Emergence AI Raised $97 Million to Make Agents Behave - Then Its Own Test Exposed the Hard Part

The most useful number Emergence AI has published is not the $97.2 million it raised when it stepped out of stealth. It is 49 percent. That was the approximate success rate of its Multi-Agent Orchestrator on 45 enterprise-style tasks in a company-designed benchmark. The result beat three alternatives in the same test. It also meant the orchestrator failed more often than it succeeded.

There, in one awkward statistic, sits the company’s entire case. Businesses have hundreds of applications, uneven data and decades of accumulated process. A single AI agent clicking through screens is not enough. Emergence wants a supervisor that can break work into pieces, pick specialist agents, decide whether to use an interface or an API, and check the result. The better architecture made a marked difference. Reality remained stubborn.

That candor is more interesting than another frictionless demo. Emergence is building what it calls verified autonomy: software that can act across an organization while staying inside policies, permissions and proof-like constraints. The customer is not a person looking for a cheerful chatbot. It is a data team, platform group or industry operator with an expensive workflow and a low tolerance for improvisation.

It began with a browser and an IBM reunion

Satya Nitta, Ravi Kokku and Sharad Sundararajan had worked at IBM Research and later co-founded Merlyn Mind, whose voice-enabled classroom technology reached more than 10,000 classrooms. Nitta saw that the same ingredients could operate workplace software and web applications. The trio formed Emergence around that observation. It was less “invent a digital person” than “teach software to use the software people already have.”

The early expression was Agent-E, an open-source web-navigation agent. It used a hierarchy of planners and workers, stripped clutter from a page’s document structure and watched for changes after an action. On the WebVoyager benchmark, the team reported that it beat prior text and multimodal agents by 10 to 30 percent in most categories. It could search, fill forms and navigate changing sites.

Then broader work exposed the limitation. Agent-E operated through the user interface. That is flexible, but it is also slow and brittle. Buttons move. Pop-ups appear. Travel tasks acquire one more constraint. A page changes while the agent is thinking. In Emergence’s later enterprise benchmark, Agent-E landed at 21 percent success with a median latency of 4.7 minutes. That was what failed first: the browser-only route buckled under compound work.

The change of mind was architectural

The fix was not simply a larger model. In December 2024, Emergence introduced a Multi-Agent Orchestrator that paired a web agent with an API agent. The orchestrator plans a job, delegates it and chooses the best route. An API can create a Jira issue quickly and precisely; a web agent can retrieve information from a portal with no useful API. Cross-vendor neutrality is the sales pitch. Microsoft, Salesforce and Amazon each have agent systems, but an enterprise rarely lives inside one vendor’s garden.

By spring 2025, the system could also create agents. A user describes an outcome in ordinary language. The orchestrator checks its registry, generates a specialist if none fits, assembles a team and retires redundant creations as reusable agents emerge. Emergence calls this recursive intelligence. The phrase sounds grand. The practical function is recognizable: generate the integration code and workflow that consultants or developers would otherwise assemble by hand.

Humans have not disappeared from the loop. Users can interrupt a run, add instructions and approve consequential steps. Access controls and verification rubrics constrain what agents may do. This is not a philosophical concession. It is an enterprise requirement. An agent that writes a plausible report is useful; one that updates a production system on a plausible guess is a liability.

CRAFT starts before the agent gets clever

The current product line moves further down the stack. CRAFT assesses whether company data is ready for agents, enriches metadata, generates quality rules and supports natural-language analysis. Underneath are shared identity, authorization, secrets and asset services. It can run on any conformant Kubernetes cluster in a customer’s cloud, data center or both. That matters to regulated buyers who cannot ship sensitive data to a black box.

The platform now spans data insights, governance and semiconductor analytics. In the semiconductor pitch, agents connect design files, waveforms, test systems, quality tools and the unwritten knowledge in tickets and reports. A workflow can triage an issue, gather evidence, propose hypotheses, validate them and present a fix for review. The company claims five to ten times faster execution across parts of the manufacturing lifecycle and recommends starting with one or two high-impact workflows. Those are company claims, not a published customer case study, but the narrow starting point is sensible.

Ian Eslick is now CEO, while Nitta serves as executive chairman, Kokku as CTO and Sundararajan as CPO. Eslick brings a different kind of enterprise scar tissue: he was CTO of U.S. Bank, led AI adoption at SoFi and earlier co-founded Silicon Spice, sold to Broadcom for $1.2 billion. The leadership change fits the product’s migration from a frontier-agent story toward controlled infrastructure for large institutions.

The customer buys less glue - and more accountability

Emergence sells to enterprises whose work crosses databases, SaaS applications, old portals and policy boundaries. Supply-chain teams can combine SAP data with supplier websites. Financial groups can merge internal history and regulatory documents. Quality teams can turn browser interactions into repeatable software tests. Data users can ask a question in plain English and receive generated SQL, charts and a trail back to the evidence.

The Scratch Foundation is a publicly named user, working with the Orchestrator on platform moderation. Andela is training engineers on the technology and building deployment playbooks. In early 2026, e& enterprise became a distribution and implementation partner for the Middle East, North Africa and Türkiye, with cloud, on-premises and air-gapped options. The partnerships reveal the business model: software plus the implementation muscle required to reach production.

The distinction is not permanent. Every large platform vendor is adding orchestration, governance and model choice. Open-source frameworks are improving quickly. Emergence’s defensible ground, if it has one, is the combination of cross-vendor execution, semantic data preparation, customer-controlled deployment and formal verification research. Its lab is exploring neuro-formal systems that pair probabilistic neural models with mathematical constraints. Its Emergence World project studies what multi-agent populations do over long time horizons, when small decisions compound and tidy demos stop being informative.

What a reader can copy tomorrow

The method will not work everywhere. A stable, deterministic task may be cheaper and safer as an ordinary script or classic robotic process automation. A workflow with no objective definition of success cannot be reliably verified. If errors are irreversible, permissions are broad or the underlying data is incoherent, adding more agents compounds the risk. Even Emergence’s stronger benchmark result is a warning against an unsupervised rollout.

That is why 49 percent is the right number to remember. It is high enough to suggest that coordination and mixed execution paths matter. It is low enough to puncture the fantasy that autonomous enterprise work is solved. Emergence AI has raised a large sum to occupy the distance between those facts. The company wins if that distance becomes infrastructure. It loses if buyers decide the last 51 percent still requires too much human labor, or if the platform giants close the gap first.