Technology experts reveal the traits of AI innovators, dispel common misperceptions, and share value-creation strategies during SymphonyAI event
Bold claims about productivity and process improvements possible with enterprise AI can make it difficult to separate what’s real right now and what might be possible tomorrow. While there has been considerable hype around AI, the debate has shifted from whether it creates value to how much, how fast and what are the right moves companies should be making today.
To bring clarity to those issues, SymphonyAI assembled four of the sharpest minds in the technology world for a virtual roundtable discussion on the topic of, “Cutting through the hype to unlock true enterprise AI value.”
The conversation was moderated by R “Ray” Wang, who is chairman, founder and principal analyst with Constellation Research and featured panelists included Microsoft Azure Senior Director Eve Psalti, Dr. Daniela Rus, director of MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), and SymphonyAI CTO Raj Shukla.
Wang and the panelists addressed topics such as unique enterprise AI requirements, barriers to adoption, whether to build or buy, top use cases, proof of concept success factors, AI misperceptions and emerging trends. The group covered a lot of ground, but the initial focus was on the state of AI in the enterprise and practical steps for success.
“AI has been around for a long time, but it’s back again and the number one topic everywhere, especially with generative AI taking all the spotlight at the moment,” Wang said as he asked panelists to share their views on the reality of AI in the enterprise today.
“There are a lot of expectations and sometimes unrealistic expectations,” said Psalti. “Companies presume there is going to be immediate ROI, and the reality is that a lot of companies are in the early stages of AI adoption.”
While enthusiasm about generative AI is well-deserved, Rus believes companies have begun to realize there is more to AI than generative AI. For AI to live up to the hype requires combining generative AI with other types of AI which are less frequently discussed, such as predictive, optimizing and deciding AI.
“I’m very bullish about AI and the value it can bring to companies, but we have to think about it more holistically and companies have to think very carefully about their playbook with AI,” Rus said.
Whatever type of AI is involved, it requires getting the data layer right to unlock its effectiveness against use cases which vary by industry vertical, according to SymphonyAI’s Shukla. So unlike ChatGPT or other large language models (LLMs) in the market that are trained on publicly available data sets, the value of enterprise AI derives from access to and training on vertical- and company-specific data, so the model can be used for reasoning and to connect the dots across the enterprise while adhering to good data governance and privacy guidelines.
“There is all this enterprise data available, but have you built the foundation, add the data layer, add the API layer to make all of it available to a generative AI model to access as required,” noted Shukla . “That’s the key difference that everybody realizes versus the consumer applications where one model has been trained and it answers questions for you.”
Build or buy?
Deciding whether to build or buy enterprise AI solutions begins with being clear on use cases and whether they are specific to a company or industry vertical. Shukla addressed this point by explaining how building a solution may be the only option in situations where the problem being solved is unique to one company. However, when a use case is common across multiple companies in an industry vertical, then buy is the preferred approach. He offered an example common in the retail industry of evaluating transferable demand when products are added or deleted from a retailer’s assortment.
“The idea of demand transfer is common and SymphonyAI products will let retailers solve that use case, so it makes sense to buy something like that,” Shukla said. “But if a problem is very specific to particular store locations, then maybe you are better off building a solution. This analogy applies to other industry verticals as well.”
Cost is also a major consideration, according to other speakers who shared the view that building something is going to be more costly than buying so the economic justification needs to be compelling.
“In very, very few cases should you go and build your own,” said Psalti. “A common misconception that we see is Microsoft customers jump to the conclusion that they need to develop their own generative AI model and that can be a very expensive and time consuming operation.”
It’s also often unnecessary. Psalti noted that when using generative AI, it is important to begin customizing it through prompt engineering (the process of giving a model instructions to improve results). Then companies can bring the LLM up to date with retrieval augmented generation methodologies (RAG) to ground the LLM in a more factual way for specific use cases.
“It doesn’t require a whole team and a lot of time to do this, and only then would I say go and fine tune, which basically is changing the weights of the model, but that requires a little bit more skill and know-how,” Psalti said.
Moving from pain point to POC to production systems
The process of buying an enterprise AI product typically involves a proof of concept (POC) evaluation phase before moving into a full-blown project. As Wang noted, POCs may have conversion rates as low as 20% so he pressed the panelists on how to overcome barriers to getting enterprise grade POCs to the actual deployment phase.
MIT’s Dr. Rus believes more understanding and planning is needed at the front end of the process to establish whether the use of AI is appropriate and having a clear objective and a clear strategy for achieving it.
“Defining a clear objective means understanding what problem you are trying to solve,” Rus said. “It also means defining and understanding the stakeholders, the roles, the data, understanding the measures of success, understanding all the aspects of delivery that will go from the idea of the problem to a small POC and to an actual enterprise-level solution.”
To aid that process, Psalti shared the view that pain point identification is essential. Once companies identify areas where they are spending a lot of time because automation isn’t in place or data and processes are siloed, it helps prioritize which POCs to pursue because they are most likely to achieve desired outcomes.
“Another thing we’re seeing across companies that have successfully integrated generative AI into their operations is that they all have a center of excellence. So they can spin up POCs in multiple departments or groups, share some of the best practices and move fast,” Psalti said.
A center of excellence can help an enterprise choose optimal use cases to target, because as SymphonyAI’s Shukla noted, some of the problem areas have existed for years or even decades.
“When the problem exists, and we can put a number on the value, the POC would be defined as what is the desired outcome,” Shukla said. “In our industrial vertical, we would look at predictive maintenance and the issue of detecting a machine failing before it actually does. Our Iris Foundry product does a great job at this.”
Financial services is another area where enterprise AI helps solve a compelling use case – in this instance, the longstanding industry problem of detecting and preventing fraud and financial crime. With a properly trained model powering predictive capabilities, early detection of illegal activity is possible in a way that is more effective than traditional rules-based approaches that produce large numbers of false positive warnings. Generative AI also helps dramatically accelerate generation of suspicious activity reports (SARs).
“We work with a large bank in South Africa and have shown that replacing their rules-based system with a predictive model reduced false positives by more than 70%,” Shukla said. “With another financial services customers we have reduced the time of investigations by 40%.”
Dispelling common misperceptions
When it comes to expectations of enterprise AI, there are those that relate to the creation of business value as well as others involving potentially negative effects. The panelists addressed this topic by focusing on the more frequently raised concerns, including that AI is a job-killer.
The speakers cited industry research that shows headcounts generally remain stable with adoption of enterprise AI, but productivity, effectiveness and employee and customer engagement improve. That stands to reason as the availability of better tools means people can focus on things that matter rather than undesirable repetitive tasks. However, that can also lead to concerns that workers are displaced as companies move through Wang’s “Five A’s of AI,” maturity framework which includes the phases of augmentation, acceleration, automation, agents, and advisors.
The panelists agreed that most companies are in the augmentation and acceleration phases with a few dabbling in the automation phase. Regardless of where companies fall, Microsoft’s Psalti believes the common thread across all is that significant human oversight is needed.
“These AI models, as innovative and effective as they are, are not really autonomous. They need human oversight, they need fine tuning, they need interventions,” Psalti said. “I agree that we need to employ them and start with low risk use cases just to train and continue to customize things and make sure that we are comfortable with the fact that they’re not going to yield results immediately as it takes time to prepare the data sets to do the model training and iterations that need to happen to provide value.”
AI’s connected future
As each of the speakers looked ahead to a future where AI is more integral to a high functioning enterprise leading to accelerated value creation, several factors stood out as drivers of future success.
For MIT’s Rus, she sees evolving roles at organizations and a blurring of responsibilities as employees become bilingual in the needs of the business and the specifics of enterprise AI.
“You can’t just be an AI expert who’s learned the theory. You really need to understand your business. You need to understand the market, you need to understand the products,” Rus said. “If you’re only trained in the business domain, then you don’t know AI and it’s difficult to understand what the possibilities are. If all you know is AI, you don’t really know how AI can be useful. It’s important to have people who understand both sides.”
It’s also important to pursue a new vision of the connected enterprise, as that holds the key to AI in all its forms exceeding today’s lofty expectations. As SymphonyAI’s Shukla explained, the connected philosophy is evident in one of the company’s flagship products for retail called CINDE Gen AI. The product unifies disparate sources of data to provide a comprehensive picture of retail operations to help retailers make more informed decisions about operations, merchandising, and supply chains.
“The true power of AI models is they become better if they are given connected data,” Shukla said. “Sometimes a POC will fail because you didn’t connect the story and to connect the story you have to go top down, and the data teams and the operational teams have to come together. That’s how we have developed our solutions, and it is resonating in the industry.”
Learn how enterprise AI can create value for your organization. Get in touch today!
