- The Augmented Advantage
- Posts
- AI For Business: Comparing the Custom, Prebuilt, and Fine-tuned Approach
AI For Business: Comparing the Custom, Prebuilt, and Fine-tuned Approach
Pick the option that fits your needs
Tobias Zwingmann
March 10, 2023
Read time: 6 minutes
Hey there,
When you want to integrate AI into your business, you have 3 options:
Build AI from scratch
Use AI-as-a-Service
Fine-tune an existing model
The reason I'm highlighting this today is because many people don't realize that there are multiple approaches to integrating AI into their products, services, or workflows.
They just go with the one they know.
However, once you understand all the options available, you can make an informed decision and pick the one that fits your goals.
Let's dive in!
Want to level up your analytics skills? Join my upcoming Business Analytics with Python Bootcamp - LIVE with O'Reilly!
Participation requires an O'Reilly subscription. Use this promo link to try it free for 30 days!
You don't have to build AI to use AI
To integrate AI into your products, first make sure you don't make some of the most common mistakes:
Assuming that AI is only for big companies and not suitable for small businesses
Expecting AI to solve all the problems without considering the limitations and potential biases
Thinking that your company needs to be super advanced in its analytics maturity before it can benefit from AI
AI is still a new technology, and there are many misconceptions and myths around it.
For example, many people still believe that they need to be able to build their own AI service before they can benefit from AI.
Much like building a car before learning to drive.
That's not how it works.
So let's explore your options:
Option 1: Build an AI service from scratch
Building an AI from scratch means creating a new AI model and training it from the ground up to meet your specific business needs.
This is still the most common way people think about when they hear "AI"
If I want to use AI, I have to build AI, right?
Well, that's definitely ONE option.
But be aware of the complexity.
Here are some key components you need:
Data: Sufficient historical and labeled data to train a custom AI model. The amount of data required depends on the use case and can range from a few thousand to a few million examples.
Compute Power: Computing infrastructure is needed to train AI models, which can be provided in-house or rented from cloud computing providers.
Workflow: The workflow involves different stages such as data preparation, model training, evaluation, and deployment. Data scientists or machine learning engineers are usually required for this process, but Auto ML can be used to simplify the workflow.
Deployment: Finally, an application or a service must be built and maintained that allows the model to make predictions on new data and interact with users.
Building an AI service from scratch gives you the highest level of customization and control over the AI model.
For example, if you're developing a personalized recommendation system for your e-commerce platform, you may need to train an AI model with your own customer data and product catalog, which may not be possible with pre-built solutions.
When to use the from-scratch approach?
Building an AI service from scratch is a viable option when there is a strategic, high-impact use case where your organization has a competitive advantage in terms of training data and expertise.
However, it requires a significant investment in training and ongoing maintenance, so it's important to weigh the benefits of customization and control against the costs.
Option 2: Use AI-as-a-Service
This is the opposite of building your custom AI service.
In fact, you're not building anything, you're integrating a service based on an existing pre-trained AI model from vendors like Google, Microsoft, OpenAI, Amazon, IBM, etc.
This offering is called AI-as-a-Service.
The models are typically offered to you on a pay-per-use basis, mostly on cloud infrastructure, and very rarely on-premises.
AI-as-a-Service provides ready-to-use, generic AI capabilities that can be added to products and services without having to build everything from scratch.
Here are some examples:
Computer Vision AI as a Service:
Recognize entities from images (e.g. cars, people, faces, animals etc.)
Extract text from documents (e.g. tables from PDFs)
Moderate content (e.g. flag explicit images)
...
NLP AI as a Service:
Generate and complete text
Extract key phrases
Recognize entities
Translation
...
Audio AI as a Service:
Text to Speech
Speech to Text
Examples: OpenAI Whisper, Amazon Polly
Predictive Analytics as a Service:
Time-Series Forecasting
Anomaly Detection
Recommender Systems
...
In the short term, AI-as-a-Service can save you time, money, and resources.
In the long run, of course, this approach creates dependencies on the AI service provider.
The biggest myth here is that AI-as-a-Service is too generic and cannot be tailored to your specific needs.
The key is to find the right AI platform that offers the capabilities you need and can be easily integrated into your existing infrastructure.
Many platforms also offer fine-tuning capabilities for their models so you can improve performance on your data (more on this below).
When to use AI-as-a-Serice?
AI-as-a-Service is a great option if:
you're working on very generic tasks that don't need custom labeled data
you want to get started quickly to validate (prototype) feasibility and impact
you're working on non-strategic use cases where your company does not have a clear competitive advantage (e.g. you don't have any custom training data anyway)
Given the low up-front investment and plenty of options, AI-as-a-Service is a great starting point for many low profile workflows.
Option 3: Fine-tune existing models
Fine-tuning is an exciting middle ground between building an AI service from scratch and using a pre-built service, especially for many small and medium-sized businesses.
You can still leverage the capabilities of a large, pre-trained AI model while being able to customize and control your solution.
The workflow generally looks like this:
Select a pre-trained model that is relevant to your use case
Prepare your own labeled dataset that will be used to fine-tune the model
Run the fine-tuning process using a framework of your choice
Evaluate the performance of the fine-tuned model and iterate as needed
Deploy the model into your product or service to begin generating predictions or insights based on your data.
Many people believe that fine-tuning requires extensive knowledge of machine learning algorithms and programming skills.
The truth is that fine-tuning has also become more accessible through many platforms and no-code or low-code services.
For example, Huggingface offers probably the world's largest unified interface for accessing and fine-tuning pre-trained models - with or without code.
Let's say you want to build an image recognition system for your products. You could start with a pre-built model like Microsoft's ResNet-50, and then fine-tune it with your own product data to recognize specific objects.
If you're looking for a managed solution, you can try services like landing.ai or check out the big cloud providers - they all offer fine-tuning solutions.
When to use fine-tuning?
Fine-tuning is a great option if you have at least some technical expertise and some data to customize a pre-trained AI model based on your specific business needs.
It works well for both strategic/high-impact use cases and smaller projects or quick proofs of concept.
I truly believe that fine tuning will remain the most lucrative option for small and medium businesses.
Conclusion
As you have seen, integrating AI into your products and services doesn't have to be a daunting task.
By understanding the three options available to you - AI-as-a-Service, fine-tuning existing models, and building an AI service from scratch - you can make informed decisions that best fit your business needs and goals.
On the one hand, plan and design your AI journey carefully, but on the other hand, don't hesitate to get started and get hands-on experience.
The tools are out there waiting for you!
I hope you enjoyed today's newsletter.
Drop me a reply if you like!
See you next Friday,
Best,
Tobias
Want to learn more? Here are 3 ways I could help:
Read my book: If you want to further improve your AI/ML skills and apply them to real-world use cases, check out my book AI-Powered Business Intelligence (O'Reilly).
Book a meeting: If you want to pick my brain, book a coffee chat with me so we can discuss more details.
Follow me: I'm regulary sharing free content on LinkedIn and Twitter.
If you liked this content then check out my book AI-Powered Business Intelligence (O’Reilly).
Reply