Ema’s $77M Raise Highlights a Bigger Shift: AI Agents Are Challenging Per-Seat Software Pricing

Enterprise AI is moving into a phase where the most disruptive question is no longer which chatbot employees prefer. It is whether software should still be sold by the number of humans who log into it.

Ema, a company building multi-agent systems for corporate work, announced a $77 million Series B on September 23. TechCrunch reports that the round brings total funding to $140 million and that Ema is positioning its so-called AI employees to execute multi-step processes across HR, IT, finance and other business functions.

The funding itself is ordinary startup news. The more interesting signal is Ema’s commercial model. CEO Surojit Chatterjee told TechCrunch that the company does not primarily charge by seats or AI tokens. Instead, pricing is tied to tasks and business outcomes.

That model points toward a potentially important change in enterprise software economics. If an agent can complete work across several applications without a person opening each application, the relationship between software usage and employee headcount starts to weaken.

For CIOs and software buyers, this does not mean the SaaS model is about to disappear. It does mean that procurement teams need new ways to measure value, reliability and risk when software acts rather than merely providing an interface.

Why per-seat pricing made sense for traditional SaaS

The seat has been a convenient unit because most enterprise applications are designed around human operators. A salesperson logs into a CRM. An HR employee logs into a human-resources system. A finance analyst uses an expense or planning platform.

Charging for each authorized user roughly connects the vendor’s revenue to organizational adoption. It is easy to understand and easy to budget.

Agents complicate that model. One automated system may perform work that previously required dozens or hundreds of people to interact with several applications. It may run continuously and use APIs rather than graphical interfaces. Counting human logins can become disconnected from the actual value produced.

This is one reason outcome pricing is attractive. Instead of charging for access, a vendor can charge for a completed support case, processed request, reconciled invoice or another measurable unit of work.

But that simplicity is deceptive. Defining a successful outcome is harder than counting a seat.

Outcome pricing needs precise definitions

Imagine an AI system that processes an employee onboarding request. Did it complete the task when it created the account? When the employee received equipment? When all policy acknowledgements were signed? What happens if a human had to correct two fields halfway through?

Those questions become commercial questions when billing depends on completion.

Buyers should define success criteria, exception handling and duplicate work before signing an outcome-based contract. They should also establish how failed attempts are counted and whether a process that requires human rescue is billed at the same rate as a fully automated one.

A useful contract should make the denominator visible. If a vendor says an agent successfully completes 95% of a workflow, the customer needs to know what the other 5% looks like and how much human labor it consumes.

Without that information, outcome pricing can obscure costs rather than align them.

Multi-agent systems target the gaps between applications

Ema’s product strategy is also notable because it does not assume companies will immediately remove their existing software. The system initially works across applications, coordinating agents that interact with the tools and data already present.

This is where enterprise automation has historically been difficult. A business process rarely lives inside one application. Hiring can involve an HR system, identity provider, email, document storage, ticketing, payroll and messaging. Finance processes can cross procurement, ERP, banking and spreadsheets.

Traditional integrations connect systems, but somebody still has to define the workflow. Agentic systems promise to handle more of the interpretation and decision-making between those systems.

That is also where risk increases. An assistant that only drafts text has limited ability to cause operational damage. An agent with credentials to several production systems can create accounts, modify records or trigger external actions.

Organizations should therefore evaluate agents as privileged automation, not as chatbots.

Permissions become an architecture problem

Least privilege is especially important for agents that span departments. A system should not receive a broad administrator credential simply because several workflows are easier to implement that way.

Instead, each action should use the narrowest practical permission. High-impact operations should require explicit approval, and organizations should maintain logs showing what the agent attempted, which tools it used and what changed.

Identity also needs to be attributable. If an automated action appears in an audit log as though it were performed by a human employee, investigation becomes unnecessarily difficult. Agent identities should be distinguishable from human identities while still mapping back to the workflow and responsible owner.

This becomes more important as organizations deploy many specialized agents. Without governance, the result can resemble the service-account sprawl that security teams already struggle to manage.

Reported growth deserves context

TechCrunch reports that Ema says its revenue has grown 50-fold over two years and that bookings have exceeded $150 million. The company also reports net dollar retention of roughly 180% and says more than 90% of customers have expanded beyond their first use case.

Those are meaningful indicators of customer expansion if they hold, but they should be interpreted carefully. Ema is privately held, and bookings are not the same as annual recurring revenue. Chatterjee told TechCrunch that the bookings figure includes the total value of multiyear contracts.

That distinction matters because the enterprise AI market is moving quickly. Long contracts can show demand without revealing how much revenue is recognized in a particular year or what the long-term cost of serving those customers will be.

The stronger signal for buyers is that organizations appear willing to test agents on real business workflows rather than restricting AI to document generation and search.

AI may also pressure IT services

The second-order effect extends beyond software licenses. Enterprise applications often generate consulting and integration work. Companies pay services firms to configure products, map processes, build connectors and maintain implementations.

If agents can perform some integration and workflow adaptation themselves, part of that services spending could shift as well. Ema argues that AI can absorb some implementation work traditionally performed by people.

That does not eliminate the need for systems integrators. Complex organizations still need architecture, governance, change management and domain expertise. But the labor mix could change. A consultant may supervise a fleet of automation tools rather than manually configure every step.

The right pilot measures work, not demos

Companies evaluating agentic platforms should select a workflow with a clear start, finish and measurable baseline. Record current labor hours, processing time, error rates, escalation frequency and software costs before introducing the agent.

Then measure the same outcomes after deployment. Include the time humans spend reviewing, correcting and recovering failed runs. Also include model, integration and observability costs.

This prevents a polished demonstration from being mistaken for an operational improvement.

The larger trend behind Ema’s funding is that AI vendors increasingly want to sell completed work rather than software access. If that model succeeds, enterprise technology budgets may gradually shift from paying for tools people operate toward paying for verified outcomes machines help produce. The winners will not simply be the agents that can perform the most actions. They will be the systems that make those actions measurable, governable and economical.