Enterprise AI gets real with voice agents, workflows, and the new ROI math

At a time when the world is pouring trillions into AI infrastructure, one question looms large: Is enterprise AI truly delivering value, or are we still living in the hype cycle?

At TechSparks 2025, a panel featuring Raghav Chandra, Co-founder & CPTO, Urban Company; Khadim Batti, Co-founder & CEO, Whatfix; and Bala Gorthi, COO, Teleradiology Solutions (TRS) dove straight into the question. Moderated by Pankaj Mitra, Partner, Bessemer Venture Partners, the discussion blended humour, candour, and rare behind-the-scenes detail on what’s really working in enterprise AI.

Inside Urban Company’s AI ops engine

Chandra opened with the use case that excites him most, AI-driven partner operations. “The bar to solve problems on the partner side is much higher. You can’t solve this with text. You have to wait for voice to become real,” he said.

Urban Company today handles everything from onboarding and training to dispute resolution using AI-driven voice agents. But it wasn’t easy.

He played a real two-minute clip of a live AI–partner helpline interaction: full of interruptions, clarifications, back-and-forth, and parallel workflows. This wasn’t a demo. It was messy, real, operational complexity, and the AI agent handled it end to end.

“Six months ago, this was impossible. The models weren’t sophisticated enough, and tonality was nowhere close,” Chandra said.

The hidden challenge, according to him, lies in ROI measurement. Chandra broke down AI ROI into three buckets: revenue growth, quality improvement, and human replacement. Human replacement is the clearest ROI model but also the hardest to execute.

“You don’t replace a full human. You replace 8% of 10 different roles. That means headcount reduction only happens after a year,” he noted.

What would Urban Company do differently now?

Chandra shared key early decisions, some turned out to be huge advantages:

• No proprietary model fine-tuning: “Foundation models will get better. If we build too much on one, we’ll get stuck,” he said.
• Vendor agnosticism: “Every week Twitter says a different model is winning. The graphs look like my college hairstyle - spiky and oscillating,” said Chandra.
• Prepare for rapid tech-debt accumulation: “Last year there was nothing called tool calls, AI directly invoking a company function. Today it’s production critical. Everything we built a year ago already feels archaic,” he revealed.

Selling AI to the enterprise

As a horizontal SaaS player powering over 85 Fortune 500 companies, Whatfix sits closest to the enterprise buyer. Batti’s big reveal was their AI agent Seek, an agent capable of autonomously performing tasks a Salesforce developer does.

“Seek completed the entire Salesforce Trailhead beginner + intermediate task list, around 130 tasks, autonomously in a few hours,” he said.

When asked if Salesforce developers should worry, he smiled, “Before someone disrupts us, we want to disrupt ourselves.”

Enterprise buyers have transformed in 24 months. “Two years ago,” Batti said, “CIOs said, this looks good… but don’t enable it in my environment.” One year ago, “Suddenly there were AI councils and task forces.” And now, “These councils have budgets. They have identified core use cases they must solve to stay relevant.”

For instance, pharmaceuticals want faster molecule discovery. Automotive wants AI-enhanced product design and engineering. Industrials want predictive maintenance and automation. But horizontal solutions still face slower adoption.

“Vertical AI use cases have skin in the game. Horizontal AI is still fighting the ROI battle,” he said.

AI in high-stakes healthcare

While Urban Company focuses on consumer-scale operations and Whatfix on enterprise productivity, Gorthi of TRS deals in one of the highest-stakes sectors: healthcare.

Radiology has long been labelled “the first job AI will replace.” Gorthi laughed at the prediction. “2018, Geoffrey Hinton said radiologists would be out of jobs in five years. Today, I saw a radiologist driving to work in a regular car.”

TRS is using AI primarily for flagging critical values, reducing reporting times, and ensuring no anomaly slips through. What sets their model apart is the integration into the workflow, from scanning to reporting to prioritisation.

“I think the spectrum is slightly different for us,” Gorthi explained. “Cost matters, of course, but in our world, risk matters more. And we always have a human in the loop. In the worst-case scenario, if everything else fails, a trained radiologist will make the final call.”

He emphasized that AI’s primary role in healthcare is augmenting accuracy first, and scale second.

“If I have a PACS system receiving 100 cases every hour, I’m juggling priorities to make sure every patient gets timely attention. If the AI flags even one stroke, one bleed—anything critical—that alert pushes the case to the top of the queue,” he said. “It’s not replacing the human. It’s helping the human see sooner what they might have otherwise seen later.”

He described AI as documentation and triage intelligence, not automation, but an assistive layer that helps radiologists perform at their best.

A sector-agnostic reality check

Across the panel, a few truths emerged clearly:

1. AI is no longer optional, but ROI must be justified: Boards are pushing AI, CIOs are allocating budgets, but the last-mile ROI remains the hardest metric.

2. The biggest breakthroughs are happening in workflows no one sees: Partner operations, radiology workflow automation, Salesforce integration tasks which are gritty, but massively impactful.

3. AI build cycles are collapsing: Every three months, new capabilities rewrite the stack.

4. Hybrid AI + human models are the future: AI is augmenting, accelerating, and automating; not replacing entire roles wholesale.

The panel delivered a clear message: Enterprise AI is real, but only for those who build with discipline, understand operational complexity, and are willing to rebuild every few months as the technology shifts. As Chandra put it, “This is as cutting-edge as tech allows today. But six months from now, this will look primitive again, and that’s the exciting part.”

And perhaps the best closing line of the session came from Gorthi, “AI won’t replace radiologists. But radiologists who don’t adopt AI? They might just replace themselves.”