock, and Microsoft’s Azure Open AI. While the cost of the AI itself can be reasonable, there is additional overhead in managing the AI that should be considered.
Level of Ownership
- Control: Moderate. Users can modify how the model is applied but cannot alter the core architecture.
- Customization: Limited to prompt tuning, API settings, and implementation-specific adjustments.
- Infrastructure & Maintenance: Requires some computational resources if running locally; otherwise, managed by API providers.
- Cost: Varies; some models are free, while API usage incurs costs. Also requires more specialized resources who understand core AI concepts like prompt engineering, vector embedding, model parameters like temperature and Top P, and more.
Ideal For
- Developers who need powerful AI capabilities without the cost and complexity of training a model.
- Businesses seeking more control than public services allow but without the burden of training and fine-tuning.
4. Fine-Tuning a Model on Your Own Data
Fine-tuning involves taking an existing AI model and adapting it to specific tasks using proprietary datasets. This allows organizations to improve the model’s performance on domain-specific applications while leveraging pre-trained capabilities.
Level of Ownership
- Control: High. Users can refine model behavior but still depend on the base model.
- Customization: Significant. Organizations can train models to better align with their specific needs.
- Infrastructure & Maintenance: Requires access to computing resources, such as GPUs or cloud-based AI platforms.
- Cost: Higher than using a pre-trained model due to fine-tuning computational costs and data preparation.
Ideal For
- Enterprises and researchers who need domain-specific performance improvements.
- Organizations with proprietary datasets requiring fine-tuned model adaptations.
5. Training a Model from Scratch with Your Own Data
Organizations that require maximum control and customization may choose to train their own generative AI models from the ground up. This approach requires large datasets, significant computing power, and deep expertise in AI model training.
Level of Ownership
- Control: Complete. Organizations have full control over architecture, training data, and fine-tuning.
- Customization: Unlimited. Every aspect of the model can be tailored.
- Infrastructure & Maintenance: Requires substantial investment in computing resources, data engineering, and ongoing maintenance.
- Cost: Very high due to infrastructure needs, training time, and expertise requirements.
Ideal For
- Companies with highly specialized AI needs that cannot be met by existing models.
- Research institutions developing new generative AI architectures.
- Organizations concerned about data privacy and ownership.
The Generative AI Scoping Matrix
| Category | Control | Customization | Infrastructure & Maintenance | Cost |
| Public Generative AI Services | Low | Minimal | None | Low (subscription-based) |
| App with Gen AI Features | Moderate | Limited | None | Subscription-based |
| Using a Pre-Trained Model | Moderate | Some (via prompts and API settings) | Some (if self-hosted) | Varies |
| Fine-Tuning a Model | High | Significant | Requires compute resources | High |
| Training a Model from Scratch | Complete | Unlimited | Requires full infrastructure | Very high |
Consider the trade-offs before choosing
The Generative AI Scoping Matrix provides a clear guide to understanding the trade-offs between control, customization, infrastructure requirements, and cost. The decision on which approach to take depends on the specific needs of an organization, its technical capabilities, and the level of AI integration required.
For most businesses, leveraging pre-trained models or fine-tuning existing models offers the best balance between capability and complexity. However, enterprises with proprietary requirements and substantial resources may choose to train their own models for maximum control.
By understanding these different levels of ownership, organizations can make informed decisions about how to integrate generative AI effectively into their operations.