Ema raises $77M Series B to automate enterprise workflows with AI employees
AriaEma's $77M bet that enterprise SaaS is just a database now
Ema just raised $77 million on a simple premise: the future of enterprise software looks nothing like software. It looks like agents.
The Series B, led by Bengaluru-based Creaegis, brings Ema's total funding to $140 million source. Existing investors Accel, Section 32, and Prosus doubled down. The startup's pitch: AI employees that wrap around your existing SaaS stack, then slowly replace it.
The wrap-and-replace strategy
Co-founder Surojit Chatterjee has a background at Google and Coinbase. He describes Ema's approach in plain terms. First you wrap agents around applications in HR, IT, and finance. Then you realize those applications have become little more than data stores with a dashboard on top. The agent owns the workflow. The UI becomes decoration.
It is a bet on structural shift. If the agent layer controls the logic of how work gets done, the traditional application loses its pricing power. Why pay for seats when the agent just reads and writes to the database directly?
The numbers so far
Ema reports more than 50 active enterprise customers and over 1 million active users source. Names include NTT DATA, Hitachi, ADP, PwC, Google, KPMG, Wipro, and Microsoft.
Revenue bookings have crossed $150 million across multiyear contracts source. The company says revenue grew 50x over two years. Net retention sits around 180%, meaning existing customers expand their spend substantially after the first deployment source.
More than 90% of customers expand beyond their initial use case source. That is the number that should get your attention. It means the wedge works. Once an agent handles one workflow, organizations find more workflows to automate.
Pricing tied to outcomes, not tokens
Ema charges for task completion and business outcomes, not per-seat or per-token source. This sounds simple but it is not. Most enterprise AI bills are a function of usage volume, which creates anxiety every time a model call takes longer than expected. Aligning price to completed work changes the conversation entirely.
Gross margins sit near 80% source. Ema claims those margins improve over time as AI systems learn from deployments and require less human support. Whether that holds at scale is an open question, but the trajectory is clear.
The model lab problem
Anthropic and OpenAI are pushing into enterprise. Both have launched forward-deployed engineering teams. Both have announced joint ventures focused on implementation services source.
Chatterjee argues model labs want to sell access, not business process automation. Ema draws on more than 150 models, including open-source options source. The integration depth and domain knowledge become the moat.
This is a reasonable argument. Building a model is different from wiring one into NTT DATA's procurement process. But the model labs have talent and capital in abundance. They will not ignore enterprise margins forever.
What's missing
Ema declined to disclose its annualized revenue run rate. The valuation, while reported to have quadrupled since 2024, was not given source. We also have no breakdown by geography or vertical, no customer acquisition cost data, and no payback period metrics. The 180% net retention is impressive, but without knowing the base revenue it is hard to calculate absolute growth.
My take
Ema is playing the right game at the right time. Enterprises have spent two years trying AI pilots. Now they need someone to make it work in production, and they are willing to pay for outcomes rather than experiments.
The wrap-and-replace framing is clever because it gives SaaS vendors a reason to cooperate rather than compete. If your product becomes a database backend, you still get paid. You just stop pretending the UI was the value.
The risk is execution. Scaling agent orchestration across 150 models, dozens of integrations, and messy enterprise environments is genuinely hard. Chatterjee knows this. His investors are betting he can do it anyway.
Sources: