Zooming in on the generative AI value chain

Let’s clarify the terminology first.

Gen AI differs from traditional AI technologies in its ability to produce new digital content, be it legal documents, report summaries, images, product designs, etc., while classic AI focuses on predictive analytics like forecasting demand.

Generative AI is also different when it comes to implementation and usage. Gen AI models are typically larger and need more computational power to build, train, and operate. They are immensely powerful, but they also come with unique challenges, such as hallucinations. You can learn more about the pros and cons of Gen AI on our blog.

Our analysts also prepared detailed articles on how to use this technology in different sectors. Check if you can find your industry below:

Here are the six key links of the Gen AI ecosystem.

Let’s take a closer look at each of these building blocks.

Generative AI models usually need enormous computational power, high-bandwidth memory, large storage systems, and efficient cooling equipment to function. As these models have grown exponentially for the past decade, traditional computer hardware isn’t up to the task anymore.

For example, the ELMo model for image recognition that was trained in 2018 contained over 20 million parameters. Google’s BERT, which came shortly after, already exceeded 300 million parameters. And the more recent large language models (LLMs) can easily contain hundreds of billions of parameters. For instance, GPT-3 has 175 billion parameters. OpenAI trained this model on 45 terabytes of data, which equates to a million feet of bookshelf space. GPT-4, which was released in 2023, comprises 1.8 trillion parameters, while the upcoming GPT-NeXT is expected to be 100 times more powerful than its predecessor.

Gen AI needs advanced chips and computational resources. It takes innovative processors, such as graphic processing units (GPUs) and even tensor processing units (TPUs) with accelerator chips to build and train these tools.

Such hardware is rather expensive. You can consider buying this equipment if your company falls into one of the following categories:

For other organizations, it makes sense to implement Gen AI in the cloud.

Cloud infrastructure provides access to expensive computing and storage resources. It enables companies to use hardware on demand and scale rapidly as their business expands. So, instead of buying pricey GPUs and TPUs and installing comprehensive cooling systems, many organizations turn to cloud computing.

Currently, there are three major cloud service providers on the market—Amazon AWS, Microsoft Azure, and Google Cloud.

Keep in mind that you can combine on-premises and cloud deployment. For example, you can retrain a Gen AI model on your proprietary data on premises to avoid uploading the data to a third party. Afterwards, you can run the mode in the cloud.

Foundation models are built and pre-trained on massive datasets of private or public data, making them suitable for a variety of general-purpose tasks, such as generating realistic images and summarizing lengthy texts. One foundation model can power several independent applications developed by different companies.

Organizations can fine-tune these models on proprietary datasets to perform more specialized tasks. You can use a commercially available model and pay license fees, or you can opt for an open-source solution. The second option gives more room for flexibility and customization.

Building and training a foundation model from scratch is an extremely expensive process. OpenAI is believed to have spent at least $4 million on training its GPT-3 large language model, which drives many text-generating applications today. But price is not the only prohibitor. Building Gen AI models takes diverse expertise, including AI consultants who will design and build the model, data scientists to prepare the data, and domain experts to verify the output and give feedback.

You can find more information on how much Gen AI costs on our blog.

Apps serve as an interface between Gen AI models and the end user. Even though foundation models can complete dedicated tasks, they can’t deliver value without applications.

An LLM that was trained to generate high-quality text will just sit idle until someone develops an app that activates it. One company can utilize the same LLM to create applications for different use cases. For instance, an HR department can use this Gen AI model to generate vacancy descriptions, while customer support specialists can envelop the model in a chatbot app that interacts with customers, and yet another application can use this model to summarize documents.

You can hire an app development company to design and build an application that leverages a foundation model of your choice and fits seamlessly into your workflow. We also encourage you to visit our blog to learn more about the application development process and the associated costs.

Companies need dedicated tools to deploy and maintain Gen AI models and adapt them if needed. And that’s where MLOps comes in.

MLOps tools and technologies enable AI teams to maintain and interact with the model. For example, the ITRex MLOps consulting services include aggregating and preparing data for model retraining, validating the model, implementing tools for performance tracking, building APIs to allow applications to interact with the model, deploying the model, and more.

You can learn more about what MLOps can do for your business on our blog.

No matter how powerful, Gen AI is just a technology, and you need people to operate it. Skilled professionals are still in the driver’s seat for innovation, reliability, and ethical standards. A talented workforce will give you a fresh perspective on emerging opportunities, spot and correct the mistakes AI makes, and ensure that AI models are ethical and free of bias.

If you don’t have the required expertise in-house and don’t want to recruit new people on a full-time basis, you can opt for the dedicated team hiring model. You can contact a specialized outsourcing company that will suggest a list of trusted professionals for you to choose from. The people that you select will work for your company on a flexible schedule for the duration of your project.

After learning about the six pillars of a generative AI value chain, let’s dive into the steps that organizations can take to maximize the value of their Gen AI endeavors.

Did you know that the top five Gen AI use cases constitute 50%-80% of the overall value the technology can bring to your business? How can you identify these?

For every company, these mission-critical applications will differ. There are two approaches that you can follow to identify the relevant use cases for your company.

One option is to focus on short-term benefits and consider use cases that will give a rapid return on investment. Or you can look into the technology’s long-term potential and search for ways to transform your processes entirely. In this case, your CTO and the tech team will work closely with business stakeholders to take a holistic approach to change and rethink your business processes. In the end, this team will come up with a global technical roadmap for possibly disrupting the business in its current state.

If you take the second approach, generative AI can enable your company to deliver value in novel ways, leading to an exponential revenue increase.

Make a strategic assessment of the potential value that each use case can add to your businesses and the risks and difficulties associated with AI implementation. Besides values and risks, you can also consider other factors, such as deployment time, the associated costs, scalability, and complexity.

Additionally, it’s important to consider your company’s corporate culture, existing workflows, and core products and services and evaluate your business partnerships, competitive landscape, and regulations.

Now it’s time to choose the foundation models, cloud providers, AI consultants, and any other partners, vendors, and tools that you might need during your Gen AI journey.

When it comes to foundation models, it’s not feasible for most organizations to build them from scratch. Especially since there are many off-the-shelf solutions that were trained on large datasets to perform specific tasks. You can select one of these. Should you follow this path, it’s recommended that you retrain the readily available models on your proprietary data to achieve superior performance. But you could also use a ready-made Gen AI tool as is under the following circumstances:

There are two types of off-the-shelf Gen AI models that you can retrain and customize:

Let’s take a closer look at each type.

During this step, you also need to decide on the architecture approach and address questions, such as:

As mentioned above, a ready-made generative AI model is seldom enough as is. Companies, in most cases, need to familiarize the selected foundation model with the specifics of their business. This will also give you an edge over the competition that took this model without change.

To fine-tune a Gen AI model on a proprietary dataset, firms need to collect and aggregate this data, prepare it for AI consumption, and make sure it’s bias-free and representative of the target population. Also, address any ethical concerns, including data privacy, and obtain consent when needed.

If you already have a data management framework, it will save you time and money. If you don’t, then this is the right time to establish one.

After deciding which foundation models you want to use, it’s time to think about where to host them and how to scale in the future.

If you opted for a proprietary Gen AI model, the vendor will deploy the model on their premises and scale the allocated resources as your operations expand. You will just have access to an API through which you can interact with the model. But if you choose an open-source solution, you have several options:

As both technology and your business evolve, you might want to systematically reassess your generative AI tools to make sure they are still fit for their purpose. Besides business relevance, it’s preferable to audit the models for accuracy and prejudice, such as bias. If left unnoticed, these can cause inconveniences and, in the worst case, hefty fines and reputational damage.

If you find that the model produces erroneous output, which can happen if you introduce new, unfamiliar data, you can retrain it with an updated dataset. If that won’t be enough, you can go back to Step 3 to look for a different Gen AI model.

After you successfully deploy Gen AI for one application, you can look for other related use cases that can benefit from this Gen AI tool. Scaling the technology to the next application will be much cheaper and faster.

Even though you can expand Gen AI to other existing applications, you might also consider reimagining some of your workflows with the help of the technology.

As we hear about Gen AI everywhere, it might seem that most companies have implemented the technology and are already reaping the rewards. Or are they? In its recent study, the Boston Consulting Group (BCG) surveyed senior executives across ten sectors and found that only 10% of the companies are scaling their Gen AI initiatives, with 50% being stuck in the piloting stage and 40% still observing and not taking any action.

Companies that haven’t experimented with the technology yet can still start their Gen AI journey and, according to BCG, even catch up with their piloting competitors. But they need to act fast. The longer they postpone the decision to collaborate with generative AI consultants and put the technology to work, the wider the gap becomes.

Here are valuable tips from ITRex that you can use along with the path to generative AI value delivery described above.

This tip is not limited to AI. You can use the same approach with any cutting-edge technology.

Investing in the generative AI value chain building blocks and following the six steps described in this article will help you with Gen AI deployment. But the success of such projects also depends on the people involved. ITRex will make a great Gen AI partner that will assist you every step of the way. Here is why.

And don’t be afraid to experiment with generative AI. As ITRex CEO Vitali Likhadzed says: