Mphasis Rewrites the Economics of Legacy Modernisation with AI

Mphasis, a mid-tier IT services firm, using AI agents to directly decode and modernise legacy COBOL applications, is reducing modernisation costs from about $7 per line of code to $3, it said.

This approach breaks from the IT industry's usual practice of wrapping artificial intelligence around decades-old software, according to Chief Executive Officer (CEO) Nitin Rakesh, who described cosmetic AI overlays on legacy systems as little more than ‘window dressing’.

"Software will eat software," he said, referring to using artificial intelligence to continuously dismantle and re-engineer legacy systems rather than simply layering new technologies over them.

Across the industry, many enterprises are adding artificial intelligence as a layer over existing systems rather than rebuilding core platforms. Consulting firms including Accenture and Deloitte are deploying AI chatbots and analytics tools over existing SAP and Oracle ERP systems for banks and manufacturers, delivering quick efficiency gains while leaving aging infrastructure largely untouched, according to analysts.

Mphasis in contrast said, its AI NeoIP platform can reverse-engineer applications within three to six months with 90% to 95% accuracy, compared to traditional timelines of two to three years. A typical 30-million-line modernization that would cost around $210 million over seven years can now be done in three years for under $90 million, Rakesh said at a recent press meet held in Bengaluru, after the post-earnings call.

"We're not using AI as a cosmetic layer," Rakesh said. "We're using it to find, preserve and constantly update the business logic embedded in existing applications."

The IT services company reported that over half its revenue now comes from AI-infused work, with more than 60% of new deals worth $428 million in the third quarter being AI-led. Net profit rose 3.3% year-on-year to ₹442 crore, while revenue climbed 12.4% to ₹4,002 crore, though margins took a hit from talent and AI investments.

The mid-tier IT services firm works extensively with COBOL, a programming language developed in the 1960s, because many global banks and insurers still run their core transaction systems on it. According to Reuters, 43% of banking systems are built on COBOL, with 220 billion lines of COBOL code still in use today, making safe modernisation critical for systems processing millions of transactions daily.

Rakesh pointed to one example where a client with one of the world's largest mainframe installations spent five years decommissioning hardware but couldn't modernize the applications. A complete rewrite would have cost between $400 million and $900 million with no guarantee of success.

Training AI to Read Before Writing

"Everyone's so excited about ChatGPT writing software. How about we train AI to read software?" Rakesh said. "It's a 180-degree flip. Input is software and your output is English. Your input is not English and output code."

This flips how most companies approach generative AI. Instead of using tools like GitHub Copilot or Claude to generate code faster, Mphasis uses AI to decode decades-old systems first.

The company's NeoZeta agent extracts business rules from legacy code and feeds them into Ontosphere, Rakesh said. The platform can pull rules from COBOL, documents, user manuals, operations manuals, or any document where knowledge exists, he said.

This creates what Rakesh called "an intermediary asset that never goes stale". The system keeps updating its understanding of the software every time someone makes a change, avoiding what he called the accumulation of technical debt.

The NeoIP platform uses several specialized AI agents. One called NeoSaba acts like a virtual business analyst, using the Ontosphere knowledge graph to write software requirements, Rakesh said. Another agent, NeoIra, focuses on redesigning applications for modern systems. Other AI agents help generate code using tools like Copilot, Gemini and Claude, and manage the development process.

Operations Over Coding

While the tech industry promotes AI coding tools, Rakesh argues the real cost savings will come from AI-driven IT operations, not products like Copilot or Gemini.

He pointed to a fundamental problem with how IT services work today. The traditional "break-fix" business model means companies profit from resolving problems rather than preventing them. "We wait for something to break so we can fix it and charge for it," he said. "I have zero incentive as an industry to eliminate incidents and increase availability."

He compared today's IT operations unfavorably to modern cars. "Even your car has a yellow light that flashes when something is starting to go wrong. But our IT operations don't do that even today."

AI-driven operations that predict and prevent failures rather than fix them after they occur could upend this model, potentially cutting the need for expensive emergency repairs that currently generate revenue for IT service providers.

Mphasis said its NeoIP platform delivers 60% efficiency gains in development and cuts the time to detect and resolve IT incidents in half, with three to five hours of advance warning before major outages.

Shifting to Outcome-Based Pricing

Clients are moving from headcount-based pricing to outcome-based contracts, Rakesh said. One high-tech client now tracks eight productivity metrics for contractors. Another is unbundling a software-as-a-service deal to separate software from services. "They didn't know how many people are engaged. It was a bundled deal," he said. The client wants to understand what they're paying for and get results rather than just more headcount.

Mphasis tested its approach two years ago with a client running 40 million lines of COBOL for payments processing. The client needed to cut customer onboarding from 18-24 months to two months to compete with digital-native firms. Traditional tools estimated the modernization would take seven years.

The company doesn't break out AI revenue separately because AI is woven into everything it does now. "If everyone is touching AI, what exactly am I breaking out?" Rakesh said. The goal is to have AI in every single project.

Companies that can show measurable results rather than simply add more bodies to a project may have an advantage. Mphasis has already changed how it hires, moving away from campus recruitment to internships and hackathons. The company now looks for what Rakesh calls "learnability" over static technical skills that become outdated quickly as AI reshapes the industry.