Ema Raises $77M; AI Employees Keep Human Approval

Ema announced a $77 million Series B on September 23, 2026, led by Creaegis and bringing its total funding to $140 million. Existing investors Accel, S32 and Prosus increased their investments. Ema plans to use the capital to expand sales and develop the enterprise agents it calls “AI Employees” for HR, IT and finance work.

Those agents can plan and execute tasks in existing business applications, but sensitive actions can still require a person’s decision. SiliconANGLE’s account describes agents that check their work and notify human overseers when approval is needed. The division of labor depends on how a customer configures the workflow: the software can carry out routine steps, while an assigned person decides whether a designated action proceeds.

What the round is meant to expand

Ema’s pitch is that an agent can finish a process across applications instead of answering a question and leaving the remaining steps in a support queue. It can gather information, select the next action and work within systems a company already uses. That gives the funding announcement a concrete workplace claim: the product is meant to take on coordination and execution, not merely generate responses.

The commercial figures describe a different kind of progress. TechCrunch’s funding report says Ema has surpassed $150 million in bookings, including the full value of multiyear contracts, while its chief executive declined to disclose the current annualized revenue run rate. Bookings show contracted business over time; they do not reveal current revenue or establish how consistently agents complete tasks without help.

The financing therefore supports an expansion plan, while the operational case rests on what customers are already using. Ema’s agents work around existing applications rather than requiring every underlying system to be replaced. That arrangement matters to the approval question: an agent may initiate or prepare an action in a connected system, but authority over a sensitive decision can remain with the organization operating that system.

What Wipro’s deployment shows

An Ema-hosted Wipro case study describes an assistant serving about 240,000 associates across more than 65 countries, handling roughly 2.9 million employee queries a year and enabling more than 100 live actions. The account names leave and timesheet submission, payslip access, benefits management and service-ticket resolution. It also describes the assistant as a common entry point for HR, IT and payroll services connected to existing enterprise systems.

That is a more specific production example than the broad “AI employee” label. An employee can ask for help and, for supported requests, the assistant can take an action in another application. Yet the annual query count measures interactions, not the number of processes completed autonomously. The case study does not break out which requests ended in an automated action, which were handed to staff, or which required approval before an action took effect.

The case study credits the deployment with faster resolution and higher employee satisfaction, but its published account does not provide an independent audit or enough methodology to assess those outcomes separately from the vendor and customer’s claims. It describes real categories of work and a named deployment; it does not establish that every supported task succeeds without intervention. That distinction is especially important when a query can range from a straightforward policy question to a request that changes an employee record.

Finance remains less visible in this named example. Ema markets agents for finance alongside HR and IT, but the Wipro account chiefly details employee services and payroll access. Accessing a payslip is not evidence that an agent independently authorizes a payment or approves a financial transaction. The public production example supports a claim about employee-service workflows, with a narrower basis for judging finance operations.

Where human approval enters the process

Ema’s human-in-the-loop documentation describes a workflow that pauses, assigns a request to a user or role, and resumes after that person responds; it also allows a timeout to fail the run, skip a step or approve automatically. The same mechanism can seek permission for a sensitive action, collect missing information or ask a clarifying question. An agent can also pause during execution when it needs a detail it cannot determine.

The approval point is thus a configured step in a particular process. For a routine employee request, an agent might retrieve information or submit an allowed change. A request affecting access rights, personnel records or a payment could instead be routed to someone with authority to decide. These are examples of where an organization might place a checkpoint, not a claim that Wipro or every other Ema customer uses those exact rules.

The timeout setting is a consequential detail. A workflow designed to wait for a person can also be set to continue without a response, depending on the option its operator chooses. Human sign-off is therefore available as a control for sensitive work, but the product documentation does not establish that every action receives it. The organization deploying the agent determines which actions require an approver and what happens when that approver does not respond.

What remains unmeasured

The round is confirmed, and Wipro provides a named account of an assistant operating at substantial employee-service volume. The available reporting does not disclose Wipro’s task-level completion rate, the share of requests escalated to people, or the approval rules used in that deployment. It also does not provide comparable, independently verified results for each of the HR, IT and finance functions Ema targets.

Those missing measures define the limit of the “AI employee” claim for now. The public evidence shows software taking actions across business systems and a documented way to stop for a person’s decision. It does not yet show how often the agents finish work alone or how often human judgment determines the result. As Ema expands with the new funding, those figures would reveal the division of labor more clearly than interaction volume alone.