# Ema Raised $77 Million to Replace Software Seats and Service Hours > Ema says its AI employees can absorb work sold as software and IT services. A buyer still has to prove which licenses, service hours and review costs actually leave the budget. - Published: 2026-09-28 - Author: Gene Dai - Canonical: [https://digidai.github.io/2026/09/28/ema-ai-employees-software-services-budget/](https://digidai.github.io/2026/09/28/ema-ai-employees-software-services-budget/) - Topics: Artificial Intelligence, Enterprise Software, IT Services, Future of Work, AI Agents, AI Economics, Deep Investigation --- ![A paper-cut budget table places an AI worker between stacked software seats and service-hour cards under the headline THE REPLACEMENT TEST.](/images/articles/ema-ai-employees-software-services-budget/cover-v1.jpg) _AI-generated editorial illustration._ On September 23, Ema announced a $77 million Series B and put an unusually direct claim beside the financing. The company does not want to sell one more assistant layered onto an enterprise software contract. It wants its "AI employees" to do work that companies currently buy as software access, implementation projects and service hours. In the [company announcement](https://www.ema.ai/series-b), chief executive Surojit Chatterjee reduced the pitch to six words: "Enterprises do not need more software." He told [TechCrunch](https://techcrunch.com/2026/09/23/ema-raises-77m-as-ai-starts-eating-into-enterprise-software-and-services/) that some customers were moving toward replacing large software applications, leaving the old products to function mostly as databases. He also said Ema could take on integration and consulting work that companies had paid IT-services firms to perform. The round gives that argument a fresh balance sheet. Creaegis led the financing. Accel, Section 32 and Prosus added to earlier investments. Ema said the round brings total funding to $140 million and more than quadruples its 2024 valuation, although it did not disclose the new valuation. TechCrunch reported that the transaction consisted of primary equity rather than debt or sales by existing holders. Ema plans to spend much of the money on sales and marketing after building a company of nearly 200 people. Creaegis managing partner Prakash Parthasarathy described Ema as an execution backbone for autonomous work. That is an investor's account of why the firm led the round, informed by access that public readers do not share. It is not an independent performance result. The description is useful because it names the layer Ema wants to own: not a single HR or IT feature, but the place where work crosses applications and reaches a final state. Capital is the easiest number in the story. Replacement is harder to price. An AI worker may answer an employee without opening the HR portal, yet the employer can still pay for the HR system that holds the record. It may reset a password without a service-desk analyst touching the ticket, while the telephony, identity and ticketing contracts remain. It may complete an invoice check and still require a person to approve the payment, investigate exceptions and repair a bad write. A new outcome charge can arrive before any old license or service commitment leaves the budget. Ema reports production deployments and rapid commercial growth. Most detailed results come from Ema or customer stories hosted on its site. The company has not published a rate card, audited customer cohort, annual recurring revenue figure or contract showing which incumbent bill disappeared. Those omissions do not disprove the product. They set the work a buyer must do before describing automation as replacement. This article examines public information available as of September 28, 2026. Company metrics are labeled as company metrics. No Ema contract, paid product test, customer event log or workforce file was available for this analysis. The replacement test below is an editorial procurement framework, not an independent product benchmark, customer case or legal opinion. ## September 23 put $77 million behind the replacement claim One week before Ema disclosed its round, Gartner's [September forecast](https://www.gartner.com/en/newsroom/press-releases/2026-09-16-gartner-forecasts-worldwide-ai-spending-to-grow-49-point-5-percent-in-2026) put worldwide AI spending at $2.67 trillion in 2026, up 49.5% from 2025. Infrastructure accounts for more than half of that forecast, but the operating categories around Ema are still substantial. Gartner estimated $576.5 billion for AI services, $461.6 billion for AI software and $29.2 billion for agents and assistants. Forecast categories do not describe money already collected, and they do not map neatly onto a buyer's chart of accounts. A consulting firm can include software in a managed service. A software vendor can include implementation credits. An agent company can charge for an outcome while relying on a customer's existing applications and a services partner's people. One workflow can support revenue in all three categories. Ema's financing pitch crosses each boundary. The product coordinates specialized agents, draws on multiple language models and takes actions in existing systems. The company says customers can start in HR, IT or finance, then extend to other functions without another services-heavy deployment. Chatterjee told TechCrunch that Ema charges for completed tasks and business outcomes rather than seats or tokens. No public page defines a standard billable outcome or gives the price of one. The ambition is larger than adding AI features to an incumbent suite. Ema first wraps the applications that a company already owns. If the agent becomes the place where workers ask questions and start processes, Chatterjee argues, some of those applications can later be reduced or removed. The agent company would then capture part of the software budget. If its builder and forward-deployed work also automate integration and support, it could capture part of the services budget as well. Incumbents have their own answer. Software companies are embedding agents inside the systems of record, and services firms are using AI to change delivery. Accenture's [third-quarter filing](https://www.sec.gov/Archives/edgar/data/1467373/000146737326000032/acn-20260531.htm) reported approximately 799,000 workers as of May 31, 2026, 93% utilization and 5% local-currency growth in managed services. The company said clients continued to buy application, infrastructure and operations work while seeking productivity and cost savings from AI. That is one company's disclosure, not a market verdict. It is still a useful counter-signal to the idea that services hours vanish as soon as an agent reaches production. Gartner analyst John-David Lovelock described a more gradual redistribution. In the same September forecast, he said enterprises were turning to services providers less often for broad business transformation and more often for smaller projects that exploit AI features inside incumbent software. If that forecast holds, services work does not simply disappear. Its scope shifts toward integration, cost tracking and specific operating changes, exactly the work an orchestration vendor also wants to absorb. This is unlikely to produce a clean transfer from one category to another. Software vendors will defend their records, permissions and embedded workflows. Services firms will defend their customer knowledge, integration work and exception handling. Agent companies will try to own the execution layer that sits above both. Buyers may save money, but they can also end up paying all three while the new layer matures. The $77 million backs a technical claim that a network of agents can coordinate work across many applications. A product demonstration can show that coordination. Cancelling old spend is a separate commercial claim, visible only in renewal records, service scopes and the human work left after automation runs. ## An AI employee starts by wrapping the old stack "AI employee" sounds like a replacement for a person. Ema's public material describes an execution layer assembled from models, tools, permissions, workflows and human checkpoints. It can present one conversational front door, but it depends on the systems behind that door. The [company website](https://www.ema.ai/) advertises more than 250 integrations and a model-routing layer that can use more than 100 models. TechCrunch reported more than 150 frontier and open-source models. The pages may count different sets or reflect different update dates; neither explains the difference. More important, neither figure establishes that a customer has connected a specific system, permitted a specific action or tested every failure path. Ema's own [custom-integration documentation](https://builder.ema.ai/builder/integrations-data/custom-integrations) makes the operating work visible. A customer defines inputs, writes workflow logic and declares outputs. Custom code can call an external API through Ema's connector, while a custom agent can classify data, extract fields or make a language-based decision. Credentials are injected at runtime. A customer can require human approval before a tool acts. The same documentation lists boundaries that matter after a demo. An API call has a 60-second deadline. A script can make no more than 20 connector calls. Non-success responses do not automatically throw an error, so the customer's logic must inspect status and handle failure. Updated actions take effect after saving, and the current documentation says there is no version history for those updates. A saved action still needs a live connection and real credentials before it can run. These are normal engineering constraints, not evidence of a defective product. They show why orchestration does not erase operations. Someone still owns authentication, field mapping, error behavior, approvals, tests and changes. When an underlying API changes, the agent layer needs a response. When an agent sends an incorrect field or receives a timeout, the business process needs a recoverable state. Behind the simpler employee experience, the old application may become more important as a system of record. A password-reset agent needs an identity store. An HR assistant needs worker and policy data. A finance agent needs a payables or enterprise-resource-planning system that can post the approved result. Removing screens is different from removing records, access controls and statutory retention. A credible case exists for wrapping first. A buyer can avoid a multiyear rip-and-replace project, preserve validated records and give employees a simpler route through fragmented applications. An agent can also coordinate across vendor boundaries that no single incumbent wants to concede. If a worker gets one correct answer instead of searching five portals, value exists even when every old contract remains. That is an experience improvement, though, not yet software displacement. Displacement requires the buyer to name a license, module or support tier that can be reduced without losing a required record, permission or recovery path. If the old product stays at the same quantity and price, the agent has changed how work enters the stack. It has not replaced the stack. For a company with a small technology team, that front door may be the main value. The alternative is not always a well-staffed internal platform group building clean integrations. It can be email, spreadsheets, brittle scripts and employees who know which colleague can repair a stuck request. A packaged agent layer can make that informal system visible and repeatable. Buyers should compare it with the process they actually operate, while keeping the replacement label for costs that leave. Services follow the same logic. An agent builder may generate part of an integration, but someone still selects the workflow, maps authority, tests edge cases and accepts the result. An IT-services partner may use fewer people on a routine queue and more on platform engineering, quality review or new deployments. The work mix changes before the contract total necessarily falls. ## Fifty-fold growth still hides the revenue base For a private AI company, Ema disclosed a substantial set of commercial metrics. The numbers use different denominators, so they cannot be added into one adoption claim. TechCrunch reported more than 50 active enterprise deals, over 1 million active enterprise users and more than 5 million actions and queries handled. Ema said revenue grew fiftyfold in two years. Chatterjee said revenue bookings had passed $150 million, that net dollar retention was about 180%, and that more than 90% of customers expanded beyond the first use case. He also put gross margin near 80%. Each figure can be meaningful. Together they still leave several basic questions open. | Disclosed metric | What it can show | What remains missing | | --------------------------------------- | ------------------------------------------------------- | --------------------------------------------------------------------------------------- | | 50 active enterprise deals | A count of commercial relationships described as active | Contract size, paid production scope, start date and customer concentration | | More than 1 million active users | Potential reach across enterprise populations | Activity definition, period, repeat use and depth of workflow completion | | More than 5 million actions and queries | Product volume | Split between questions and writes, retry count, accepted outcomes and correction rate | | Fiftyfold revenue growth | Fast growth from the starting period | Starting revenue, current revenue, revenue recognition and durability | | More than $150 million in bookings | Contracted multiyear value reported by the company | Annual recurring revenue, cancellation terms, implementation obligations and collection | | About 180% net dollar retention | Expansion among the measured customer cohort | Cohort definition, period, churn, pricing effects and whether services are included | | Close to 80% gross margin | Software-like unit economics claimed by management | Audited calculation, hosting allocation, implementation labor and customer-support cost | TechCrunch added one useful distinction: the $150 million bookings figure includes two- and three-year contract value and is not annual recurring revenue. Ema declined to provide its annualized revenue run rate. A buyer or investor cannot turn a booking into annual revenue by dividing it casually, because contract timing, ramp, termination rights and delivery obligations can differ. Action counts have a similar problem at the workflow level. A query can be answered without changing a record. An action may be attempted, retried, reversed or corrected. Five tool calls can support one employee request. A workflow can complete technically and still produce an outcome the business rejects. Volume proves use. It does not, by itself, show accepted work or displaced cost. Expansion deserves a balanced reading. If more than 90% of customers move beyond an initial use case and net dollar retention reaches 180%, customers are buying more from Ema. That is a stronger commercial signal than a list of pilots. It may also mean the new layer grows before old contracts shrink. Expansion revenue tells Ema that the account is working. It does not tell the buyer that the combined software, services and labor budget fell. Gross margin points in the same direction. An 80% figure would look like software economics, even if the product captures work previously billed as consulting or managed services. But margin is the vendor's relationship between revenue and cost of delivery. Buyer savings are a different calculation. A high-margin provider can deliver a valuable reduction, charge most of that value back to the customer, or sit on top of unreduced incumbent spend. None of these gaps make the reported growth false. They prevent a category error. Funding, bookings, user reach, actions, retention, gross margin and customer savings describe different objects. A replacement claim needs one more object: the old cost that actually left. ## Service desks provide the cleanest case A password reset is easier to count than a broad promise to improve finance or HR. Tickets, calls, escalations and resolutions give Ema its strongest public evidence. A person has a problem, the system takes an action, and the issue either stays closed or returns. Ema's financing announcement highlighted three deployments. One employee assistant supports more than 240,000 people across 65 countries, automates more than 100 workflows and handles about 2.9 million queries a year. Ema reported that response time fell from days to seconds, satisfaction rose 20%, about 60% of tickets were avoided and the people- operations team ran about 50% leaner. Another deployment handles more than 1 million IT-service tickets and more than 1 million calls in over 15 languages, with up to 95% accuracy and fewer than 10% of interactions escalated to a person. A third, described only as a $50 billion global conglomerate, reportedly reached production in four weeks, integrated more than 20 systems, served over 40,000 employees, improved efficiency 70% and cut ticket volume 30%. Those figures are detailed enough to inspect and too incomplete to treat as audited outcomes. The public announcement does not define the satisfaction instrument, the before-and-after periods, the composition of the avoided tickets or the denominator behind a leaner team. It does not say whether the 95% accuracy figure was measured on all calls, a sample, intents the system chose to answer, or a set that excluded transfers and failures. The named [NTT DATA customer story](https://www.ema.ai/customer/ntt-data) supplies more process detail. Ema says NTT DATA uses voice AI as the first line for service desks, with a Fortune 100 insurer as an initial customer. The agent connects to telephony and ticketing systems, resets passwords, provisions access and closes tickets during calls. The page reports more than 1 million calls at scale, 15 or more languages and 90% to 95% of employee queries resolved from end to end. Abhinav Rajan, NTT DATA's global head of business, frames the change as moving from routing an issue to resolving it. His name and role give the claim an accountable sponsor. They do not answer how NTT DATA audits the resolution rate, how many issues reopen or how the Fortune 100 customer defines a security exception. The story appears on Ema's site and promotes Ema's product. It names and quotes an NTT DATA executive. It does not publish the raw call set, severity mix, repeat-contact window, security-event rate or independent audit. A low human- request rate can be positive if issues stay fixed. It can be negative if people cannot reach a person when the system fails. The missing quality window decides which reading is correct. Ema's [Wipro customer story](https://www.ema.ai/customer/wipro) associates the 240,000-person deployment with 2.9 million annual queries, a 20% satisfaction increase and a 50% reduction in HR-operations cost. Ateet Jayaswal, Wipro's chief culture and employee experience officer, is the named executive on that case. Wipro is both a customer and an investor through Wipro Ventures, a relationship buyers should know when weighing the results. Public materials do not show whether the cost reduction means actual spending, avoided future hiring, reassigned workers, a smaller vendor contract or a modeled estimate. Service-desk automation has a strong counterargument in its favor. Password resets and access requests are repetitive, employees dislike waiting, and multilingual coverage is expensive to staff around the clock. If an agent closes a routine issue correctly, the employee gets time back and the human desk can focus on complex incidents. Requiring an independent randomized trial before making any operational change would set an unrealistic bar. A buyer can still run a serious before-and-after test. Keep the same eligible issue types. Count every attempted call, not only those the agent accepts. Measure repeat contact after 24 hours and seven days. Separate human-requested escalation from security-mandated escalation. Record analyst minutes spent reviewing, correcting and recovering work. Track incidents that arise from an incorrect action. Then compare loaded cost per issue that remains resolved. HR and finance require an even higher bar. An employee-policy answer can be fast and still be wrong for a location. A payroll change can close a case and still miss the next pay run. A recruiting workflow can advance or reject a person, which carries a different consequence from resetting a password. The European Commission's [AI Act overview](https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) identifies some employment and worker-management systems as high-risk, with the relevant Annex III rules scheduled for December 2, 2027. The actual classification depends on the system and its use. A broad compliance badge cannot decide that question for every workflow. Employees need a route that aggregate metrics cannot supply. A person who receives the wrong policy answer, access change or payroll action must know that software acted. That person also needs to reach a human, pause the downstream step and obtain a correction tied to the same work ID. Escalation is not automatically a product failure in these cases. It can be the intended safety path. A buyer that rewards only lower escalation can pressure the system to hide the human intervention that protects the employee. ## Outcome pricing moves the argument into the contract A seat is a familiar invoice unit. An outcome is not. Ema's pricing claim is attractive because it appears to align payment with completed work. A seat charges for access whether a person uses the product or not. A service contract may charge for hours even when the process stays slow. A token bill charges for computation without asking whether the answer helped. Outcome pricing seems to move the vendor closer to the buyer's result. The contract has to define the result. Suppose an AI employee resets a password. Completion could mean the identity system accepted the write, the employee logged in once, the employee stayed able to work for seven days, or the issue never reopened. Each definition creates a different bill. If a failed first attempt triggers three retries and then a human resolves the case, the parties must decide whether any outcome fee is due and who pays for the underlying model and system calls. Public Ema materials reviewed for this article do not provide a standard price, rate card, outcome definition, credit policy or dispute window. Enterprise contracts are often negotiated and need not be public. The absence means a reader cannot infer Ema's customer economics from the phrase "outcome based." Buyers already report difficulty forecasting AI-related software charges. Zylo's [2026 analysis](https://zylo.com/blog/software-pricing-volatility), based on its SaaS-management research, says 78% of surveyed IT leaders experienced unexpected AI or consumption charges over the prior year. It also reports that 79% saw renewal price increases and 77% encountered unexpected costs after signing. Zylo sells software-spend management, so the survey supports its commercial argument. Its findings still identify the budget conditions under which a new pricing unit arrives. Outcome pricing can reduce one form of volatility while introducing another. A buyer may know the price of a completed task but not how many tasks employees will route to the agent after adoption. The definition can also reward shallow closure. If the vendor earns money when a ticket closes, repeat contact and correction must sit inside the quality window. Otherwise the vendor can bill twice for one unresolved problem. Hiding tokens on an invoice does not remove the agent's internal cost. McKinsey's [August analysis](https://www.mckinsey.com/capabilities/quantumblack/our-insights/where-AI-agents-pay-off-a-practical-guide-to-the-economics-of-agentic-workflows) estimated a $20,000 to $30,000 operating cost for one customer-facing banking agent and $100,000 to $200,000 for a multiagent team. These are consultancy examples, not Ema bills. They show how model choice, number of runs, tool calls and workflow design can create a wide cost distribution beneath one business result. An outcome contract must connect technical completion to a result that remains valid for a defined period. Human work needed to approve, correct and recover exceptions belongs beside that result, not in an unpriced footnote. Dependent software and service costs need a separate reconciliation. The contract should state whether they are included, reduced or simply billed elsewhere. There is a strong case for letting the vendor manage the model mix. A specialist can choose a cheap model for routine classification, reserve expensive reasoning for difficult cases and switch when a provider fails. The customer should not have to price every token. But hiding implementation detail is different from hiding economic detail. The buyer still needs the eligible-work denominator, accepted-outcome count, quality window, total charge and retained human effort. NIST describes its [AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) as voluntary guidance for the design, development, use and evaluation of AI products, services and systems. That three-part scope fits the commercial problem. An AI employee is not only a model or product. It changes a service process and a human operating system. A buyer that tests only the answer misses the workflow around it. ## A seven-line replacement test for software and services The replacement claim can be tested without waiting for a perfect market benchmark. A buyer needs one bounded workflow, a before period, an after period and seven lines that procurement, the functional owner and finance can read together. ![A top-down paper map connects an accepted work item to a quality clock, a human exception desk and two contract folders.](/images/articles/ema-ai-employees-software-services-budget/replacement-map-v1.jpg) _AI-generated editorial illustration. The map separates accepted work, review effort and costs that can actually be removed._ | Line | Minimum evidence | A replacement claim fails when | | ------------------ | ------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------- | | Eligible work | Fixed workflow, population, channels, exclusions and baseline volume | The vendor reports only tasks the agent chose to accept | | Accepted outcome | Stable work ID, intended final state and business-owner acceptance | Tool calls, answers or closed tickets are counted as final work | | Quality window | Reopen, correction, complaint, policy and incident checks after a defined period | The clock stops at technical completion | | Human load | Review, approval, escalation, exception, recovery and training minutes | Human work moves to managers or employees and disappears from the model | | Loaded cost | Agent fee, models, integration, retained software, services, review and correction | The new fee is compared with only one old line item | | Removed commitment | Cancelled seats or modules, reduced support tier, fewer contracted service hours or documented avoided hiring | An interface disappears while every contract renews unchanged | | Exit path | Exportable workflow, event and decision history; tested fallback; owner for unfinished work | The buyer cannot leave without losing process evidence or operating capacity | Two calculations keep the table honest: `Accepted outcome rate = outcomes still valid after the quality window / all eligible attempts` `Loaded cost per accepted outcome = all platform, model, integration, service, review and correction cost / accepted outcomes` The denominator must include work the agent rejected, timed out on or handed to a person. Otherwise the system improves its apparent success by selecting easy cases. The numerator must exclude a result that was reopened or corrected inside the agreed window. A technical action can remain in the event log without earning the same status as accepted work. Human time belongs in the calculation even when it sits outside the team that bought the agent. If managers approve every action, employees rephrase failed requests, security teams investigate bad writes or analysts maintain policy content, those minutes are part of delivery. They may still be lower than the old process. Counting them makes the saving defensible. Loaded cost also prevents a narrow comparison. An AI employee may reduce a service-desk contract while increasing model spend and requiring a new integration team. It may reduce application seats while the company keeps a premium system-of-record tier. It may avoid hiring during growth without producing a visible headcount reduction. Each can be a valid economic outcome, but they need different evidence. Removed commitment is the decisive line. Buyers should name the actual contract, module, support tier, statement of work, contractor hours or approved future positions that changed. An annual plan can count avoided hiring if the baseline demand and authorized positions existed before deployment. It should not turn an aspirational workforce plan into a cash saving. Exit matters because a replacement is more consequential than an add-on. Ema says it provides audit trails, role-based permissions, on-premises options and human approval. Buyers should verify which event, workflow, configuration and decision records they can export, how unfinished work is transferred and whether the underlying applications can resume direct operation. A successful pilot with no exit test can become a new dependency before it retires the old ones. The table is intentionally small. It does not require a new committee or vocabulary. Existing owners bring their records to one review. Operations defines eligible work, the function accepts outcomes, risk and legal set consequence windows, finance reconciles loaded cost, procurement proves contract removal, and technology tests exit. ## Renewal decides what actually disappeared By renewal, the funding headline has no place in the spreadsheet. Ema, founded in 2023, reports more than 50 enterprise deals, million-user reach, multiyear bookings and expansion across multiple functions. Its product documentation describes an execution system, not a thin chat window. Customer stories put large volumes and named organizations behind the pitch. Evidence remains asymmetric. Ema can count its funding, customers, queries, actions and expansion. A customer must count the software seats, service hours, employee effort and exceptions on the other side. The vendor sees revenue grow when a workflow expands. The buyer sees replacement only when total operating cost or capacity changes without quality falling. Consider an illustrative procurement review, not a reported Ema customer case. A CIO arrives at renewal with three folders. The first contains the agent invoice and its accepted-outcome report. The second contains incumbent software renewals. The third contains a managed-services statement of work and the internal hours spent on approvals and exceptions. Six months earlier, a demonstration showed one conversational route through all of them. The review should now be physical. Which folder became thinner? Which line ended? Which service hours moved? Which employee work disappeared, and which work shifted into review? Did the outcome remain correct long enough to matter? Can the company export the record and recover if the agent is unavailable? A buyer may find that the new layer is worth paying for even before it replaces anything. Faster answers, one front door and fewer handoffs can justify a contract. That is an augmentation case, and it should be approved on those terms. Calling it replacement too early hides the transition cost and gives every old supplier time to renew. Ema's $77 million finances a bid to become the execution layer above enterprise systems. Its share of the budget will be decided one renewal at a time. If the seats, service hours and exception work remain, the buyer did not replace them. It bought another vendor. ## Continue reading - [AI Agents Can Skip the UI, Not the SaaS Bill](https://digidai.github.io/2026/05/15/agent-action-tax-saas-workflow-pricing/): Follow how agent work can remove interface clicks while keeping the underlying software bill. - [HR AI Vendors Want Outcome Pricing. CFOs Want Refund Rights.](https://digidai.github.io/2026/05/19/hr-ai-outcome-pricing-refund-rights/): Compare vendor-defined outcomes with refund windows, corrections and buyer dispute rights. - [HR Agents Crowd the Budget Meeting](https://digidai.github.io/2026/06/20/hr-ai-agents-budget-meeting/): Map the budget owners who must reconcile software, labor, legal exposure and review work. - [Software Seats Meet Salesforce's Agentic Work Units](https://digidai.github.io/2026/08/27/salesforce-agentic-work-units-software-seats/): See why seats, agent activity, credits and accepted outcomes need separate denominators.