The Realities of Successful AI Implementation at Scale with

Advances in AI have opened a myriad of doors in unlocking what data workers can achieve while adhering to the “do more with less” mantra dominating all modern enterprises. Notably, in the world of building data products, gone are the days of complex, resource-intensive data mining as AI leads the way toward easily surfacing insights from unstructured data. However, AI implementation is far from a “one and done” process to realize provable business value.

Jon Osborn, field CTO at, with moderator Paul Lacey, head of product marketing at, joined DBTA’s webinar, Innovating with AI: Harnessing the Power of LLMs, NLP, and OCR for Modern Data Products, to discuss how enterprises can successfully leverage AI technologies through a robust, comprehensive understanding of your business’s data needs and infrastructure.

Osborn and Lacey highlighted three steps to building modern data products, which included:

  1. Review AI use cases
  2. Choose the right method
  3. Build the pipeline

Osborn elaborated on these steps, explaining that, “I think a lot of folks don’t realize how many options are available for AI. When we say AI, what do we actually mean? From a method standpoint, how do you and when do you introduce it to your environment? How do you manage it, how do you contain it, how do you deploy it, how do you test it; those topics are really important to go through.”

When it comes to actually putting that AI into practice, “it’s one thing to have a cool Python program on your laptop that does something amazing; it's a whole other thing to put that thing into production on gigabytes or terabytes of huge data sets,” Osborn added.

He offered a humbling statistic: 80% of AI/machine learning (ML)/LLM projects fail to produce business value. When enterprises embark on an MLOps journey and they’ve created a trained model that has some sort of perceived value, many are unable to put that into production at a large scale.

The inability to go into production with a trained model is further compounded by a lack of understanding of data maturity. While many enterprises think they are ready for AI initiatives, they are, in reality, in the early stages of data maturity.

“It is true that data maturity is a real thing, even if you want to ignore it,” said Osborn. “Most people go through this progression of maturity journey…and everyone wants to be on the right side. When these new things like GenAI come up, like ChatGPT…you have to really pay attention [to data maturity].”

“This [data maturity] is one of the bases for failure in delivering business value, but it also speaks to the capabilities that need to exist within the data architectures of companies that want to do AI and ML at scale, in production, in a serious way,” he added.

Acknowledging where your organization is at in terms of data maturity will inevitably inform how the business will—or, more accurately, should—reform their infrastructures to realize true AI business value.

Lacey introduced  various options to implement LLMs in the real world: early-pipeline, mid-pipeline, and post-pipeline.

In the early-pipeline stage, LLMs are useful for unstructured data extraction. For image, PDF documents, and flare files, OCR is crucial; similarly, for unstructured text analysis, NLP will come into play.

Mid-pipeline implementation focuses on data enrichment and synthesis, where unstructured and semi-structured data is passed to LLMs for analysis. This analysis can consist of processes like classification, sentiment scoring, and summarization.

Finally, for post-pipeline implementation, LLMs can be leveraged to provide more accurate responses to queries through model fine-tuning and retrieval augmented generation (RAG).

Osborn stressed that while most organizations focus on the post-pipeline implementation, early and mid LLM usage can drive significant cost savings and increase overall data quality.

While AI technology has been around for decades, “there’s so many more accessible features available [that were not] five or ten years ago,” explained Osborn. “But there is still all this other work that you have to do in your data environment to produce an actual business value output.”

“The business value is not the execution of this model as a singular function, it’s the combination of the inputs to that execution, the other data that comes in to combine with the outcome of that data, the cleaning of that, and then the actual formulation of the business outcome,” he continued.

For an in-depth discussion of how to best implement AI for realizing tangible business value, you can view an archived version of the webinar here.