In our latest episode of Deployed, we talked with Surojit Chatterjee, CEO and founder of Ema, who brings a unique perspective to building AI products. Before founding Ema in early 2023, Surojit spent a combined 11 years at Google (including as VP of Product for Google Shopping, and leading mobile ads from zero to tens of billions in revenue) and he also served as Chief Product Officer at Coinbase.

We were excited to talk to Surojit because of the timing of Ema's founding. Surojit started building enterprise AI agents in early 2023 – well before "agentic AI" became the buzzword it is today, and at a time when most enterprise technology leaders viewed the idea of trusting an LLM to make complex business decisions as dangerously risky.

Two years later, Ema is working with some of the world's largest enterprises like Hitachi to transform critical business processes across HR, finance, and customer support. Surojit's experience building agents in these environments gives him valuable insights for anyone working on AI products, particularly around what it takes to get AI agents to a level of production quality that enterprises actually trust.

The episode is live on Spotify, Apple Podcasts and YouTube, and you can watch the whole thing here. Read on for some of our favorite highlights.

Why "AI Employees" Instead of "Agents?"

One of the most thought-provoking aspects of our conversation was Surojit's choice to frame Ema's product as "AI Employees" rather than agents. This isn't just about marketing: It reflects a fundamentally different way of thinking about how AI systems should be built, deployed, and improved.

"We want our business users to really think of these as employees, ‘AI employees,’ not software at all. The way they are giving feedback is just natural language. Like they will give feedback to another human. Think, do more of this, do less of that, whatever…"

He’s not the first enterprise AI leader we’ve heard frame things this way. Treating AI agents like employees has practical implications for product design, and for the concept of evaluations and quality improvement. Especially as more business users ramp up building and deploying their own agents for business-critical work, they need to learn some of the concepts of observability and evals without learning those technical skills.

If you treat agents as employees, it means that you can set them up just as you would onboard a new human employee with documentation, training, and feedback, Ema's AI employees are designed to receive similar treatment. Users can provide natural language feedback on any action the AI employee takes with a thumbs up/down, written feedback, or more detailed corrections.

Companies already have well-developed systems for evaluating human employee performance, providing feedback, and defining trusted reviewers. Why not apply those same frameworks to AI systems?

This creates an interesting opportunity for those of us working on evals and AI product optimization. As Surojit put it about Freeplay: "You are building the HR system for the new workforce." The challenge becomes: how do you create evaluation and feedback systems that work with natural language, and without requiring deep technical expertise?

The Reality of Enterprise AI Transformation

Surojit is refreshingly candid about what it actually takes to build enterprise AI systems that work at scale. When we asked about common challenges, he emphasized that true enterprise AI transformation isn't about just adding conversational interfaces to existing software. It’s often about re-thinking inefficient internal processes, and redesigning them to work better with the help of AI.

He shared a compelling example: a large company approached Ema to automate their travel and expense process, which involved nine different email inboxes that humans monitored and triaged. The initial request was simple: create nine agents to replace the humans watching those inboxes.

Surojit's team pushed back: 

"Why do you have nine inboxes? ... This is your opportunity to transform the whole process. You need one inbox. Why does someone need to send an email? Maybe you just need one interface…."

This illustrates a broader challenge facing enterprise AI adoption. Many companies are under pressure to "sprinkle some AI on top" but haven't thoroughly considered how AI enables fundamentally different ways of working. The result, as Surojit notes: "There's a lot of reports of customers who have implemented generative AI not seeing ROI."

Multi-Agent Systems and Composability

Ema's architecture is built around multi-agent systems rather than monolithic single agents. Surojit explains this is essential for explainability and handling enterprise complexity:

"Our AI employees are basically multi-agent systems. We think a single agent system is not necessarily workable to solve complex enterprise use cases... Our ‘AI employee’ is (actually) a whole team of experts."

For a use case like Hitachi's HR transformation, they had to build something that worked across 50,000 employees in multiple portfolio companies with different systems of record (SuccessFactors, Workday, different cloud providers, etc.). Ema created specialized agents to orchestrate it all: some that understand specific platforms, some that handle document processing, some that create dynamic workflow plans, etc.

This composable approach allows Ema to build pre-configured AI employees for common workflows (like hire-to-retire processes) that can be quickly customized for each customer's specific needs. As Surojit describes it, their platform lets them "(turn) services into software very quickly."

The result is implementations that would traditionally take years and tens of millions of dollars in consulting fees for traditional SaaS models can instead be completed in weeks.

The Forward-Deployed Approach

Like other successful enterprise AI companies, Ema has developed what Surojit calls "agentic clinics": collaborative sessions where they work with customers to identify the right problems to solve and design appropriate solutions.

This approach often involves educating customers about what's actually possible and steering them away from suboptimal implementations. The team includes “Customer Value Engineers” with consulting backgrounds who can advise on process transformation, not just feature implementation.

This forward-deployed model reflects a broader shift in enterprise AI. As Surojit notes: "SaaS is kind of dead, or will be dead... The new age of software has to be tied to actual ROI." Embedding deeply and consultatively with customers is the only way to get there (at least today).

Technical Challenges That Remain

When we asked about what's still not working well, Surojit was honest about the limitations:

"The most complex dynamic planning, we are kind of pushing the boundaries on how can we dynamically plan more... I would love to see like a chain of 20 tools dynamically called and planned. We can probably do five, six tools today."

LLM hallucination remains a challenge, which is why Ema built their own small model (MFusion) that sits on top of commercial LLMs to improve accuracy and reduce costs for enterprise-specific tasks.

The longer the context window, the more tools involved, and the more turns in a conversation, the harder it becomes to maintain reliability — even with frontier models. This is consistent with what we hear from Freeplay customers working on complex agent systems.

Advice for Builders

Surojit's advice for teams building AI products is simple but powerful:

"Focus on the customer and just blank out the noise... There's just so much noise in the market... In the end, creating real customer value wins and focus on actually making it work in a sustainable way, not like, okay, it worked one time or two times. Can it work at scale?"

This emphasis on sustainable, scalable solutions over demos resonates with our experience at Freeplay. The teams that succeed with AI aren't the ones with the flashiest demos. They're the ones who invest in robust evaluation systems, clear feedback loops, and processes for continuous improvement.

Looking Forward

Surojit's journey with Ema illustrates what it takes to build enterprise AI systems that work at production scale. The key insights:

• Design AI systems in ways that help users understand how to interact with and improve them

• Focus on transforming processes, not just automating existing workflows

• Build composable systems that can handle enterprise complexity quickly (vs. big monoliths)

• Consider forward-deployed teams who can guide customers to the right solutions

• Focus on sustainable, scalable implementations rather than one-off demos

For anyone building AI products for the enterprise, Surojit's experience offers a valuable roadmap, and a reminder that the fundamentals of product building (focus on customer value, build for scale, solve real problems) haven't changed, even in the age of AI.

Want to hear more conversations like this? Subscribe to Deployed on Spotify, Apple Podcasts, or YouTube.