How pre-trained AI models increase ROI

The availability of pre-trained AI models is enabling lower cost, low risk development of AI applications.

Why is this important now?

Over the past 5 years, AI initiatives have failed or failed to produce business value in 3 ways: 1) they took longer than expected, 2) they cost more than expected, or 3) they did not deliver the results that were expected.

During this time some of the influential players (NVIDIA, Google, Meta) were making significant investments (think billions of dollars) that are now starting to produce results that benefit the entire industry. Those investments were centered around creating and publishing pre-trained AI models.

How does their investment improve my ROI?

With the newfound availability of pre-trained models, we can now take on AI initiatives at more moderate costs. This means that smaller investments can lead to substantial returns, and higher rates of return. In the past, even a small AI initiative required such a significant investment that anything short of a largescale business transformation was impractical. $10M to $50M investments required enormous returns. But with pre-trained models, $1M-$2M investments can yield multi-million-dollar returns. Risk is reduced, costs are recoupled, and positive returns are realized in a fraction of the time.

What can I expect to spend?

Three years ago, a moderate AI effort would take a year and require us to spend $10M to $100M. With the availability of pretrained models, similar capabilities could be developed in 6-9 months for $1M to $2M.

What types of projects should I invest in?

Few mid-sized companies will spend $1M to update back-office systems, even though those investments may pay off in a few years. Improvements like that are always difficult to fund. Product improvements, on the other hand, certainly warrant significant investments, particularly when it is necessary to retain and grow market share. AI driven capabilities are appearing in products, and for good reason – they have the potential to improve the customer experience and to reduce production costs.

Where do the savings come from?

On average, using a pretrained model saves most of the cost associated with AI application development. These savings come from: 1) Eliminating training time and compute resources, 2) Eliminating large dataset collection and maintenance, and 3) Reducing in the number of PhD AI Engineers required to complete the work.

What pre-trained models are out there for me to use?

Pretrained models for image and video analysis, object recognition and classification, face recognition, and numerous natural language processing tasks are available for free download*. With a moderate level of science to validate the model and assess the training set, these models can be integrated into applications that are compatible with corporate infrastructures.

*Examples:  https://huggingface.co/, https://tfhub.dev/, https://modelzoo.co/categories

How should I adjust to take advantage of pre-trained models?

Tailoring AI initiatives to leverage pretrained models can lead to significant returns in the short term, delivering the promise of AI while reducing the risk of a science failure.