Boston Consulting Group: To unlock enterprise AI value, start with the data you’ve been ignoring

When building a business AI, some companies find that the most difficult part is sometimes to decide what to build and how to approach the different processes involved.

HAS Venturebeat Transform 2025The quality of the data and governance were at the center and the center, the companies beyond the experimental phase of the AI ​​and explore the means of productivating and the failure of agents and other applications.

>>See all of our Transform 2025 coverage here<

Organizations face pain to think about how technology is believed with people, processes and design, said Braden Holstege, Managing Director and Partner of Advice group in Boston. He added that companies must think about a range of complexities related to exposure to data, AI budgets per person, access authorizations and external and internal risk management.

Sometimes new solutions involve means to use data previously unusable. Speaking Tuesday afternoon on stageHolstege gave an example of a client who used large languages ​​(LLM) models to analyze millions of ideas on people who are transformed, product complaints and positive comments – and discover ideas that were not possible a few years ago with the treatment of natural language (NLP).

“The wider lesson here is that the data is not monolithic,” said Holstege. “You have everything, from transaction records to documents to customer comments, including trace data that is produced during the development of applications and one million other types of data.”

Some of these new possibilities are thanks to improvements in data ready for IT, said Susan Ellinger, principal director of Microsoft d’Azure A AI.

“Once you are there, you start to get this sense of art of the possible,” said Elinger. “It is a balance between this and to come with a clear meaning of what you are trying to resolve. Let’s say that you are trying to resolve the customer experience. This is not an appropriate case, but you don’t always know. You can find something else in the process.”

Why the data practiced is essential for the adoption of companies

The data ready at AI is an essential step in the adoption of AI projects. In a separate Gartner investigationMore than half of the 500 medium -sized business CIOs have said they expect the adoption of infrastructure ready at ITA would contribute to faster and more flexible data processes.

It could be a slow process. Until 2026, Gartner predict Organizations will abandon 60% of AI projects which are not supported by data practiced by AI. When the research firm questioned the data management managers last summer, 63% of respondents said that their organizations did not have good data management practices, or that they were not sure of the practices.

As deployments become more mature, it is important to consider the means to meet the challenges in progress such as the drift of the AI ​​model over time, Awais Sher Bajwa, head of data and IA banks said in Bank of America. He added that companies do not always need to precipitate something to put an end to users who are already quite advanced in the way they think of the potential of cat -based applications.

“We all in our daily life are cat users,” said Sher Bajwa. “Users have become quite sophisticated. In terms of training, you do not need to push it to end users, but this also means that it becomes a very collaborative process. You must understand the elements of implementation and scaling, which become the challenge.”

The growing pain and complexities of AI calculate

Companies must also take into account the opportunities and challenges of applications on cloud, on -site and hybrid applications. Compatible with clouds AID Applications Authorizing the test of different technologies and scaling in a more abstract manner, said Sher Bajwa. However, he added that companies must take into account various infrastructure problems such as security and cost – and that suppliers like NVIDIA and AMD facilitate tests of different models and different methods of deployment deployment

Decisions concerning cloud suppliers have become more complex than they were a few years ago, Holstege said. Although new options such as neoclouds (offering servers and virtual machines supported by GPU) can sometimes offer cheaper alternatives to traditional hyperscalers, it has noted that many customers will probably deploy the AI ​​where their data already resides – which will make changes of infrastructure major less likely. But even with cheaper alternatives, Holstege sees a compromise with IT, cost and optimization. For example, he pointed out that open source models like Llama and Mistral can have higher IT requests.

“Does the cost of calculation are worth it to incur the headaches of the use of open source models and migration of your data?” Holstege asked. “Just the border of the choices that people are now confronted with is much wider than it was three years ago.”