Evos 完成 $9M 種子前輪融資，開發承接工業營運工作的自主 AI 智能體

At many industrial businesses, the most valuable operating manual is a person. An experienced employee knows which invoice needs another look, which carrier exception can wait and when a routine decision should go to a manager. Evos is betting that capturing those judgments can turn AI from another tool employees manage into software that completes operational work.

The New York startup, formally Evolutionary Operating System, Inc., has emerged from stealth with $9 million in pre-seed funding. The round was co-led by Uncork Capital and Field Ventures, with participation from Crosslink Capital, Utopia Capital Management, Sandwith Ventures and angel investors.

Founded in 2026 by CEO Urav Shah, Evos targets small and mid-sized businesses in physical industries. It says it passed $1 million in annual recurring revenue four weeks after going live. The funding will support hiring and bringing its operators to more businesses across the physical economy. A formal launch of its Workhorse platform is planned for the coming weeks.

From Recording Work to Completing It

Evos operators are software agents, rather than industrial robots. Their work includes answering broker calls, assembling freight loads, pursuing missing documents and reconciling invoices inside the systems a business already uses.

Those tasks can look straightforward until an exception appears. An invoice may disagree with a transportation management system, a required document may be missing, or a customer’s circumstances may justify a different response. Experienced staff often resolve such cases using knowledge that never becomes a formal procedure.

Evos says its operators learn from those decisions and the patterns behind them. The objective is to give an operator responsibility for an ongoing function, rather than requiring a person to initiate every task with another prompt.

That approach fits the broader shift toward coordinating AI agents across industrial workflows. Completing a business process requires more than producing a plausible answer: the system must work with current information, carry out permitted actions and recognize when the decision belongs to a human.

Autonomy Has to Be Earned

The company’s deployment model starts with an approval gate. People approve, reject or modify proposed decisions, and their feedback becomes training data. An operator receives more autonomy for particular tasks and scenarios as it demonstrates agreement with the team. Unfamiliar situations return to a person.

This makes autonomy a property of a defined scope of work. Success on routine invoice checks does not automatically establish competence to negotiate a disputed charge or make a new commercial commitment. The boundary between familiar work and a new exception is therefore central to the product.

Evos describes an initial expert session, a working operator the following day and measurable results within a month. Its deployment documentation adds a sequence of discovery, design, strategy, building, deployment and optimization. Teams map their processes and select use cases, then agree on what stays with people before operators enter production.

The advertised 24-hour deployment is a company claim about getting an operator working. It should not be read as a promise that an entire business function becomes safely autonomous in one day. The release explicitly ties independent operation to demonstrated performance and the customer’s chosen level of oversight.

The Technology Behind an Operator

Evos’s architecture page describes three connected elements: a synchronized data layer over existing business systems, a controller coordinating more than 300 specialized subagents, and an action model tuned for each company.

The data layer connects information from transportation, enterprise resource planning and warehouse systems, email and customer portals. The controller distributes work across narrower reasoning and automation components. Evos describes the company-specific action model as a small language model that improves through use and is not shared across customers.

The same documentation describes structured elicitation of expert knowledge and a progression from shadow mode to supervised and autonomous operation, with guardrails and an audit trail. These are the company’s descriptions of its system, rather than an independent evaluation of its architecture.

The operational challenge is keeping all of those elements aligned. A sound decision based on stale information can still be wrong. A technically successful system update can still exceed the operator’s authority. Reliability depends on the combination of data freshness, decision quality, permissions and escalation behavior.

What the Early Results Show

The announcement’s most concrete example is a freight brokerage using an Evos operator to audit less-than-truckload carrier invoices against its transportation management system. Evos says the operator made nearly 3,500 audit decisions at 97% accuracy during its first two months. Over the latest five weeks, it closed out more than $600,000 in invoice value.

Invoice value is the amount of work processed, not a measure of money saved or revenue generated. The stated accuracy also applies to that audit use case; it is not a platform-wide guarantee.

In the accompanying briefing, Evos reports roughly 82% decision accuracy and about 70% of work handled without human intervention more broadly. The materials do not provide a common testing protocol or enough detail to compare those figures directly with the 97% invoice-audit result. They should be understood as separate company-reported observations.

MyFreightWorld president Mike Chalfant describes improved billing accuracy and efficiency following deployment. For prospective customers, the next questions are how performance is measured, how costly errors are detected and what proportion of exceptions still requires experienced staff. Aggregate accuracy alone cannot answer those questions.

Turning Institutional Knowledge Into Repeatable Operations

The ambition extends beyond reducing the time spent on individual tasks. Evos wants to retain the reasoning that makes a business run well, then apply it consistently as work arrives. That could be valuable when operations depend heavily on a small number of experienced employees.

But capturing existing practice also requires deciding which practices should be preserved. Historical decisions can include outdated assumptions or exceptions that no longer apply. Human corrections and ongoing oversight remain important even after an operator becomes proficient at familiar work.

The pre-seed round backs an early attempt to make operational expertise reusable without asking smaller industrial companies to replace their existing systems. Evos’s commercial test is whether its operators can complete useful work reliably, keep people in control of consequential exceptions and retain those advantages as customers broaden their scope.