Wipro Leader Makes the Case for Smaller, Specialized AI Models to Solve for Enterprise Costs

A costly blind spot is derailing enterprise AI. While most companies build powerful systems, few have a sustainable way to pay for them. Today, the problem is so widespread that some experts estimate 95% of AI projects fail to deliver ROI. Now, the challenge for leaders is to scale a brand new technology that comes with an unanticipated, unpredictable, and often exponential price tag.

For an expert's perspective, we spoke with Brad Kaufman, Global Consulting Head for R&D in Life Sciences at IT consulting firm Wipro. In his current role, Kaufman advises companies in one of the world's most data-intensive sectors. But before that, he served as a Global Technology Strategy Partner at Cognizant and as an Executive Director at Morgan Stanley, where he built an AI and NLP-based system that increased sales revenue by 40%.

The path to sustainable growth begins by treating AI-related costs as a fundamental design constraint, according to Kaufman. His advice to other leaders is straightforward: plan for token consumption from day one. Because once architectural decisions are made, they are difficult and expensive to reverse.

"You have to view token costs as a design constraint from day one. If you wait until later, you'll find that it's almost impossible to pull those costs back in. It's like trying to put toothpaste back in the tube after you've already squeezed it out. Once you’ve made all those API calls, the design is set," Kaufman says. To avoid this trap, he recommends applying standard engineering discipline.

• Simulate to accumulate: Before scaling, teams must model a workflow's financial footprint to understand its potential costs and returns on investment, Kaufman explains. "Build working prototypes at the outset and run a sample workflow to estimate your token consumption. You can simulate a lot of this with the right tools. You can even use AI to help you predict the costs. We're starting to see tools that can do this much better now."

Leaders must look past the technical symptoms to understand why so many projects go off the rails, Kaufman says. For him, the issue is a people problem: the lack of in-house expertise creates a "fear factor" that drives poor, and expensive, decisions.

• A failure of nerve: Many organizations carry scar tissue from getting "burned" by early open-source models, Kaufman says. "I see a 'fear factor' that leads to bad assumptions. Teams assume they can't train their own models effectively, so they default to using the big frontier models from Anthropic, Microsoft, or OpenAI. But in many cases, you don't need them. You absolutely can train your own models on your own data. And you should, because that data is the lifeblood of your organization."

• The expert advantage: An experienced architect can see the hidden complexity a novice usually misses, Kaufman explains. "This is why you need experts. They are the ones who know how to ask the right questions. They see a simple plan on a whiteboard and know that in reality, you're going to need to pull in data from 20 different sources. They understand how to apply good engineering to simplify that complexity and choose the right tools for the job. That's where the magic comes in."

The right tool for the job can break this fear-driven cycle, Kaufman continues. But instead of larger systems, he champions an approach that leverages smaller, specialized models. By treating components like modular "Lego blocks" and building systems with clearly defined "abstraction layers," organizations can create a flexible architecture, he says.

• Precision over power: As an example, Kaufman points to real-world processes, like medical, legal, and regulatory (MLR) review in biopharma. Here, content is pulled from dozens of siloed sources, he says. "The instinct is often to think you need one giant API where all the content lives, but that’s the wrong approach. The key to success is training small, specialized models. Fine-tune them on your data, do it quickly, know what 'good enough' looks like, and know how to test it effectively."

• Future-proofing the stack: The AI landscape evolves at a blistering pace, he explains, with new, better models emerging every few months. "Keep your teams small and agile, and make sure your architects are involved from the start. Especially the ones who know the different AI models out there. Your goal should be to build a modular system. That way, if a better model comes along in six months, you can plug it in without breaking everything."

• Let the AI help: It might seem intimidating, but pairing experts with modern tools can accelerate the process significantly, Kaufman says. "This doesn't have to be an overwhelming task. You can literally talk to an AI like Claude and say, 'Here's my code base. Show me how to build an abstraction layer on top of it.' It can get you 80% of the way there."

With a philosophy honed in life sciences R&D, Kaufman sees the industry as an ideal proving ground for AI. Because the data is notoriously disorganized and security needs are high, the value of embedding AI-aware developers directly into business units is clear. Now, that blueprint can be repurposed for almost any enterprise.

• Into the labs: For Kaufman, the goal is to uncover "unarticulated needs": the bottlenecks and manual workarounds that business users might not recognize as problems an AI could solve. "Think about R&D and lab digitalization. Every data source comes from a different machine, and each one has a different format. The only way to solve that is by embedding your best developers directly into the labs. Let them see firsthand what the researchers need and where their workflow slows down. That strategy works really well."

• The data deluge: The explosion of data is an irreversible reality, he says. The toothpaste is out of the tube. "The amount of data is only increasing. Sure, it's gotten faster to integrate third-party data. But that has only made the appetite for more data more significant. In this environment, more data wins, but only if you know how to synthesize, aggregate, and redact it effectively. There's a real art to that." That's precisely why a disciplined, modular, and expert-driven approach is no longer optional, Kaufman explains.

Ultimately, his message is anchored by a single question: What will deliver the most business value? The power of AI, he says, is best harnessed when its limitations are respected. Offering a simple golf analogy for this "respect mentality," he concludes by comparing a large, expensive model (the "driver") with a smaller, more precise tool (the "five iron"). “Don't always pull the driver out of the bag just because you can. Sometimes, the right play is to pull out the three wood or the five iron. It’s about knowing which tool to use and when. You have to care, have respect for the complexity, and manage your token costs. There are ways to do this right, but it requires an appreciation for the art of it.”