TLDR
- Ema CEO Surojit Chatterjee believes AI employees can reduce the use of – and eventually replace – many traditional enterprise software applications.
- EmaFusion routes tasks among different AI models based on cost, complexity and performance rather than relying on a single model.
- Ema faces growing competition as Salesforce, ServiceNow and other SaaS giants build agents that can also work across applications.
Surojit Chatterjee has an ambitious target for his AI startup Ema: Replace many of the software-as-a-service applications enterprises use today.
The founder and CEO of Ema, who is a former Google and Coinbase executive, believes there’s an easier way to get work done than employees accessing dozens of applications daily for their tasks. Use Ema’s “AI employees” instead, which can work across existing systems such as Workday, ServiceNow and Salesforce, carrying out jobs in areas including HR, IT and finance.
By using Ema, Chatterjee’s goal is that over time, companies could reduce – or even eliminate – their use of SaaS apps and decrease the number of software licenses they pay for today.
“Many of the applications will not be needed anymore,” Chatterjee told The AI Innovator in an interview. “That’s what we are going after squarely.”
What’s more is that the AI employee learns about the company over time while incorporating human feedback.
“They also learn on the job by automatically connecting, discovering pathways to automate things, and remembering,” he said. “They are building a context graph of every enterprise – just getting more intelligent, more adaptable inside the enterprise and gets better at the work.”
“That’s a fundamental difference in how software existed before and what we have built,” Chatterjee continued. “It writes its own instructions, writes its own code and dynamically evolves as the enterprise is evolving.”
Founded in 2023, Ema – which stands for Enterprise Machine Assistant – has raised $140 million, including a recent $77 million Series B. It employs around 200 people. Chatterjee said revenue has increased 50-fold over the past two years, without disclosing specific revenue figures.
Ema is embedded into a company’s systems; employees can interact with it through voice, chat or email in 15 languages. Ema can also dynamically create a dashboard if it’s more helpful, say, for things like invoice processing. “It can create on the fly,” he said. “You can pre-configure or create, either way.”
From supplementing to replacing
Chatterjee describes Ema as a digital workforce rather than another AI copilot.
The distinction is important to his argument. A copilot generally helps an employee perform work inside an application. Ema’s AI employees are intended to perform multistep jobs themselves, reaching into the different applications needed to complete them.
“The idea is it’s not about a single assistant or copilot, which assists an individual employee in writing an email or to do something else, but a whole synthetic or digital workforce that can, end-to-end, run business processes,” he said.
That reflects the messy reality of large corporate technology environments. Chatterjee said large enterprises can have thousands of applications, with employees effectively serving as the connection. “Humans are the glue between those applications,” he said. “Can you automate all of that? That’s what we do.”
Ema initially sits above those applications rather than replacing them. Chatterjee describes the strategy as “wrap, reduce, replace”: Ema first operates across existing software, then potentially allows customers to reduce licenses and eventually eliminate some applications.
At Wipro, for example, Chatterjee said Ema is being used for more than 100 employee “hire-to-retire” use cases spanning systems including ServiceNow, Workday and internally developed applications.
The replacement phase remains early. Chatterjee acknowledged that only a small number of Ema customers have eliminated SaaS applications so far. Large companies cannot simply rip out systems overnight, he said, making replacement potentially a multiyear process. However, most clients have been discussing their replacement plans with him and his team.
Picking the right AI for the job
A key piece of Ema’s technology is EmaFusion, which is designed to avoid relying on a single foundation model for every task.
An April 2025 research paper co-authored by Chatterjee describes EmaFusion as a framework that can break a job into smaller problems and determine which AI models are best suited to handle them.
For familiar tasks, a taxonomy-based router identifies appropriate models based on characteristics such as the type and complexity of the job. For ambiguous tasks, a learned router predicts which models are most likely to perform well.
The system can then start with a less expensive model and evaluate its answer before deciding whether to use a more capable – and more expensive – one. A series of AI judges assesses the output, allowing the system to escalate difficult problems while avoiding the expense of using a premium model for every request.
In effect, EmaFusion acts as an AI traffic controller: It decides which models should do which pieces of work and whether their answers are good enough. The architecture can work with both open- and closed-source foundation models rather than tying Ema to a single model provider.
However, the paper also underscores that model routing itself is not unique to Ema. It discusses earlier approaches that similarly route requests to different models and cascade requests from cheaper models to more powerful ones. Ema’s approach attempts to distinguish itself by combining taxonomy-based and learned routing with its judging and cascading system.
The SaaS companies fight back
Chatterjee argues that Ema has another advantage over incumbent SaaS providers: Their existing business models could make it difficult to embrace agents that could cannibalize their existing businesses.
“Can Salesforce take a 90% revenue cut and go from a seat-based model to a fully usage-based model?” he said. “That’s the question for every public (SaaS) company to reckon with.”
But the SaaS companies are not conceding the agent layer – and their agents are increasingly able to operate beyond their own applications.
Salesforce is extending Agentforce beyond its own applications through MuleSoft, which the company says lets its agents retrieve data and take actions in third-party systems using APIs and connectors.
ServiceNow is pursuing a similarly broad strategy. Its AI Agent Fabric connects third-party agents and tools, while its Action Fabric allows agents to reach external tools, data and systems using protocols including MCP and A2A. ServiceNow says its Otto system can route requests and coordinate work across systems.
That sets up a contest over which technology becomes the orchestration layer through which enterprise work gets done. Ema is betting that an independent platform capable of operating across vendors can occupy that position. Incumbent SaaS providers are betting they can extend their own platforms across many of the same systems while retaining the data, workflows, permissions and customer relationships they already control.
Chatterjee maintains that Ema’s horizontal approach is the advantage.
“Work is not done in Salesforce alone,” he said. “Work is done across hundreds of applications.”
AI employees instead of software seats
Ema is also trying to change how companies pay for enterprise software.
Rather than charging for each employee who has access to the system or for the number of AI tokens consumed, Chatterjee said Ema increasingly prices around completed work.
He gave the example of an enterprise receiving one million IT requests. If Ema resolves 80% of the tickets, the customer can pay based on those completed outcomes rather than the number of employees using the system.
“Tickets are no longer tickets. They are handled automatically by our AI employees,” Chatterjee said. As such, “they never become tickets.”
That model fits the broader thesis behind Ema: As AI agents take over more work, the per-user pricing structure becomes a less useful measure of the software’s value.
It also raises one of the biggest questions facing the enterprise software industry. If employees interact primarily with AI agents rather than directly with applications, companies may need fewer conventional SaaS seats even if the underlying software remains in place as a system of record.
For Ema, that intermediate stage could be as important as outright replacement. A company might continue running Salesforce, Workday or ServiceNow underneath while employees increasingly interact with Ema instead.
Chatterjee expects Ema to deepen its presence in HR, IT and finance while expanding into areas including sales and marketing.
The larger bet is that enterprise software is moving from an application-centric model – in which people learn how to operate dozens of individual systems – toward an agent-centric model in which AI coordinates those systems and increasingly performs the work.
“Over time, many of those applications become redundant because they are mostly being used as a database,” Chatterjee said. “Ema is orchestrating, making decisions, creating plans and executing on those plans across those applications for a multitude of business processes.”
Author
Get the latest insights about enterprise AI.
Subscribe to our newsletter. Thank you.
